feat: support litellm-sdk for reranker endpoint

chore: remove dead code
2026-02-12 14:43:35 +01:00 · 2026-02-12 14:21:53 +01:00 · 2026-02-12 14:21:26 +01:00 · 2026-02-12 14:13:04 +01:00 · 2026-02-12 13:14:57 +01:00 · 2026-02-12 11:26:12 +01:00
435 changed files with 55658 additions and 6380 deletions
@@ -31,6 +31,12 @@ HINDSIGHT_API_HOST=0.0.0.0
 HINDSIGHT_API_PORT=8888
 HINDSIGHT_API_LOG_LEVEL=info

+# Base Path / Reverse Proxy Support (Optional)
+# Set these when deploying behind a reverse proxy with path-based routing
+# Example: To deploy at example.com/hindsight/, set both to "/hindsight"
+# HINDSIGHT_API_BASE_PATH=/hindsight
+# NEXT_PUBLIC_BASE_PATH=/hindsight
+
 # Database (Optional - uses embedded pg0 by default)
 # HINDSIGHT_API_DATABASE_URL=postgresql://user:pass@host:5432/db
 # HINDSIGHT_API_DATABASE_SCHEMA=public  # PostgreSQL schema name (default: public)
@@ -50,3 +56,18 @@ HINDSIGHT_API_LOG_LEVEL=info
 # HINDSIGHT_API_RERANKER_LOCAL_MODEL=cross-encoder/ms-marco-MiniLM-L-6-v2
 # For TEI provider:
 # HINDSIGHT_API_RERANKER_TEI_URL=http://localhost:8081
+
+# Observability & Tracing (Optional - disabled by default)
+# Enable OpenTelemetry tracing for LLM calls (GenAI semantic conventions)
+# HINDSIGHT_API_OTEL_TRACES_ENABLED=true
+#
+# Local development with Grafana LGTM stack (recommended - see scripts/dev/grafana/README.md)
+# HINDSIGHT_API_OTEL_EXPORTER_OTLP_ENDPOINT=http://localhost:4318
+#
+# Cloud backends (Grafana Cloud, Langfuse, DataDog, etc.)
+# HINDSIGHT_API_OTEL_EXPORTER_OTLP_ENDPOINT=https://your-backend-url
+# HINDSIGHT_API_OTEL_EXPORTER_OTLP_HEADERS="Authorization=Bearer your-token"
+#
+# Custom service name and environment (optional, defaults: hindsight-api, development)
+# HINDSIGHT_API_OTEL_SERVICE_NAME=hindsight-production
+# HINDSIGHT_API_OTEL_DEPLOYMENT_ENVIRONMENT=production
@@ -188,6 +188,55 @@ jobs:
        path: hindsight-integrations/openclaw/*.tgz
        retention-days: 1

+  release-ai-sdk-integration:
+    runs-on: ubuntu-latest
+    environment: npm
+
+    steps:
+    - uses: actions/checkout@v4
+
+    - name: Set up Node.js
+      uses: actions/setup-node@v4
+      with:
+        node-version: '22'
+        registry-url: 'https://registry.npmjs.org'
+
+    - name: Install dependencies
+      working-directory: ./hindsight-integrations/ai-sdk
+      run: npm ci
+
+    - name: Build
+      working-directory: ./hindsight-integrations/ai-sdk
+      run: npm run build
+
+    - name: Publish to npm
+      working-directory: ./hindsight-integrations/ai-sdk
+      run: |
+        set +e
+        OUTPUT=$(npm publish --access public 2>&1)
+        EXIT_CODE=$?
+        echo "$OUTPUT"
+        if [ $EXIT_CODE -ne 0 ]; then
+          if echo "$OUTPUT" | grep -q "cannot publish over"; then
+            echo "Package version already published, skipping..."
+            exit 0
+          fi
+          exit $EXIT_CODE
+        fi
+      env:
+        NODE_AUTH_TOKEN: ${{ secrets.NPM_TOKEN }}
+
+    - name: Pack for GitHub release
+      working-directory: ./hindsight-integrations/ai-sdk
+      run: npm pack
+
+    - name: Upload artifacts
+      uses: actions/upload-artifact@v4
+      with:
+        name: ai-sdk-integration
+        path: hindsight-integrations/ai-sdk/*.tgz
+        retention-days: 1
+
  release-control-plane:
    runs-on: ubuntu-latest
    environment: npm
@@ -291,6 +340,7 @@ jobs:
        retention-days: 1

  release-docker-images:
+    name: Release Docker (${{ matrix.image_name }}${{ matrix.tag_suffix }})
    runs-on: ubuntu-latest
    permissions:
      contents: read
@@ -300,10 +350,28 @@ jobs:
        include:
          - target: api-only
            image_name: hindsight-api
+            tag_suffix: ""
+            build_args: ""
+          - target: api-only
+            image_name: hindsight-api
+            tag_suffix: "-slim"
+            build_args: |
+              INCLUDE_LOCAL_MODELS=false
+              PRELOAD_ML_MODELS=false
          - target: cp-only
            image_name: hindsight-control-plane
+            tag_suffix: ""
+            build_args: ""
          - target: standalone
            image_name: hindsight
+            tag_suffix: ""
+            build_args: ""
+          - target: standalone
+            image_name: hindsight
+            tag_suffix: "-slim"
+            build_args: |
+              INCLUDE_LOCAL_MODELS=false
+              PRELOAD_ML_MODELS=false

    steps:
    - uses: actions/checkout@v4
@@ -341,6 +409,9 @@ jobs:
      uses: docker/metadata-action@v5
      with:
        images: ghcr.io/${{ github.repository_owner }}/${{ matrix.image_name }}
+        flavor: |
+          latest=auto
+          suffix=${{ matrix.tag_suffix }}
        tags: |
          type=semver,pattern={{version}},value=${{ steps.get_version.outputs.VERSION }}
          type=semver,pattern={{major}}.{{minor}},value=${{ steps.get_version.outputs.VERSION }}
@@ -366,7 +437,7 @@ jobs:
    # - name: Smoke test - verify container starts
    #   env:
    #     GROQ_API_KEY: ${{ secrets.GROQ_API_KEY }}
-    #   run: ./scripts/docker-smoke-test.sh "${{ matrix.image_name }}:test" "${{ matrix.target }}"
+    #   run: ./docker/test-image.sh "${{ matrix.image_name }}:test" "${{ matrix.target }}"

    # Build multi-platform and push to release tags
    - name: Build and push release images
@@ -375,6 +446,7 @@ jobs:
        context: .
        file: docker/standalone/Dockerfile
        target: ${{ matrix.target }}
+        build-args: ${{ matrix.build_args }}
        push: true
        platforms: linux/amd64,linux/arm64
        tags: ${{ steps.meta.outputs.tags }}
@@ -415,7 +487,7 @@ jobs:

  create-github-release:
    runs-on: ubuntu-latest
-    needs: [release-python-packages, release-typescript-client, release-openclaw-integration, release-control-plane, release-rust-cli, release-docker-images, release-helm-chart]
+    needs: [release-python-packages, release-typescript-client, release-openclaw-integration, release-ai-sdk-integration, release-control-plane, release-rust-cli, release-docker-images, release-helm-chart]
    permissions:
      contents: write

@@ -444,6 +516,12 @@ jobs:
        name: openclaw-integration
        path: ./artifacts/openclaw-integration

+    - name: Download AI SDK Integration
+      uses: actions/download-artifact@v4
+      with:
+        name: ai-sdk-integration
+        path: ./artifacts/ai-sdk-integration
+
    - name: Download Control Plane
      uses: actions/download-artifact@v4
      with:
@@ -487,6 +565,8 @@ jobs:
        cp artifacts/typescript-client/*.tgz release-assets/ || true
        # OpenClaw Integration
        cp artifacts/openclaw-integration/*.tgz release-assets/ || true
+        # AI SDK Integration
+        cp artifacts/ai-sdk-integration/*.tgz release-assets/ || true
        # Control Plane
        cp artifacts/control-plane/*.tgz release-assets/ || true
        # Rust CLI binaries
@@ -9,42 +9,11 @@ concurrency:
  cancel-in-progress: true

 jobs:
-  build-python-packages:
-    runs-on: ubuntu-latest
-    strategy:
-      matrix:
-        include:
-          - name: hindsight-all
-            path: hindsight
-          - name: hindsight-api
-            path: hindsight-api
-          - name: hindsight-client
-            path: hindsight-clients/python
-          - name: hindsight-embed
-            path: hindsight-embed
-
-    steps:
-    - uses: actions/checkout@v4
-
-    - name: Install uv
-      uses: astral-sh/setup-uv@v5
-      with:
-        enable-cache: true
-
-    - name: Set up Python
-      uses: actions/setup-python@v5
-      with:
-        python-version-file: ".python-version"
-
-    - name: Build ${{ matrix.name }}
-      working-directory: ./${{ matrix.path }}
-      run: uv build
-
  build-api-python-versions:
    runs-on: ubuntu-latest
    strategy:
      matrix:
-        python-version: ['3.11', '3.12', '3.13']
+        python-version: ['3.11', '3.12', '3.13', '3.14']

    steps:
    - uses: actions/checkout@v4
@@ -105,6 +74,29 @@ jobs:
      working-directory: ./hindsight-integrations/openclaw
      run: npm run build

+  build-ai-sdk-integration:
+    runs-on: ubuntu-latest
+
+    steps:
+    - uses: actions/checkout@v4
+
+    - name: Set up Node.js
+      uses: actions/setup-node@v4
+      with:
+        node-version: '22'
+
+    - name: Install dependencies
+      working-directory: ./hindsight-integrations/ai-sdk
+      run: npm ci
+
+    - name: Run tests
+      working-directory: ./hindsight-integrations/ai-sdk
+      run: npm test
+
+    - name: Build
+      working-directory: ./hindsight-integrations/ai-sdk
+      run: npm run build
+
  build-control-plane:
    runs-on: ubuntu-latest

@@ -285,16 +277,35 @@ jobs:
      run: helm lint helm/hindsight

  build-docker-images:
+    name: Build Docker (${{ matrix.name }})
    runs-on: ubuntu-latest
    strategy:
      matrix:
        include:
          - target: api-only
            name: api
+            variant: full
+            build_args: ""
+          - target: api-only
+            name: api-slim
+            variant: slim
+            build_args: |
+              INCLUDE_LOCAL_MODELS=false
+              PRELOAD_ML_MODELS=false
          - target: cp-only
            name: control-plane
+            variant: full
+            build_args: ""
          - target: standalone
            name: standalone
+            variant: full
+            build_args: ""
+          - target: standalone
+            name: standalone-slim
+            variant: slim
+            build_args: |
+              INCLUDE_LOCAL_MODELS=false
+              PRELOAD_ML_MODELS=false

    steps:
    - uses: actions/checkout@v4
@@ -313,20 +324,31 @@ jobs:
    - name: Set up Docker Buildx
      uses: docker/setup-buildx-action@v3

-    - name: Build ${{ matrix.name }} image
+    - name: Build ${{ matrix.name }} image (${{ matrix.variant }})
      uses: docker/build-push-action@v6
      with:
        context: .
        file: docker/standalone/Dockerfile
        target: ${{ matrix.target }}
+        build-args: ${{ matrix.build_args }}
        push: false
-        load: false
+        load: ${{ matrix.variant == 'slim' }}
+        tags: hindsight-${{ matrix.name }}:test
+        # Removed GitHub Actions cache (type=gha) - it frequently returns 502 errors
+        # causing buildx to fail with "failed to parse error response 502"
+        # Build will be slower but more reliable

-    # TODO: Re-enable smoke test when disk space issue is resolved
-    # - name: Smoke test - verify container starts
-    #   env:
-    #     GROQ_API_KEY: ${{ secrets.GROQ_API_KEY }}
-    #   run: ./scripts/docker-smoke-test.sh "hindsight-${{ matrix.name }}:test" "${{ matrix.target }}"
+    # Only test slim variants to save disk space (they're much smaller)
+    # Slim variants require external embedding providers
+    - name: Smoke test - verify container starts
+      if: matrix.variant == 'slim'
+      env:
+        GROQ_API_KEY: ${{ secrets.GROQ_API_KEY }}
+        HINDSIGHT_API_EMBEDDINGS_PROVIDER: openai
+        HINDSIGHT_API_EMBEDDINGS_OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
+        HINDSIGHT_API_RERANKER_PROVIDER: cohere
+        HINDSIGHT_API_COHERE_API_KEY: ${{ secrets.COHERE_API_KEY }}
+      run: ./docker/test-image.sh "hindsight-${{ matrix.name }}:test" "${{ matrix.target }}"

  test-api:
    runs-on: ubuntu-latest
@@ -749,9 +771,9 @@ jobs:
  test-embed:
    runs-on: ubuntu-latest
    env:
-      HINDSIGHT_EMBED_LLM_PROVIDER: groq
-      HINDSIGHT_EMBED_LLM_API_KEY: ${{ secrets.GROQ_API_KEY }}
-      HINDSIGHT_EMBED_LLM_MODEL: openai/gpt-oss-20b
+      HINDSIGHT_API_LLM_PROVIDER: groq
+      HINDSIGHT_API_LLM_API_KEY: ${{ secrets.GROQ_API_KEY }}
+      HINDSIGHT_API_LLM_MODEL: openai/gpt-oss-20b
      # Prefer CPU-only PyTorch in CI
      UV_INDEX: pytorch=https://download.pytorch.org/whl/cpu

@@ -782,10 +804,62 @@ jobs:
          ${{ runner.os }}-huggingface-embed-
          ${{ runner.os }}-huggingface-

+    - name: Run unit and integration tests
+      working-directory: ./hindsight-embed
+      run: uv run pytest tests/ -v
+
    - name: Run smoke test
      working-directory: ./hindsight-embed
      run: ./test.sh

+  test-hindsight-all:
+    runs-on: ubuntu-latest
+    env:
+      HINDSIGHT_API_LLM_PROVIDER: groq
+      HINDSIGHT_API_LLM_API_KEY: ${{ secrets.GROQ_API_KEY }}
+      HINDSIGHT_API_LLM_MODEL: openai/gpt-oss-20b
+      # For test_server_integration.py compatibility
+      HINDSIGHT_LLM_PROVIDER: groq
+      HINDSIGHT_LLM_API_KEY: ${{ secrets.GROQ_API_KEY }}
+      HINDSIGHT_LLM_MODEL: openai/gpt-oss-20b
+      # Prefer CPU-only PyTorch in CI
+      UV_INDEX: pytorch=https://download.pytorch.org/whl/cpu
+
+    steps:
+    - uses: actions/checkout@v4
+
+    - name: Install uv
+      uses: astral-sh/setup-uv@v5
+      with:
+        enable-cache: true
+        prune-cache: false
+
+    - name: Set up Python
+      uses: actions/setup-python@v5
+      with:
+        python-version-file: ".python-version"
+
+    - name: Build hindsight-all
+      working-directory: ./hindsight
+      run: uv build
+
+    - name: Install dependencies
+      working-directory: ./hindsight
+      run: uv sync --frozen --extra test --index-strategy unsafe-best-match
+
+    - name: Cache HuggingFace models
+      uses: actions/cache@v4
+      with:
+        path: ~/.cache/huggingface
+        key: ${{ runner.os }}-huggingface-all-${{ hashFiles('hindsight/pyproject.toml') }}
+        restore-keys: |
+          ${{ runner.os }}-huggingface-all-
+          ${{ runner.os }}-huggingface-
+
+    - name: Run unit tests
+      working-directory: ./hindsight
+      run: uv run pytest tests/ -v
+
  test-doc-examples:
    runs-on: ubuntu-latest
    needs: test-rust-cli
@@ -958,6 +1032,18 @@ jobs:
      working-directory: ./hindsight-dev
      run: uv run pytest upgrade_tests/ -v --tb=short

+    - name: Show upgrade test logs
+      if: always()
+      run: |
+        echo "=== Upgrade Test Server Logs ==="
+        for log in /tmp/upgrade-test-*.log; do
+          if [ -f "$log" ]; then
+            echo ""
+            echo "--- $log ---"
+            tail -500 "$log"
+          fi
+        done
+
  verify-generated-files:
    runs-on: ubuntu-latest
    env:
@@ -1028,4 +1114,52 @@ jobs:
          git diff --stat
          exit 1
        fi
-        echo "✓ All generated files are up to date"
+        echo "✓ All generated files are up to date"
+
+  check-openapi-compatibility:
+    runs-on: ubuntu-latest
+    env:
+      UV_INDEX: pytorch=https://download.pytorch.org/whl/cpu
+
+    steps:
+    - uses: actions/checkout@v4
+      with:
+        fetch-depth: 0  # Fetch full git history to access base branch
+
+    - name: Install uv
+      uses: astral-sh/setup-uv@v5
+      with:
+        enable-cache: true
+
+    - name: Set up Python
+      uses: actions/setup-python@v5
+      with:
+        python-version-file: ".python-version"
+
+    - name: Install hindsight-dev dependencies
+      run: |
+        cd hindsight-dev && uv sync --frozen --index-strategy unsafe-best-match
+
+    - name: Check OpenAPI compatibility with base branch
+      run: |
+        # Get the base branch (usually main)
+        BASE_BRANCH="${{ github.base_ref }}"
+
+        if [ -z "$BASE_BRANCH" ]; then
+          echo "⚠️  Warning: No base branch found (not a PR?). Skipping compatibility check."
+          exit 0
+        fi
+
+        echo "Checking OpenAPI compatibility against base branch: $BASE_BRANCH"
+
+        # Extract the old OpenAPI spec from base branch
+        git show "origin/$BASE_BRANCH:hindsight-docs/static/openapi.json" > /tmp/old-openapi.json
+
+        if [ ! -s /tmp/old-openapi.json ]; then
+          echo "⚠️  Warning: Could not find OpenAPI spec in base branch. Skipping compatibility check."
+          exit 0
+        fi
+
+        # Check compatibility using our tool
+        cd hindsight-dev
+        uv run check-openapi-compatibility /tmp/old-openapi.json ../hindsight-docs/static/openapi.json
@@ -53,4 +53,6 @@ hindsight-clients/rust/target
 whats-next.md
 TASK.md
 # Changelog is now tracked in hindsight-docs/src/pages/changelog.md
-# CHANGELOG.md
+# CHANGELOG.md
+
+blog-post*
@@ -45,6 +45,7 @@ cd hindsight-control-plane && npm run dev
 ./scripts/dev/start-docs.sh
 ```

+
 ### Generating Clients/OpenAPI
 ```bash
 # Regenerate OpenAPI spec after API changes (REQUIRED after changing endpoints)
@@ -237,26 +238,61 @@ def process(data: UserData) -> str:

 ### Adding New API Configuration Flags

-When adding a new environment variable configuration:
+Configuration follows a hierarchical system: **Global (env vars) → Tenant (via extension) → Bank (database)**.
+
+Fields must be categorized as either **hierarchical** (can be overridden per-tenant/bank) or **static** (server-level only).
+
+#### Adding a New Configuration Field

 1. **config.py** (`hindsight-api/hindsight_api/config.py`):
-   - Add `ENV_*` constant for the environment variable name
+   - Add `ENV_*` constant for the environment variable name (e.g., `ENV_MY_SETTING = "HINDSIGHT_API_MY_SETTING"`)
   - Add `DEFAULT_*` constant for the default value
-   - Add field to `HindsightConfig` dataclass
+   - Add field to `HindsightConfig` dataclass with type annotation
+   - **Mark as hierarchical or static** by adding to `_HIERARCHICAL_FIELDS` set (hierarchical) or leaving it out (static)
   - Add initialization in `from_env()` method

+   ```python
+   # Hierarchical field (can be overridden per-bank)
+   _HIERARCHICAL_FIELDS = {
+       ...,
+       "my_setting",  # Add here for hierarchical
+   }
+
+   # Static field - just don't add to _HIERARCHICAL_FIELDS
+   ```
+
 2. **main.py** (`hindsight-api/hindsight_api/main.py`):
   - Add field to the manual `HindsightConfig()` constructor call (search for "CLI override")

-3. **Use the config** in code:
+3. **Use hierarchical config in MemoryEngine**:
+   ```python
+   # Config is resolved automatically per bank via ConfigResolver
+   config_dict = await self._config_resolver.get_bank_config(bank_id, context)
+   value = config_dict["my_setting"]
+   ```
+
+4. **Use static config** (non-hierarchical):
   ```python
   from ...config import get_config
   config = get_config()
-   value = config.your_new_field
+   value = config.my_static_field
   ```

-4. **Documentation** (`hindsight-docs/docs/developer/configuration.md`):
+5. **Documentation** (`hindsight-docs/docs/developer/configuration.md`):
   - Add to appropriate section table with Variable, Description, Default
+   - Mark if it's hierarchical (can be overridden per-bank)
+
+#### Hierarchical vs Static Guidelines
+
+**Hierarchical** (per-bank overridable):
+- LLM settings (provider, model, API key, base URL)
+- Operation-specific settings (retain mode, chunk size, etc.)
+- Feature flags that vary by customer/bank
+
+**Static** (server-level only):
+- Infrastructure settings (database URL, port, host)
+- Global limits (max concurrent operations)
+- System-wide feature flags

 ## Environment Setup

@@ -280,3 +316,4 @@ Optional (uses local models by default):
 - `HINDSIGHT_API_EMBEDDINGS_PROVIDER`: local (default) or tei
 - `HINDSIGHT_API_RERANKER_PROVIDER`: local (default) or tei
 - `HINDSIGHT_API_DATABASE_URL`: External PostgreSQL (uses embedded pg0 by default)
+- `HINDSIGHT_API_ENABLE_BANK_CONFIG_API`: Enable per-bank config API (default: false, disabled for security)
@@ -2,7 +2,7 @@

 ![Hindsight Banner](./hindsight-docs/static/img/hindsight-github-banner.png)

-[Documentation](https://hindsight.vectorize.io) • [Paper](https://arxiv.org/abs/2512.12818) • [Cookbook](https://hindsight.vectorize.io/cookbook) • [Hindsight Cloud](https://vectorize.io/hindsight/cloud)
+[Documentation](https://hindsight.vectorize.io) • [Paper](https://arxiv.org/abs/2512.12818) • [Cookbook](https://hindsight.vectorize.io/cookbook) • [Hindsight Cloud](https://ui.hindsight.vectorize.io/signup)

 [![CI](https://github.com/vectorize-io/hindsight/actions/workflows/release.yml/badge.svg)](https://github.com/vectorize-io/hindsight/actions/workflows/release.yml)
 [![Slack Community](https://img.shields.io/badge/Slack-Join%20Community-4A154B?logo=slack)](https://join.slack.com/t/hindsight-space/shared_invite/zt-3nhbm4w29-LeSJ5Ixi6j8PdiYOCPlOgg)
@@ -42,27 +42,51 @@ If you need more control over how and when your agent stores and recalls memorie

 ![Hindsight Banner](./hindsight-docs/static/img/migration-code.png)

+---
+
+> 🤖 **Using a coding agent?** Install the Hindsight documentation skill for instant access to docs while you code:
+> ```bash
+> npx skills add https://github.com/vectorize-io/hindsight --skill hindsight-docs
+> ```
+> Works with Claude Code, Cursor, and other AI coding assistants.
+
+---
+

 ## Quick Start

 ### Docker (recommended)

 ```bash
-export OPENAI_API_KEY=your-key
+export OPENAI_API_KEY=sk-xxx

 docker run --rm -it --pull always -p 8888:8888 -p 9999:9999 \
  -e HINDSIGHT_API_LLM_API_KEY=$OPENAI_API_KEY \
-  -e HINDSIGHT_API_LLM_MODEL=o3-mini \
  -v $HOME/.hindsight-docker:/home/hindsight/.pg0 \
  ghcr.io/vectorize-io/hindsight:latest
 ```

+>API: http://localhost:8888
+>UI: http://localhost:9999
+
 You can modify the LLM provider by setting `HINDSIGHT_API_LLM_PROVIDER`. Valid options are `openai`, `anthropic`, `gemini`, `groq`, `ollama`, and `lmstudio`. The documentation provides more details on [supported models](https://hindsight.vectorize.io/developer/models).

-API: http://localhost:8888  
-UI: http://localhost:9999

-Install client:
+
+### Docker (external PostgreSQL)
+
+```bash
+export OPENAI_API_KEY=sk-xxx
+export HINDSIGHT_DB_PASSWORD=choose-a-password
+cd docker/docker-compose
+docker compose up 
+```
+
+
+>API: http://localhost:8888
+>UI: http://localhost:9999
+
+### Client

 ```bash
 pip install hindsight-client -U
@@ -70,7 +94,7 @@ pip install hindsight-client -U
 npm install @vectorize-io/hindsight-client
 ```

-Python example:
+#### Python

 ```python
 from hindsight_client import Hindsight
@@ -87,7 +111,29 @@ client.recall(bank_id="my-bank", query="What does Alice do?")
 client.reflect(bank_id="my-bank", query="Tell me about Alice")
 ```

-### Python (embedded, no Docker)
+#### Node.js / TypeScript
+
+```bash
+npm install @vectorize-io/hindsight-client
+```
+
+```javascript
+const { HindsightClient } = require('@vectorize-io/hindsight-client');
+
+const main = async () => {
+  const client = new HindsightClient({ baseUrl: 'http://localhost:8888' });
+
+  await client.retain('my-bank', 'Alice loves hiking in Yosemite');
+
+  const results = await client.recall('my-bank', 'What does Alice like?');
+  console.log(results);
+}
+
+main();
+```
+
+
+### Python Embedded (no server required)

 ```bash
 pip install hindsight-all -U
@@ -107,20 +153,6 @@ with HindsightServer(
    results = client.recall(bank_id="my-bank", query="Where does Alice work?")
 ```

-### Node.js / TypeScript
-
-```bash
-npm install @vectorize-io/hindsight-client
-```
-
-```javascript
-const { HindsightClient } = require('@vectorize-io/hindsight-client');
-
-const client = new HindsightClient({ baseUrl: 'http://localhost:8888' });
-
-await client.retain('my-bank', 'Alice loves hiking in Yosemite');
-await client.recall('my-bank', 'What does Alice like?');
-```

 ---

@@ -0,0 +1,54 @@
+# Docker Compose file for Hindsight with PostgreSQL and pgvector
+#
+# Make sure to set the required environment variables before running:
+# - HINDSIGHT_DB_PASSWORD: Password for the PostgreSQL user
+# - Configure LLM provider variables as needed (see below in the hindsight service)
+#
+# Usage:
+# docker compose up -d
+#
+# Optional environment variables with defaults:
+# - HINDSIGHT_VERSION: Hindsight application version (default: latest)
+# - HINDSIGHT_DB_USER: PostgreSQL user (default: hindsight_user)
+# - HINDSIGHT_DB_NAME: PostgreSQL database name (default: hindsight_db)
+# - HINDSIGHT_DB_VERSION: PostgreSQL version (default: 18)
+
+services:
+  db:
+    # Use a PostgreSQL-Image with pgvector extension pre-installed
+    # see https://hub.docker.com/r/pgvector/pgvector 
+    image: pgvector/pgvector:pg${HINDSIGHT_DB_VERSION:-18}
+    container_name: hindsight-db
+    restart: always
+    # Expose PostgreSQL port
+    # ports:
+      # - "5432:5432"
+    environment:
+      POSTGRES_USER: ${HINDSIGHT_DB_USER:-hindsight_user}
+      POSTGRES_PASSWORD: ${HINDSIGHT_DB_PASSWORD:?Please set the HINDSIGHT_DB_PASSWORD env variable}
+      POSTGRES_DB: ${HINDSIGHT_DB_NAME:-hindsight_db}
+    volumes:
+      - pg_data:/var/lib/postgresql/${HINDSIGHT_DB_VERSION:-18}/docker
+    networks:
+      - hindsight-net
+
+  hindsight:
+    image: ghcr.io/vectorize-io/hindsight:${HINDSIGHT_VERSION:-latest}
+    container_name: hindsight-app
+    ports:
+      - "8888:8888"
+      - "9999:9999"
+    environment:
+      - HINDSIGHT_API_LLM_API_KEY=${OPENAI_API_KEY?Please set the OPENAI_API_KEY env variable}
+      - HINDSIGHT_API_DATABASE_URL=postgresql://${HINDSIGHT_DB_USER:-hindsight_user}:${HINDSIGHT_DB_PASSWORD:?Please set the HINDSIGHT_DB_PASSWORD env variable}@db:5432/${HINDSIGHT_DB_NAME:-hindsight_db}
+    depends_on:
+      - db
+    networks:
+      - hindsight-net
+
+networks:
+  hindsight-net:
+    driver: bridge
+
+volumes:
+  pg_data:
@@ -0,0 +1,96 @@
+# Nginx Reverse Proxy with Custom Base Path
+
+Deploy Hindsight API under `/hindsight` (or any custom path) using Nginx reverse proxy.
+
+## Quick Start (Published Image - API Only)
+
+```bash
+docker-compose up
+```
+
+- **API:** http://localhost:8080/hindsight/docs
+- **Control Plane:** http://localhost:9999 (direct access, not proxied)
+
+## Full Stack with Custom Base Path (Requires Build)
+
+**Important:** You cannot rebuild from the published image with build args. You must build from source.
+
+### Build from Source with Custom Base Path
+
+1. **Clone the repository** (if you haven't):
+```bash
+git clone https://github.com/vectorize-io/hindsight.git
+cd hindsight
+```
+
+2. **Build with base path**:
+```bash
+docker build \
+  --build-arg NEXT_PUBLIC_BASE_PATH=/hindsight \
+  -f docker/standalone/Dockerfile \
+  -t hindsight:custom \
+  .
+```
+
+3. **Update docker-compose.yml** to use your built image:
+```yaml
+services:
+  hindsight:
+    image: hindsight:custom  # ← Change this
+    environment:
+      HINDSIGHT_API_BASE_PATH: /hindsight
+      NEXT_PUBLIC_BASE_PATH: /hindsight
+```
+
+4. **Update nginx.conf** to handle Control Plane routes (see below)
+
+5. **Run**:
+```bash
+docker-compose up
+```
+
+### Required nginx.conf for Full Stack
+
+Replace the current `nginx.conf` with this to proxy both API and Control Plane:
+
+```nginx
+events { worker_connections 1024; }
+
+http {
+    include /etc/nginx/mime.types;
+    default_type application/octet-stream;
+
+    upstream hindsight_api { server hindsight:8888; }
+    upstream hindsight_cp { server hindsight:9999; }
+
+    server {
+        listen 80;
+
+        # API
+        location ~ ^/hindsight/(docs|openapi\.json|health|metrics|v1|mcp) {
+            proxy_pass http://hindsight_api;
+            proxy_set_header Host $http_host;
+        }
+
+        # Control Plane static files
+        location ~ ^/hindsight/_next/ {
+            proxy_pass http://hindsight_cp;
+            proxy_set_header Host $http_host;
+        }
+
+        # Control Plane UI
+        location /hindsight {
+            proxy_pass http://hindsight_cp;
+            proxy_set_header Host $http_host;
+        }
+
+        location = / { return 301 /hindsight; }
+    }
+}
+```
+
+### Why Build is Required
+
+Next.js requires `basePath` at **build time**. The published image was built without a custom base path, so you must rebuild from source with the `NEXT_PUBLIC_BASE_PATH` build arg to deploy the Control Plane under a subpath.
+
+The API works without rebuild because `HINDSIGHT_API_BASE_PATH` is a runtime environment variable.
@@ -0,0 +1,88 @@
+# Hindsight API deployment with Nginx reverse proxy (API-only)
+#
+# This example deploys Hindsight API under the path /hindsight with:
+# - Hindsight standalone image (API + Control Plane + embedded pg0)
+# - Nginx reverse proxy (API only)
+#
+# Quick Start:
+#   docker-compose -f docker/docker-compose/nginx/docker-compose.yml up
+#
+# Access:
+#   API (via nginx):         http://localhost:8080/hindsight/docs
+#   Control Plane (direct):  http://localhost:9999
+#
+# For full stack deployment (API + Control Plane both under /hindsight):
+#   See README.md in this directory for instructions on building with basePath.
+#
+# Note: This configuration uses the published image (no build required).
+#       Control Plane is served directly because Next.js basePath requires
+#       build-time configuration. See README.md for the full stack option.
+
+services:
+  # Hindsight (API + Control Plane + embedded pg0)
+  hindsight:
+    image: ghcr.io/vectorize-io/hindsight:latest
+    ports:
+      - "9999:9999"  # Control Plane (direct access, not proxied)
+    environment:
+      # API base path for reverse proxy
+      HINDSIGHT_API_BASE_PATH: /hindsight
+
+      # LLM configuration
+      # Using mock provider for testing (no API key needed)
+      # For production, set OPENAI_API_KEY or ANTHROPIC_API_KEY and use a real provider
+      HINDSIGHT_API_LLM_PROVIDER: ${HINDSIGHT_API_LLM_PROVIDER:-mock}
+      HINDSIGHT_API_LLM_API_KEY: ${OPENAI_API_KEY:-not-needed-for-mock}
+      HINDSIGHT_API_LLM_MODEL: ${HINDSIGHT_API_LLM_MODEL:-mock-model}
+
+      # Production examples (uncomment and set appropriate API key):
+      # HINDSIGHT_API_LLM_PROVIDER: openai
+      # HINDSIGHT_API_LLM_API_KEY: ${OPENAI_API_KEY}
+      # HINDSIGHT_API_LLM_MODEL: gpt-4o-mini
+
+      # HINDSIGHT_API_LLM_PROVIDER: anthropic
+      # HINDSIGHT_API_LLM_API_KEY: ${ANTHROPIC_API_KEY}
+      # HINDSIGHT_API_LLM_MODEL: claude-sonnet-4-20250514
+
+      # Server config
+      HINDSIGHT_API_HOST: 0.0.0.0
+      HINDSIGHT_API_PORT: 8888
+      HINDSIGHT_API_LOG_LEVEL: info
+
+      # Control Plane config
+      HINDSIGHT_CP_DATAPLANE_API_URL: http://localhost:8888
+    volumes:
+      # Persist embedded pg0 database
+      - hindsight_data:/app/data
+    # Note: Ports not exposed - access via Nginx at localhost:8080/hindsight/
+    # To debug directly, uncomment these ports:
+    # ports:
+    #   - "8888:8888"  # API
+    #   - "9999:9999"  # Control Plane
+    healthcheck:
+      test: ["CMD", "curl", "-f", "http://localhost:8888/hindsight/health"]
+      interval: 10s
+      timeout: 5s
+      retries: 3
+      start_period: 30s
+    networks:
+      - hindsight
+
+  # Nginx reverse proxy
+  nginx:
+    image: nginx:alpine
+    ports:
+      - "8080:80"
+    volumes:
+      - ./nginx.conf:/etc/nginx/nginx.conf:ro
+    depends_on:
+      hindsight:
+        condition: service_healthy
+    networks:
+      - hindsight
+
+volumes:
+  hindsight_data:
+
+networks:
+  hindsight:
@@ -0,0 +1,40 @@
+# Nginx configuration for API-only reverse proxy
+# Control Plane accessed directly (not through nginx)
+
+events {
+    worker_connections 1024;
+}
+
+http {
+    include /etc/nginx/mime.types;
+    default_type application/octet-stream;
+
+    # Logging
+    access_log /var/log/nginx/access.log;
+    error_log /var/log/nginx/error.log;
+
+    # Upstream - Hindsight API
+    upstream hindsight_api {
+        server hindsight:8888;
+    }
+
+    server {
+        listen 80;
+        server_name _;
+
+        # API endpoints - forward with /hindsight prefix
+        location /hindsight/ {
+            proxy_pass http://hindsight_api;
+
+            proxy_set_header Host $http_host;
+            proxy_set_header X-Real-IP $remote_addr;
+            proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;
+            proxy_set_header X-Forwarded-Proto $scheme;
+        }
+
+        # Redirect root to API docs
+        location = / {
+            return 301 /hindsight/docs;
+        }
+    }
+}
@@ -0,0 +1,93 @@
+name: hindsight
+# Docker Compose file for Hindsight with PostgreSQL and vectorchord
+# docker compose -f docker/docker-compose/docker-compose.yaml down && sleep 2 && docker compose -f docker/docker-compose/docker-compose.yaml up -d
+# Make sure to set the required environment variables before running:
+# - HINDSIGHT_DB_PASSWORD: Password for the PostgreSQL user
+# - Configure LLM provider variables as needed (see below in the hindsight service)
+#
+# Usage:
+# docker compose up -d
+#
+# Optional environment variables with defaults:
+# - HINDSIGHT_VERSION: Hindsight application version (default: latest)
+# - HINDSIGHT_DB_USER: PostgreSQL user (default: hindsight_user)
+# - HINDSIGHT_DB_NAME: PostgreSQL database name (default: hindsight_db)
+# - HINDSIGHT_DB_VERSION: PostgreSQL version (default: 18)
+
+services:
+  db:
+    # Use a PostgreSQL-Image with vectorchord extension pre-installed
+    image: tensorchord/vchord-suite:pg${HINDSIGHT_DB_VERSION:-18-latest}
+    container_name: hindsight-db
+    restart: always
+    # Expose PostgreSQL port
+    ports:
+      - "5436:5432"
+    environment:
+      POSTGRES_USER: ${HINDSIGHT_DB_USER:-hindsight_user}
+      POSTGRES_PASSWORD: ${HINDSIGHT_DB_PASSWORD:-hindsight_password}
+      POSTGRES_DB: ${HINDSIGHT_DB_NAME:-hindsight_db}
+    volumes:
+      - pg_data:/var/lib/postgresql/${HINDSIGHT_DB_VERSION:-18}/docker
+    networks:
+      - hindsight-net
+
+  vectorchord-init:
+    image: tensorchord/vchord-suite:pg18-latest
+    #container_name: vectorchord-init
+    depends_on:
+      - db
+    environment:
+      - PGPASSWORD=${HINDSIGHT_DB_PASSWORD:-hindsight_password}
+    command: >
+      bash -c "
+        echo 'Waiting for PostgreSQL to be ready...';
+        until pg_isready -h hindsight-db -p 5432 -U hindsight_user; do
+          echo 'PostgreSQL is unavailable - sleeping';
+          sleep 2;
+        done;
+        echo 'PostgreSQL is ready - creating hindsight_db database';
+        psql -h hindsight-db -p 5432 -U hindsight_user -c 'CREATE DATABASE hindsight_db;' 2>/dev/null || echo 'Database already exists';
+        echo 'Creating extensions in hindsight_db database';
+        psql -h hindsight-db -p 5432 -U hindsight_user -d hindsight_db -c 'CREATE EXTENSION IF NOT EXISTS vchord CASCADE;';
+        psql -h hindsight-db -p 5432 -U hindsight_user -d hindsight_db -c 'CREATE EXTENSION IF NOT EXISTS pg_tokenizer CASCADE;';
+        psql -h hindsight-db -p 5432 -U hindsight_user -d hindsight_db -c 'CREATE EXTENSION IF NOT EXISTS vchord_bm25 CASCADE;';
+        echo 'Creating llmlingua2 tokenizer';
+        psql -h hindsight-db -p 5432 -U hindsight_user -d hindsight_db -c \"SELECT create_tokenizer('llmlingua2', \\$\\$ model = \\\"llmlingua2\\\" \\$\\$);\" 2>/dev/null || echo 'Tokenizer already exists or creation skipped';
+        echo 'Database and extensions created successfully';
+      "
+    restart: "no"
+    networks:
+      - hindsight-net
+
+  hindsight:
+    image: ghcr.io/vectorize-io/hindsight:${HINDSIGHT_VERSION:-latest}
+    container_name: hindsight-app
+    ports:
+      - "8888:8888"
+      - "9999:9999"
+    environment:
+      # LLM Configuration (uses OpenAI for testing vchord)
+      # LLM configuration
+      HINDSIGHT_API_LLM_PROVIDER: ${HINDSIGHT_API_LLM_PROVIDER:-openai}
+      HINDSIGHT_API_LLM_API_KEY: ${OPENAI_API_KEY:-your-api-key}
+
+      # Database Configuration
+      HINDSIGHT_API_DATABASE_URL: postgresql://${HINDSIGHT_DB_USER:-hindsight_user}:${HINDSIGHT_DB_PASSWORD:-hindsight_password}@db:5432/${HINDSIGHT_DB_NAME:-hindsight_db}
+
+      # Vector and Text Search Extensions
+      HINDSIGHT_API_VECTOR_EXTENSION: vchord
+      HINDSIGHT_API_TEXT_SEARCH_EXTENSION: vchord
+
+    depends_on:
+      - db
+    networks:
+      - hindsight-net
+
+
+networks:
+  hindsight-net:
+    driver: bridge
+
+volumes:
+  pg_data:
@@ -8,6 +8,7 @@
 #                                     Set to false when using external providers (TEI, OpenAI, Cohere)
 #   PRELOAD_ML_MODELS=true/false   - Pre-download ML models during build (default: true)
 #                                     Only effective when INCLUDE_LOCAL_MODELS=true
+#                                     NOTE: tiktoken encodings are ALWAYS preloaded (required for air-gapped deployments)
 #
 # Examples:
 #   docker build -t hindsight .                                          # Both (standalone)
@@ -111,6 +112,10 @@ RUN rm -f package-lock.json && sed -i '/"@vectorize-io\/hindsight-client":/d' pa
 # Copy built SDK directly into node_modules (more reliable than npm link in Docker)
 COPY --from=sdk-builder /app/hindsight-clients/typescript ./node_modules/@vectorize-io/hindsight-client

+# Accept base path as build argument for reverse proxy deployments
+# Usage: docker build --build-arg NEXT_PUBLIC_BASE_PATH=/hindsight ...
+ARG NEXT_PUBLIC_BASE_PATH=""
+
 # Build Control Plane - run next build first, then custom standalone copy
 # (The build:standalone script expects a specific path structure that differs in Docker)
 RUN npm exec -- next build
@@ -167,6 +172,28 @@ USER hindsight

 ENV PATH="/app/api/.venv/bin:${PATH}"

+# Pre-download tiktoken encoding (ALWAYS - required for token counting even in air-gapped envs)
+# Tiktoken is a core runtime dependency, not an optional ML model
+RUN MAX_RETRIES=3; \
+    RETRY_DELAY=5; \
+    for i in $(seq 1 $MAX_RETRIES); do \
+      echo "Attempt $i/$MAX_RETRIES: Downloading tiktoken encoding..."; \
+      /app/api/.venv/bin/python -c "\
+import tiktoken; \
+print('Downloading cl100k_base encoding...'); \
+tiktoken.get_encoding('cl100k_base'); \
+print('Tiktoken encoding cached successfully')" && break; \
+      if [ $i -lt $MAX_RETRIES ]; then \
+        echo "Attempt $i failed, retrying in ${RETRY_DELAY}s..."; \
+        sleep $RETRY_DELAY; \
+        RETRY_DELAY=$((RETRY_DELAY * 2)); \
+      fi; \
+    done; \
+    if [ $i -eq $MAX_RETRIES ]; then \
+      echo "ERROR: Failed to download tiktoken encoding after $MAX_RETRIES attempts"; \
+      exit 1; \
+    fi
+
 # Pre-download ML models to avoid runtime download (conditional)
 # Only runs if both PRELOAD_ML_MODELS=true AND INCLUDE_LOCAL_MODELS=true
 # Includes retry logic with exponential backoff for transient network failures
@@ -185,7 +212,6 @@ print('Downloading embedding model...'); \
 SentenceTransformer('BAAI/bge-small-en-v1.5'); \
 print('Downloading cross-encoder model...'); \
 CrossEncoder('cross-encoder/ms-marco-MiniLM-L-6-v2'); \
-print('Downloading tiktoken encoding...'); import tiktoken; tiktoken.get_encoding('cl100k_base'); \
 print('Models cached successfully')" && break; \
      if [ $i -lt $MAX_RETRIES ]; then \
        echo "Attempt $i failed, retrying in ${RETRY_DELAY}s..."; \
@@ -208,6 +234,10 @@ ENV HINDSIGHT_API_LOG_LEVEL=info
 ENV HINDSIGHT_ENABLE_API=true
 ENV HINDSIGHT_ENABLE_CP=false
 ENV PYTHONUNBUFFERED=1
+# Suppress verbose transformers/HuggingFace model loading warnings
+ENV TRANSFORMERS_VERBOSITY=error
+ENV HF_HUB_VERBOSITY=error
+ENV TOKENIZERS_PARALLELISM=false

 CMD ["/app/start-all.sh"]

@@ -293,6 +323,28 @@ USER hindsight

 ENV PATH="/app/api/.venv/bin:${PATH}"

+# Pre-download tiktoken encoding (ALWAYS - required for token counting even in air-gapped envs)
+# Tiktoken is a core runtime dependency, not an optional ML model
+RUN MAX_RETRIES=3; \
+    RETRY_DELAY=5; \
+    for i in $(seq 1 $MAX_RETRIES); do \
+      echo "Attempt $i/$MAX_RETRIES: Downloading tiktoken encoding..."; \
+      /app/api/.venv/bin/python -c "\
+import tiktoken; \
+print('Downloading cl100k_base encoding...'); \
+tiktoken.get_encoding('cl100k_base'); \
+print('Tiktoken encoding cached successfully')" && break; \
+      if [ $i -lt $MAX_RETRIES ]; then \
+        echo "Attempt $i failed, retrying in ${RETRY_DELAY}s..."; \
+        sleep $RETRY_DELAY; \
+        RETRY_DELAY=$((RETRY_DELAY * 2)); \
+      fi; \
+    done; \
+    if [ $i -eq $MAX_RETRIES ]; then \
+      echo "ERROR: Failed to download tiktoken encoding after $MAX_RETRIES attempts"; \
+      exit 1; \
+    fi
+
 # Pre-download ML models to avoid runtime download (conditional)
 # Only runs if both PRELOAD_ML_MODELS=true AND INCLUDE_LOCAL_MODELS=true
 # Includes retry logic with exponential backoff for transient network failures
@@ -311,7 +363,6 @@ print('Downloading embedding model...'); \
 SentenceTransformer('BAAI/bge-small-en-v1.5'); \
 print('Downloading cross-encoder model...'); \
 CrossEncoder('cross-encoder/ms-marco-MiniLM-L-6-v2'); \
-print('Downloading tiktoken encoding...'); import tiktoken; tiktoken.get_encoding('cl100k_base'); \
 print('Models cached successfully')" && break; \
      if [ $i -lt $MAX_RETRIES ]; then \
        echo "Attempt $i failed, retrying in ${RETRY_DELAY}s..."; \
@@ -336,6 +387,10 @@ ENV HINDSIGHT_CP_DATAPLANE_API_URL=http://localhost:8888
 ENV HINDSIGHT_ENABLE_API=true
 ENV HINDSIGHT_ENABLE_CP=true
 ENV PYTHONUNBUFFERED=1
+# Suppress verbose transformers/HuggingFace model loading warnings
+ENV TRANSFORMERS_VERBOSITY=error
+ENV HF_HUB_VERBOSITY=error
+ENV TOKENIZERS_PARALLELISM=false

 CMD ["/app/start-all.sh"]

@@ -6,28 +6,40 @@
 # Can be run locally or in CI pipelines.
 #
 # Usage:
-#   ./scripts/docker-smoke-test.sh <image> [target]
+#   ./docker/test-image.sh <image> [target]
 #
 # Arguments:
 #   image   - Docker image to test (e.g., hindsight-api:test, ghcr.io/vectorize-io/hindsight:latest)
 #   target  - Optional: 'cp-only' for control plane, otherwise assumes API image (default: api)
 #
 # Environment variables:
-#   GROQ_API_KEY              - Required for API/standalone images (LLM verification)
-#   HINDSIGHT_API_LLM_PROVIDER - LLM provider (default: groq)
-#   HINDSIGHT_API_LLM_MODEL    - LLM model (default: llama-3.3-70b-versatile)
-#   SMOKE_TEST_TIMEOUT        - Timeout in seconds (default: 120)
-#   SMOKE_TEST_CONTAINER_NAME - Container name (default: hindsight-smoke-test)
+#   GROQ_API_KEY                                - Required for API/standalone images (LLM verification)
+#   HINDSIGHT_API_LLM_PROVIDER                  - LLM provider (default: groq)
+#   HINDSIGHT_API_LLM_MODEL                     - LLM model (default: llama-3.3-70b-versatile)
+#   HINDSIGHT_API_EMBEDDINGS_PROVIDER           - Embeddings provider (optional, for slim images: openai, cohere, tei)
+#   HINDSIGHT_API_EMBEDDINGS_OPENAI_API_KEY     - OpenAI API key for embeddings (optional)
+#   HINDSIGHT_API_RERANKER_PROVIDER             - Reranker provider (optional, for slim images: cohere, tei)
+#   HINDSIGHT_API_COHERE_API_KEY                - Cohere API key for reranking (optional)
+#   SMOKE_TEST_TIMEOUT                          - Timeout in seconds (default: 120)
+#   SMOKE_TEST_CONTAINER_NAME                   - Container name (default: hindsight-smoke-test)
 #
 # Examples:
-#   # Test a locally built image
-#   ./scripts/docker-smoke-test.sh hindsight-api:test
+#   # Test a locally built full image
+#   ./docker/test-image.sh hindsight-api:test
 #
 #   # Test a released image
-#   ./scripts/docker-smoke-test.sh ghcr.io/vectorize-io/hindsight:latest
+#   ./docker/test-image.sh ghcr.io/vectorize-io/hindsight:latest
 #
 #   # Test control plane image
-#   ./scripts/docker-smoke-test.sh hindsight-control-plane:test cp-only
+#   ./docker/test-image.sh hindsight-control-plane:test cp-only
+#
+#   # Test slim image with external providers
+#   export GROQ_API_KEY=gsk_xxx
+#   export HINDSIGHT_API_EMBEDDINGS_PROVIDER=openai
+#   export HINDSIGHT_API_EMBEDDINGS_OPENAI_API_KEY=sk-xxx
+#   export HINDSIGHT_API_RERANKER_PROVIDER=cohere
+#   export HINDSIGHT_API_COHERE_API_KEY=xxx
+#   ./docker/test-image.sh hindsight-slim:test
 #
 # Exit codes:
 #   0 - Success (container healthy)
@@ -108,12 +120,32 @@ if [ "$TARGET" = "cp-only" ]; then
        -p "${HEALTH_PORT}:${HEALTH_PORT}" \
        "$IMAGE"
 else
-    docker run -d --name "$CONTAINER_NAME" \
-        -e HINDSIGHT_API_LLM_PROVIDER="$LLM_PROVIDER" \
-        -e HINDSIGHT_API_LLM_API_KEY="${GROQ_API_KEY}" \
-        -e HINDSIGHT_API_LLM_MODEL="$LLM_MODEL" \
-        -p "${HEALTH_PORT}:${HEALTH_PORT}" \
-        "$IMAGE"
+    # Build docker run command with required and optional env vars
+    DOCKER_CMD="docker run -d --name $CONTAINER_NAME"
+    DOCKER_CMD="$DOCKER_CMD -e HINDSIGHT_API_LLM_PROVIDER=$LLM_PROVIDER"
+    DOCKER_CMD="$DOCKER_CMD -e HINDSIGHT_API_LLM_API_KEY=${GROQ_API_KEY}"
+    DOCKER_CMD="$DOCKER_CMD -e HINDSIGHT_API_LLM_MODEL=$LLM_MODEL"
+
+    # Add optional embeddings provider config
+    if [ -n "${HINDSIGHT_API_EMBEDDINGS_PROVIDER:-}" ]; then
+        DOCKER_CMD="$DOCKER_CMD -e HINDSIGHT_API_EMBEDDINGS_PROVIDER=${HINDSIGHT_API_EMBEDDINGS_PROVIDER}"
+    fi
+    if [ -n "${HINDSIGHT_API_EMBEDDINGS_OPENAI_API_KEY:-}" ]; then
+        DOCKER_CMD="$DOCKER_CMD -e HINDSIGHT_API_EMBEDDINGS_OPENAI_API_KEY=${HINDSIGHT_API_EMBEDDINGS_OPENAI_API_KEY}"
+    fi
+
+    # Add optional reranker provider config
+    if [ -n "${HINDSIGHT_API_RERANKER_PROVIDER:-}" ]; then
+        DOCKER_CMD="$DOCKER_CMD -e HINDSIGHT_API_RERANKER_PROVIDER=${HINDSIGHT_API_RERANKER_PROVIDER}"
+    fi
+    if [ -n "${HINDSIGHT_API_COHERE_API_KEY:-}" ]; then
+        DOCKER_CMD="$DOCKER_CMD -e HINDSIGHT_API_COHERE_API_KEY=${HINDSIGHT_API_COHERE_API_KEY}"
+    fi
+
+    DOCKER_CMD="$DOCKER_CMD -p ${HEALTH_PORT}:${HEALTH_PORT}"
+    DOCKER_CMD="$DOCKER_CMD $IMAGE"
+
+    eval $DOCKER_CMD
 fi

 # Wait for health endpoint
@@ -0,0 +1,51 @@
+#!/bin/bash
+#
+# Local Test Script for Slim Docker Images
+#
+# This script makes it easy to test slim images locally with external providers.
+# It expects API keys to be set in environment variables.
+#
+# Usage:
+#   export GROQ_API_KEY=gsk_xxx
+#   export OPENAI_API_KEY=sk-xxx
+#   export COHERE_API_KEY=xxx
+#   ./docker/test-slim-local.sh
+#
+# Or inline:
+#   GROQ_API_KEY=gsk_xxx OPENAI_API_KEY=sk_xxx COHERE_API_KEY=xxx ./docker/test-slim-local.sh
+#
+
+set -euo pipefail
+
+# Check for required API keys
+if [ -z "${GROQ_API_KEY:-}" ]; then
+    echo "❌ Error: GROQ_API_KEY environment variable is required"
+    echo "Set it with: export GROQ_API_KEY=gsk_xxx"
+    exit 1
+fi
+
+if [ -z "${OPENAI_API_KEY:-}" ]; then
+    echo "❌ Error: OPENAI_API_KEY environment variable is required"
+    echo "Set it with: export OPENAI_API_KEY=sk-xxx"
+    exit 1
+fi
+
+if [ -z "${COHERE_API_KEY:-}" ]; then
+    echo "❌ Error: COHERE_API_KEY environment variable is required"
+    echo "Set it with: export COHERE_API_KEY=xxx"
+    exit 1
+fi
+
+# Configuration
+IMAGE="${1:-hindsight-slim:test}"
+echo "Testing image: $IMAGE"
+echo ""
+
+# Set up external providers
+export HINDSIGHT_API_EMBEDDINGS_PROVIDER=openai
+export HINDSIGHT_API_EMBEDDINGS_OPENAI_API_KEY=$OPENAI_API_KEY
+export HINDSIGHT_API_RERANKER_PROVIDER=cohere
+export HINDSIGHT_API_COHERE_API_KEY=$COHERE_API_KEY
+
+# Run the test
+exec "$(dirname "$0")/test-image.sh" "$IMAGE" standalone
@@ -2,8 +2,8 @@ apiVersion: v2
 name: hindsight
 description: Hindsight helm chart
 type: application
-version: 0.4.6
-appVersion: "0.4.6"
+version: 0.4.10
+appVersion: "0.4.10"
 keywords:
  - ai
  - memory
@@ -127,6 +127,38 @@ API URL for control plane
 {{- printf "http://%s-api:%d" (include "hindsight.fullname" .) (.Values.api.service.port | int) }}
 {{- end }}

+{{/*
+TEI reranker labels
+*/}}
+{{- define "hindsight.tei.reranker.labels" -}}
+{{ include "hindsight.labels" . }}
+app.kubernetes.io/component: tei-reranker
+{{- end }}
+
+{{/*
+TEI reranker selector labels
+*/}}
+{{- define "hindsight.tei.reranker.selectorLabels" -}}
+{{ include "hindsight.selectorLabels" . }}
+app.kubernetes.io/component: tei-reranker
+{{- end }}
+
+{{/*
+TEI embedding labels
+*/}}
+{{- define "hindsight.tei.embedding.labels" -}}
+{{ include "hindsight.labels" . }}
+app.kubernetes.io/component: tei-embedding
+{{- end }}
+
+{{/*
+TEI embedding selector labels
+*/}}
+{{- define "hindsight.tei.embedding.selectorLabels" -}}
+{{ include "hindsight.selectorLabels" . }}
+app.kubernetes.io/component: tei-embedding
+{{- end }}
+
 {{/*
 Get the name of the secret to use
 */}}
@@ -33,7 +33,7 @@ spec:
      - name: api
        securityContext:
          {{- toYaml .Values.securityContext | nindent 10 }}
-        image: "{{ .Values.api.image.repository }}:{{ .Values.api.image.tag | default .Values.version }}"
+        image: "{{ .Values.api.image.repository }}:{{ .Values.api.image.tag | default .Values.version | default .Chart.AppVersion }}"
        imagePullPolicy: {{ .Values.api.image.pullPolicy }}
        ports:
        - name: http
@@ -60,10 +60,25 @@ spec:
        - name: HINDSIGHT_API_WORKER_ENABLED
          value: "false"
        {{- end }}
+        {{- /* Explicitly set port to override K8s service discovery env var (HINDSIGHT_API_PORT) */}}
+        - name: HINDSIGHT_API_PORT
+          value: {{ .Values.api.service.targetPort | quote }}
        {{- range $key, $value := .Values.api.env }}
        - name: {{ $key }}
          value: {{ $value | quote }}
        {{- end }}
+        {{- if .Values.tei.reranker.enabled }}
+        - name: HINDSIGHT_API_RERANKER_PROVIDER
+          value: "tei"
+        - name: HINDSIGHT_API_RERANKER_TEI_URL
+          value: "http://{{ include "hindsight.fullname" . }}-tei-reranker:{{ .Values.tei.reranker.port }}"
+        {{- end }}
+        {{- if .Values.tei.embedding.enabled }}
+        - name: HINDSIGHT_API_EMBEDDINGS_PROVIDER
+          value: "tei"
+        - name: HINDSIGHT_API_EMBEDDINGS_TEI_URL
+          value: "http://{{ include "hindsight.fullname" . }}-tei-embedding:{{ .Values.tei.embedding.port }}"
+        {{- end }}
        {{- /* Only use api.secrets when not using existingSecret (for chart-managed secrets) */}}
        {{- if not .Values.existingSecret }}
        {{- range $key, $value := .Values.api.secrets }}
@@ -84,7 +99,7 @@ spec:
      nodeSelector:
        {{- toYaml . | nindent 8 }}
      {{- end }}
-      {{- with .Values.affinity }}
+      {{- with (.Values.api.affinity | default .Values.affinity) }}
      affinity:
        {{- toYaml . | nindent 8 }}
      {{- end }}
@@ -33,7 +33,7 @@ spec:
      - name: control-plane
        securityContext:
          {{- toYaml .Values.securityContext | nindent 10 }}
-        image: "{{ .Values.controlPlane.image.repository }}:{{ .Values.controlPlane.image.tag | default .Values.version }}"
+        image: "{{ .Values.controlPlane.image.repository }}:{{ .Values.controlPlane.image.tag | default .Values.version | default .Chart.AppVersion }}"
        imagePullPolicy: {{ .Values.controlPlane.image.pullPolicy }}
        ports:
        - name: http
@@ -71,7 +71,7 @@ spec:
      nodeSelector:
        {{- toYaml . | nindent 8 }}
      {{- end }}
-      {{- with .Values.affinity }}
+      {{- with (.Values.controlPlane.affinity | default .Values.affinity) }}
      affinity:
        {{- toYaml . | nindent 8 }}
      {{- end }}
@@ -0,0 +1,56 @@
+{{- if and .Values.api.enabled .Values.api.podDisruptionBudget.enabled }}
+apiVersion: policy/v1
+kind: PodDisruptionBudget
+metadata:
+  name: {{ include "hindsight.fullname" . }}-api
+  labels:
+    {{- include "hindsight.api.labels" . | nindent 4 }}
+spec:
+  {{- if .Values.api.podDisruptionBudget.minAvailable }}
+  minAvailable: {{ .Values.api.podDisruptionBudget.minAvailable }}
+  {{- end }}
+  {{- if .Values.api.podDisruptionBudget.maxUnavailable }}
+  maxUnavailable: {{ .Values.api.podDisruptionBudget.maxUnavailable }}
+  {{- end }}
+  selector:
+    matchLabels:
+      {{- include "hindsight.api.selectorLabels" . | nindent 6 }}
+{{- end }}
+---
+{{- if and .Values.controlPlane.enabled .Values.controlPlane.podDisruptionBudget.enabled }}
+apiVersion: policy/v1
+kind: PodDisruptionBudget
+metadata:
+  name: {{ include "hindsight.fullname" . }}-control-plane
+  labels:
+    {{- include "hindsight.controlPlane.labels" . | nindent 4 }}
+spec:
+  {{- if .Values.controlPlane.podDisruptionBudget.minAvailable }}
+  minAvailable: {{ .Values.controlPlane.podDisruptionBudget.minAvailable }}
+  {{- end }}
+  {{- if .Values.controlPlane.podDisruptionBudget.maxUnavailable }}
+  maxUnavailable: {{ .Values.controlPlane.podDisruptionBudget.maxUnavailable }}
+  {{- end }}
+  selector:
+    matchLabels:
+      {{- include "hindsight.controlPlane.selectorLabels" . | nindent 6 }}
+{{- end }}
+---
+{{- if and .Values.worker.enabled .Values.worker.podDisruptionBudget.enabled }}
+apiVersion: policy/v1
+kind: PodDisruptionBudget
+metadata:
+  name: {{ include "hindsight.fullname" . }}-worker
+  labels:
+    {{- include "hindsight.worker.labels" . | nindent 4 }}
+spec:
+  {{- if .Values.worker.podDisruptionBudget.minAvailable }}
+  minAvailable: {{ .Values.worker.podDisruptionBudget.minAvailable }}
+  {{- end }}
+  {{- if .Values.worker.podDisruptionBudget.maxUnavailable }}
+  maxUnavailable: {{ .Values.worker.podDisruptionBudget.maxUnavailable }}
+  {{- end }}
+  selector:
+    matchLabels:
+      {{- include "hindsight.worker.selectorLabels" . | nindent 6 }}
+{{- end }}
@@ -0,0 +1,76 @@
+{{- if .Values.tei.embedding.enabled }}
+apiVersion: apps/v1
+kind: Deployment
+metadata:
+  name: {{ include "hindsight.fullname" . }}-tei-embedding
+  labels:
+    {{- include "hindsight.tei.embedding.labels" . | nindent 4 }}
+spec:
+  replicas: {{ .Values.tei.embedding.replicaCount }}
+  selector:
+    matchLabels:
+      {{- include "hindsight.tei.embedding.selectorLabels" . | nindent 6 }}
+  template:
+    metadata:
+      {{- with .Values.podAnnotations }}
+      annotations:
+        {{- toYaml . | nindent 8 }}
+      {{- end }}
+      labels:
+        {{- include "hindsight.tei.embedding.selectorLabels" . | nindent 8 }}
+    spec:
+      {{- if .Values.serviceAccount.create }}
+      serviceAccountName: {{ include "hindsight.serviceAccountName" . }}
+      {{- end }}
+      securityContext:
+        {{- toYaml .Values.podSecurityContext | nindent 8 }}
+      containers:
+      - name: tei-embedding
+        securityContext:
+          {{- toYaml .Values.securityContext | nindent 10 }}
+        image: "{{ .Values.tei.embedding.image.repository }}:{{ .Values.tei.embedding.image.tag }}"
+        imagePullPolicy: {{ .Values.tei.embedding.image.pullPolicy }}
+        args:
+          - "--model-id"
+          - {{ .Values.tei.embedding.model | quote }}
+          - "--hostname"
+          - "0.0.0.0"
+          {{- range .Values.tei.embedding.args }}
+          - {{ . | quote }}
+          {{- end }}
+        ports:
+        - name: http
+          containerPort: {{ .Values.tei.embedding.port }}
+          protocol: TCP
+        env:
+        - name: PORT
+          value: {{ .Values.tei.embedding.port | quote }}
+        {{- range $key, $value := .Values.tei.embedding.env }}
+        - name: {{ $key }}
+          value: {{ $value | quote }}
+        {{- end }}
+        livenessProbe:
+          {{- toYaml .Values.tei.embedding.livenessProbe | nindent 10 }}
+        readinessProbe:
+          {{- toYaml .Values.tei.embedding.readinessProbe | nindent 10 }}
+        resources:
+          {{- toYaml .Values.tei.embedding.resources | nindent 10 }}
+        volumeMounts:
+        - name: model-cache
+          mountPath: /data
+      volumes:
+      - name: model-cache
+        emptyDir: {}
+      {{- with .Values.nodeSelector }}
+      nodeSelector:
+        {{- toYaml . | nindent 8 }}
+      {{- end }}
+      {{- with .Values.affinity }}
+      affinity:
+        {{- toYaml . | nindent 8 }}
+      {{- end }}
+      {{- with .Values.tolerations }}
+      tolerations:
+        {{- toYaml . | nindent 8 }}
+      {{- end }}
+{{- end }}
@@ -0,0 +1,17 @@
+{{- if .Values.tei.embedding.enabled }}
+apiVersion: v1
+kind: Service
+metadata:
+  name: {{ include "hindsight.fullname" . }}-tei-embedding
+  labels:
+    {{- include "hindsight.tei.embedding.labels" . | nindent 4 }}
+spec:
+  type: ClusterIP
+  ports:
+  - port: {{ .Values.tei.embedding.port }}
+    targetPort: http
+    protocol: TCP
+    name: http
+  selector:
+    {{- include "hindsight.tei.embedding.selectorLabels" . | nindent 4 }}
+{{- end }}
@@ -0,0 +1,76 @@
+{{- if .Values.tei.reranker.enabled }}
+apiVersion: apps/v1
+kind: Deployment
+metadata:
+  name: {{ include "hindsight.fullname" . }}-tei-reranker
+  labels:
+    {{- include "hindsight.tei.reranker.labels" . | nindent 4 }}
+spec:
+  replicas: {{ .Values.tei.reranker.replicaCount }}
+  selector:
+    matchLabels:
+      {{- include "hindsight.tei.reranker.selectorLabels" . | nindent 6 }}
+  template:
+    metadata:
+      {{- with .Values.podAnnotations }}
+      annotations:
+        {{- toYaml . | nindent 8 }}
+      {{- end }}
+      labels:
+        {{- include "hindsight.tei.reranker.selectorLabels" . | nindent 8 }}
+    spec:
+      {{- if .Values.serviceAccount.create }}
+      serviceAccountName: {{ include "hindsight.serviceAccountName" . }}
+      {{- end }}
+      securityContext:
+        {{- toYaml .Values.podSecurityContext | nindent 8 }}
+      containers:
+      - name: tei-reranker
+        securityContext:
+          {{- toYaml .Values.securityContext | nindent 10 }}
+        image: "{{ .Values.tei.reranker.image.repository }}:{{ .Values.tei.reranker.image.tag }}"
+        imagePullPolicy: {{ .Values.tei.reranker.image.pullPolicy }}
+        args:
+          - "--model-id"
+          - {{ .Values.tei.reranker.model | quote }}
+          - "--hostname"
+          - "0.0.0.0"
+          {{- range .Values.tei.reranker.args }}
+          - {{ . | quote }}
+          {{- end }}
+        ports:
+        - name: http
+          containerPort: {{ .Values.tei.reranker.port }}
+          protocol: TCP
+        env:
+        - name: PORT
+          value: {{ .Values.tei.reranker.port | quote }}
+        {{- range $key, $value := .Values.tei.reranker.env }}
+        - name: {{ $key }}
+          value: {{ $value | quote }}
+        {{- end }}
+        livenessProbe:
+          {{- toYaml .Values.tei.reranker.livenessProbe | nindent 10 }}
+        readinessProbe:
+          {{- toYaml .Values.tei.reranker.readinessProbe | nindent 10 }}
+        resources:
+          {{- toYaml .Values.tei.reranker.resources | nindent 10 }}
+        volumeMounts:
+        - name: model-cache
+          mountPath: /data
+      volumes:
+      - name: model-cache
+        emptyDir: {}
+      {{- with .Values.nodeSelector }}
+      nodeSelector:
+        {{- toYaml . | nindent 8 }}
+      {{- end }}
+      {{- with .Values.affinity }}
+      affinity:
+        {{- toYaml . | nindent 8 }}
+      {{- end }}
+      {{- with .Values.tolerations }}
+      tolerations:
+        {{- toYaml . | nindent 8 }}
+      {{- end }}
+{{- end }}
@@ -0,0 +1,17 @@
+{{- if .Values.tei.reranker.enabled }}
+apiVersion: v1
+kind: Service
+metadata:
+  name: {{ include "hindsight.fullname" . }}-tei-reranker
+  labels:
+    {{- include "hindsight.tei.reranker.labels" . | nindent 4 }}
+spec:
+  type: ClusterIP
+  ports:
+  - port: {{ .Values.tei.reranker.port }}
+    targetPort: http
+    protocol: TCP
+    name: http
+  selector:
+    {{- include "hindsight.tei.reranker.selectorLabels" . | nindent 4 }}
+{{- end }}
@@ -32,7 +32,7 @@ spec:
      - name: worker
        securityContext:
          {{- toYaml .Values.securityContext | nindent 10 }}
-        image: "{{ .Values.worker.image.repository }}:{{ .Values.worker.image.tag | default .Values.version }}"
+        image: "{{ .Values.worker.image.repository }}:{{ .Values.worker.image.tag | default .Values.version | default .Chart.AppVersion }}"
        imagePullPolicy: {{ .Values.worker.image.pullPolicy }}
        command: ["hindsight-worker"]
        ports:
@@ -99,7 +99,7 @@ spec:
      nodeSelector:
        {{- toYaml . | nindent 8 }}
      {{- end }}
-      {{- with .Values.affinity }}
+      {{- with (.Values.worker.affinity | default .Values.affinity) }}
      affinity:
        {{- toYaml . | nindent 8 }}
      {{- end }}
@@ -1,7 +1,8 @@
 # Default values for hindsight

-# Chart version - use this to set a consistent image tag across all components
-version: "0.1.1"
+# Global version override - use this to set a consistent image tag across all components
+# If not set, defaults to Chart.appVersion from Chart.yaml
+# version: ""

 # Use an existing secret instead of creating one from values
 # When set, all keys from this secret are injected as environment variables via envFrom
@@ -57,6 +58,15 @@ api:
    timeoutSeconds: 3
    failureThreshold: 3

+  # Pod disruption budget
+  podDisruptionBudget:
+    enabled: false
+    minAvailable: 1
+    # maxUnavailable: 1
+
+  # Pod affinity/anti-affinity (overrides global affinity for this component)
+  # affinity: {}
+
  # Environment variables
  env:
    #HINDSIGHT_API_LLM_PROVIDER: "groq"
@@ -75,7 +85,7 @@ worker:
  image:
    repository: ghcr.io/vectorize-io/hindsight-api
    pullPolicy: IfNotPresent
-    # tag defaults to .Values.version if not specified
+    # tag: ""  # defaults to .Values.version, then Chart.appVersion if not specified

  service:
    # Service for metrics scraping (headless for StatefulSet)
@@ -121,6 +131,15 @@ worker:
    # HTTP port for metrics/health (matches service.targetPort)
    HINDSIGHT_API_WORKER_HTTP_PORT: "8889"

+  # Pod disruption budget
+  podDisruptionBudget:
+    enabled: false
+    minAvailable: 1
+    # maxUnavailable: 1
+
+  # Pod affinity/anti-affinity (overrides global affinity for this component)
+  # affinity: {}
+
  # Secret environment variables (inherited from api.secrets if not specified)
  secrets: {}

@@ -164,6 +183,15 @@ controlPlane:
    timeoutSeconds: 3
    failureThreshold: 3

+  # Pod disruption budget
+  podDisruptionBudget:
+    enabled: false
+    minAvailable: 1
+    # maxUnavailable: 1
+
+  # Pod affinity/anti-affinity (overrides global affinity for this component)
+  # affinity: {}
+
  # Environment variables
  env:
    NODE_ENV: "production"
@@ -262,9 +290,87 @@ nodeSelector: {}
 # Tolerations
 tolerations: []

-# Affinity
+# Affinity (applied to all components unless overridden per-component)
 affinity: {}

+# TEI (Text Embeddings Inference) - optional standalone deployments
+# for reranking and/or embedding models
+tei:
+  reranker:
+    enabled: false
+    replicaCount: 1
+    image:
+      repository: ghcr.io/huggingface/text-embeddings-inference
+      tag: cpu-1.8.3
+      pullPolicy: IfNotPresent
+    model: "cross-encoder/ms-marco-MiniLM-L-6-v2"
+    port: 8090
+    args:
+      - "--auto-truncate"
+    env:
+      PAYLOAD_LIMIT: "10000000"
+      MAX_CLIENT_BATCH_SIZE: "256"
+    resources:
+      limits:
+        cpu: 2000m
+        memory: 2Gi
+      requests:
+        cpu: 500m
+        memory: 1Gi
+    livenessProbe:
+      httpGet:
+        path: /health
+        port: 8090
+      initialDelaySeconds: 30
+      periodSeconds: 10
+      timeoutSeconds: 5
+      failureThreshold: 6
+    readinessProbe:
+      httpGet:
+        path: /health
+        port: 8090
+      initialDelaySeconds: 15
+      periodSeconds: 5
+      timeoutSeconds: 3
+      failureThreshold: 3
+
+  embedding:
+    enabled: false
+    replicaCount: 1
+    image:
+      repository: ghcr.io/huggingface/text-embeddings-inference
+      tag: cpu-1.8.3
+      pullPolicy: IfNotPresent
+    model: "sentence-transformers/all-MiniLM-L6-v2"
+    port: 8091
+    args: []
+    env:
+      PAYLOAD_LIMIT: "10000000"
+      MAX_CLIENT_BATCH_SIZE: "256"
+    resources:
+      limits:
+        cpu: 2000m
+        memory: 2Gi
+      requests:
+        cpu: 500m
+        memory: 1Gi
+    livenessProbe:
+      httpGet:
+        path: /health
+        port: 8091
+      initialDelaySeconds: 30
+      periodSeconds: 10
+      timeoutSeconds: 5
+      failureThreshold: 6
+    readinessProbe:
+      httpGet:
+        path: /health
+        port: 8091
+      initialDelaySeconds: 15
+      periodSeconds: 5
+      timeoutSeconds: 3
+      failureThreshold: 3
+
 # Autoscaling
 autoscaling:
  enabled: false
@@ -46,4 +46,4 @@ __all__ = [
    "RemoteTEICrossEncoder",
    "LLMConfig",
 ]
-__version__ = "0.4.6"
+__version__ = "0.4.10"
@@ -6,11 +6,13 @@ Create Date: 2025-11-27 11:54:19.228030

 """

+import os
 from collections.abc import Sequence

 import sqlalchemy as sa
 from alembic import op
 from pgvector.sqlalchemy import Vector
+from sqlalchemy import text
 from sqlalchemy.dialects import postgresql

 # revision identifiers, used by Alembic.
@@ -20,11 +22,79 @@ branch_labels: str | Sequence[str] | None = None
 depends_on: str | Sequence[str] | None = None


+def _detect_vector_extension() -> str:
+    """
+    Detect or validate vector extension: 'vchord' or 'pgvector'.
+    Respects HINDSIGHT_API_VECTOR_EXTENSION env var if set.
+    """
+    conn = op.get_bind()
+    vector_extension = os.getenv("HINDSIGHT_API_VECTOR_EXTENSION", "pgvector").lower()
+
+    # Validate configured extension is installed
+    if vector_extension == "vchord":
+        vchord_check = conn.execute(text("SELECT 1 FROM pg_extension WHERE extname = 'vchord'")).scalar()
+        if not vchord_check:
+            raise RuntimeError(
+                "Configured vector extension 'vchord' not found. Install it with: CREATE EXTENSION vchord CASCADE;"
+            )
+        return "vchord"
+    elif vector_extension == "pgvector":
+        pgvector_check = conn.execute(text("SELECT 1 FROM pg_extension WHERE extname = 'vector'")).scalar()
+        if not pgvector_check:
+            raise RuntimeError(
+                "Configured vector extension 'pgvector' not found. Install it with: CREATE EXTENSION vector;"
+            )
+        return "pgvector"
+    else:
+        raise ValueError(f"Invalid HINDSIGHT_API_VECTOR_EXTENSION: {vector_extension}. Must be 'pgvector' or 'vchord'")
+
+
+def _detect_text_search_extension() -> str:
+    """
+    Detect or validate text search extension: 'native' or 'vchord'.
+    Respects HINDSIGHT_API_TEXT_SEARCH_EXTENSION env var.
+    Creates the extension if needed.
+    """
+    text_search_extension = os.getenv("HINDSIGHT_API_TEXT_SEARCH_EXTENSION", "native").lower()
+
+    if text_search_extension == "vchord":
+        # Create vchord_bm25 extension if not exists
+        try:
+            op.execute("CREATE EXTENSION IF NOT EXISTS vchord_bm25 CASCADE")
+        except Exception:
+            # Extension might already exist or user lacks permissions - verify it exists
+            conn = op.get_bind()
+            result = conn.execute(text("SELECT 1 FROM pg_extension WHERE extname = 'vchord_bm25'")).fetchone()
+            if not result:
+                # Extension truly doesn't exist - re-raise the error
+                raise
+        return "vchord"
+    elif text_search_extension == "native":
+        return "native"
+    else:
+        raise ValueError(
+            f"Invalid HINDSIGHT_API_TEXT_SEARCH_EXTENSION: {text_search_extension}. Must be 'native' or 'vchord'"
+        )
+
+
 def upgrade() -> None:
    """Upgrade schema - create all tables from scratch."""

-    # Enable required extensions
-    op.execute("CREATE EXTENSION IF NOT EXISTS vector")
+    # Note: pgvector extension is installed globally BEFORE migrations run
+    # See migrations.py:run_migrations() - this ensures the extension is available
+    # to all schemas, not just the one being migrated
+
+    # We keep this here as a fallback for backwards compatibility
+    # This may fail if user lacks permissions, which is fine if extension already exists
+    try:
+        op.execute("CREATE EXTENSION IF NOT EXISTS vector")
+    except Exception:
+        # Extension might already exist or user lacks permissions - verify it exists
+        conn = op.get_bind()
+        result = conn.execute(text("SELECT 1 FROM pg_extension WHERE extname = 'vector'")).fetchone()
+        if not result:
+            # Extension truly doesn't exist - re-raise the error
+            raise

    # Create banks table
    op.create_table(
@@ -152,11 +222,23 @@ def upgrade() -> None:
    )

    # Add search_vector column for full-text search
-    op.execute("""
-        ALTER TABLE memory_units
-        ADD COLUMN search_vector tsvector
-        GENERATED ALWAYS AS (to_tsvector('english', COALESCE(text, '') || ' ' || COALESCE(context, ''))) STORED
-    """)
+    # Type depends on configured text search backend
+    text_search_ext = _detect_text_search_extension()
+
+    if text_search_ext == "vchord":
+        # VectorChord BM25: bm25vector type (no GENERATED - tokenization happens on INSERT)
+        # Note: vchord_bm25 extension creates types in bm25_catalog schema
+        op.execute("""
+            ALTER TABLE memory_units
+            ADD COLUMN search_vector bm25_catalog.bm25vector
+        """)
+    else:  # native
+        # Native PostgreSQL: tsvector with automatic generation
+        op.execute("""
+            ALTER TABLE memory_units
+            ADD COLUMN search_vector tsvector
+            GENERATED ALWAYS AS (to_tsvector('english', COALESCE(text, '') || ' ' || COALESCE(context, ''))) STORED
+        """)

    op.create_index("idx_memory_units_bank_id", "memory_units", ["bank_id"])
    op.create_index("idx_memory_units_document_id", "memory_units", ["document_id"])
@@ -186,19 +268,39 @@ def upgrade() -> None:
        ["bank_id", sa.text("event_date DESC")],
        postgresql_where=sa.text("fact_type = 'observation'"),
    )
-    op.create_index(
-        "idx_memory_units_embedding",
-        "memory_units",
-        ["embedding"],
-        postgresql_using="hnsw",
-        postgresql_ops={"embedding": "vector_cosine_ops"},
-    )
+    # Create vector index - conditional based on available extension
+    vector_ext = _detect_vector_extension()

-    # Create BM25 full-text search index on search_vector
-    op.execute("""
-        CREATE INDEX idx_memory_units_text_search ON memory_units
-        USING gin(search_vector)
-    """)
+    if vector_ext == "vchord":
+        # Use vchordrq index for vchord (supports high-dimensional embeddings)
+        op.execute("""
+            CREATE INDEX idx_memory_units_embedding ON memory_units
+            USING vchordrq (embedding vector_l2_ops)
+        """)
+    else:  # pgvector
+        # Use HNSW index for pgvector
+        op.create_index(
+            "idx_memory_units_embedding",
+            "memory_units",
+            ["embedding"],
+            postgresql_using="hnsw",
+            postgresql_ops={"embedding": "vector_cosine_ops"},
+        )
+
+    # Create full-text search index on search_vector
+    # Index type depends on text search backend
+    if text_search_ext == "vchord":
+        # VectorChord BM25 index
+        op.execute("""
+            CREATE INDEX idx_memory_units_text_search ON memory_units
+            USING bm25 (search_vector bm25_catalog.bm25_ops)
+        """)
+    else:  # native
+        # Native PostgreSQL GIN index
+        op.execute("""
+            CREATE INDEX idx_memory_units_text_search ON memory_units
+            USING gin(search_vector)
+        """)

    op.execute("""
        CREATE MATERIALIZED VIEW memory_units_bm25 AS
@@ -10,9 +10,11 @@ This migration:
 3. Adds consolidation tracking columns to the 'banks' table
 """

+import os
 from collections.abc import Sequence

 from alembic import context, op
+from sqlalchemy import text

 # revision identifiers, used by Alembic.
 revision: str = "n9i0j1k2l3m4"
@@ -27,10 +29,71 @@ def _get_schema_prefix() -> str:
    return f'"{schema}".' if schema else ""


+def _detect_vector_extension() -> str:
+    """
+    Detect or validate vector extension: 'vchord' or 'pgvector'.
+    Respects HINDSIGHT_API_VECTOR_EXTENSION env var if set.
+    """
+    conn = op.get_bind()
+    vector_extension = os.getenv("HINDSIGHT_API_VECTOR_EXTENSION", "pgvector").lower()
+
+    # Validate configured extension is installed
+    if vector_extension == "vchord":
+        vchord_check = conn.execute(text("SELECT 1 FROM pg_extension WHERE extname = 'vchord'")).scalar()
+        if not vchord_check:
+            raise RuntimeError(
+                "Configured vector extension 'vchord' not found. Install it with: CREATE EXTENSION vchord CASCADE;"
+            )
+        return "vchord"
+    elif vector_extension == "pgvector":
+        pgvector_check = conn.execute(text("SELECT 1 FROM pg_extension WHERE extname = 'vector'")).scalar()
+        if not pgvector_check:
+            raise RuntimeError(
+                "Configured vector extension 'pgvector' not found. Install it with: CREATE EXTENSION vector;"
+            )
+        return "pgvector"
+    else:
+        raise ValueError(f"Invalid HINDSIGHT_API_VECTOR_EXTENSION: {vector_extension}. Must be 'pgvector' or 'vchord'")
+
+
+def _detect_text_search_extension() -> str:
+    """
+    Detect or validate text search extension: 'native' or 'vchord'.
+    Respects HINDSIGHT_API_TEXT_SEARCH_EXTENSION env var.
+    Creates the extension if needed.
+    """
+    text_search_extension = os.getenv("HINDSIGHT_API_TEXT_SEARCH_EXTENSION", "native").lower()
+
+    if text_search_extension == "vchord":
+        # Create vchord_bm25 extension if not exists
+        try:
+            op.execute("CREATE EXTENSION IF NOT EXISTS vchord_bm25 CASCADE")
+        except Exception:
+            # Extension might already exist or user lacks permissions - verify it exists
+            conn = op.get_bind()
+            result = conn.execute(text("SELECT 1 FROM pg_extension WHERE extname = 'vchord_bm25'")).fetchone()
+            if not result:
+                # Extension truly doesn't exist - re-raise the error
+                raise
+        return "vchord"
+    elif text_search_extension == "native":
+        return "native"
+    else:
+        raise ValueError(
+            f"Invalid HINDSIGHT_API_TEXT_SEARCH_EXTENSION: {text_search_extension}. Must be 'native' or 'vchord'"
+        )
+
+
 def upgrade() -> None:
    """Create learnings and pinned_reflections tables."""
    schema = _get_schema_prefix()

+    # Detect which vector extension is available
+    vector_ext = _detect_vector_extension()
+
+    # Detect which text search extension to use
+    text_search_ext = _detect_text_search_extension()
+
    # 1. Create learnings table
    op.execute(f"""
        CREATE TABLE {schema}learnings (
@@ -57,18 +120,39 @@ def upgrade() -> None:

    # Indexes for learnings
    op.execute(f"CREATE INDEX idx_learnings_bank_id ON {schema}learnings(bank_id)")
-    op.execute(f"""
-        CREATE INDEX idx_learnings_embedding ON {schema}learnings
-        USING hnsw (embedding vector_cosine_ops)
-    """)
+
+    # Create vector index based on detected extension
+    if vector_ext == "vchord":
+        op.execute(f"""
+            CREATE INDEX idx_learnings_embedding ON {schema}learnings
+            USING vchordrq (embedding vector_l2_ops)
+        """)
+    else:  # pgvector
+        op.execute(f"""
+            CREATE INDEX idx_learnings_embedding ON {schema}learnings
+            USING hnsw (embedding vector_cosine_ops)
+        """)
+
    op.execute(f"CREATE INDEX idx_learnings_tags ON {schema}learnings USING GIN(tags)")

    # Full-text search for learnings
-    op.execute(f"""
-        ALTER TABLE {schema}learnings ADD COLUMN search_vector tsvector
-        GENERATED ALWAYS AS (to_tsvector('english', text)) STORED
-    """)
-    op.execute(f"CREATE INDEX idx_learnings_text_search ON {schema}learnings USING gin(search_vector)")
+    if text_search_ext == "vchord":
+        # VectorChord BM25: bm25vector type (no GENERATED - tokenization happens on INSERT)
+        # Note: vchord_bm25 extension creates types in bm25_catalog schema
+        op.execute(f"""
+            ALTER TABLE {schema}learnings ADD COLUMN search_vector bm25_catalog.bm25vector
+        """)
+        op.execute(f"""
+            CREATE INDEX idx_learnings_text_search ON {schema}learnings
+            USING bm25 (search_vector bm25_catalog.bm25_ops)
+        """)
+    else:  # native
+        # Native PostgreSQL: tsvector with automatic generation
+        op.execute(f"""
+            ALTER TABLE {schema}learnings ADD COLUMN search_vector tsvector
+            GENERATED ALWAYS AS (to_tsvector('english', text)) STORED
+        """)
+        op.execute(f"CREATE INDEX idx_learnings_text_search ON {schema}learnings USING gin(search_vector)")

    # 2. Create pinned_reflections table
    op.execute(f"""
@@ -94,21 +178,42 @@ def upgrade() -> None:

    # Indexes for pinned_reflections
    op.execute(f"CREATE INDEX idx_pinned_reflections_bank_id ON {schema}pinned_reflections(bank_id)")
-    op.execute(f"""
-        CREATE INDEX idx_pinned_reflections_embedding ON {schema}pinned_reflections
-        USING hnsw (embedding vector_cosine_ops)
-    """)
+
+    # Create vector index based on detected extension
+    if vector_ext == "vchord":
+        op.execute(f"""
+            CREATE INDEX idx_pinned_reflections_embedding ON {schema}pinned_reflections
+            USING vchordrq (embedding vector_l2_ops)
+        """)
+    else:  # pgvector
+        op.execute(f"""
+            CREATE INDEX idx_pinned_reflections_embedding ON {schema}pinned_reflections
+            USING hnsw (embedding vector_cosine_ops)
+        """)
+
    op.execute(f"CREATE INDEX idx_pinned_reflections_tags ON {schema}pinned_reflections USING GIN(tags)")

    # Full-text search for pinned_reflections
-    op.execute(f"""
-        ALTER TABLE {schema}pinned_reflections ADD COLUMN search_vector tsvector
-        GENERATED ALWAYS AS (to_tsvector('english', COALESCE(name, '') || ' ' || content)) STORED
-    """)
-    op.execute(f"""
-        CREATE INDEX idx_pinned_reflections_text_search ON {schema}pinned_reflections
-        USING gin(search_vector)
-    """)
+    if text_search_ext == "vchord":
+        # VectorChord BM25: bm25vector type (no GENERATED - tokenization happens on INSERT/UPDATE)
+        # Note: vchord_bm25 extension creates types in bm25_catalog schema
+        op.execute(f"""
+            ALTER TABLE {schema}pinned_reflections ADD COLUMN search_vector bm25_catalog.bm25vector
+        """)
+        op.execute(f"""
+            CREATE INDEX idx_pinned_reflections_text_search ON {schema}pinned_reflections
+            USING bm25 (search_vector bm25_catalog.bm25_ops)
+        """)
+    else:  # native
+        # Native PostgreSQL: tsvector with automatic generation
+        op.execute(f"""
+            ALTER TABLE {schema}pinned_reflections ADD COLUMN search_vector tsvector
+            GENERATED ALWAYS AS (to_tsvector('english', COALESCE(name, '') || ' ' || content)) STORED
+        """)
+        op.execute(f"""
+            CREATE INDEX idx_pinned_reflections_text_search ON {schema}pinned_reflections
+            USING gin(search_vector)
+        """)

    # 3. Add consolidation tracking columns to banks table
    op.execute(f"""
@@ -0,0 +1,60 @@
+"""Fix mental_models primary key to be scoped per bank
+
+Revision ID: w8r9s0t1u2v3
+Revises: v7q8r9s0t1u2
+Create Date: 2026-02-05
+
+This migration fixes a critical bank isolation bug where mental_models.id was
+globally unique across all banks instead of being scoped per bank. This caused
+conflicts when different banks tried to use the same custom ID.
+
+CRITICAL FIX: Changes primary key from (id) to (bank_id, id) to ensure proper isolation.
+"""
+
+from collections.abc import Sequence
+
+from alembic import context, op
+
+revision: str = "w8r9s0t1u2v3"
+down_revision: str | Sequence[str] | None = "v7q8r9s0t1u2"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+
+def _get_schema_prefix() -> str:
+    """Get schema prefix for table names (required for multi-tenant support)."""
+    schema = context.config.get_main_option("target_schema")
+    return f'"{schema}".' if schema else ""
+
+
+def upgrade() -> None:
+    """Change mental_models primary key from (id) to (bank_id, id) for proper bank isolation."""
+    schema = _get_schema_prefix()
+
+    # Drop the old primary key constraint (just id)
+    # Note: The constraint might be named differently on different DBs
+    # Try both old names (pinned_reflections_pkey from original, mental_models_pkey from rename)
+    op.execute(f"ALTER TABLE {schema}mental_models DROP CONSTRAINT IF EXISTS pinned_reflections_pkey")
+    op.execute(f"ALTER TABLE {schema}mental_models DROP CONSTRAINT IF EXISTS mental_models_pkey")
+
+    # Create the new composite primary key (bank_id, id)
+    # This ensures IDs are scoped per bank, not globally
+    op.execute(f"""
+        ALTER TABLE {schema}mental_models
+        ADD CONSTRAINT mental_models_pkey PRIMARY KEY (bank_id, id)
+    """)
+
+
+def downgrade() -> None:
+    """Revert mental_models primary key from (bank_id, id) to (id)."""
+    schema = _get_schema_prefix()
+
+    # Drop the composite primary key
+    op.execute(f"ALTER TABLE {schema}mental_models DROP CONSTRAINT IF EXISTS mental_models_pkey")
+
+    # Restore the old primary key (just id)
+    # WARNING: This downgrade will fail if there are duplicate IDs across banks
+    op.execute(f"""
+        ALTER TABLE {schema}mental_models
+        ADD CONSTRAINT mental_models_pkey PRIMARY KEY (id)
+    """)
@@ -0,0 +1,64 @@
+"""Add config JSONB column to banks table for hierarchical configuration
+
+Revision ID: x9s0t1u2v3w4
+Revises: w8r9s0t1u2v3
+Create Date: 2026-02-09
+
+This migration adds a `config` JSONB column to the banks table to support
+per-bank configuration overrides. This enables hierarchical configuration where:
+- Global config is loaded from environment variables
+- Tenant config is provided via TenantExtension
+- Bank config overrides are stored in banks.config JSONB column
+
+The config column stores overrides for hierarchical fields (LLM settings,
+retention parameters, retrieval settings, etc.) in Python field name format.
+"""
+
+from collections.abc import Sequence
+
+import sqlalchemy as sa
+from alembic import context, op
+from sqlalchemy.dialects.postgresql import JSONB
+
+revision: str = "x9s0t1u2v3w4"
+down_revision: str | Sequence[str] | None = "w8r9s0t1u2v3"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+
+def _get_schema_prefix() -> str:
+    """Get schema prefix for table names (required for multi-tenant support)."""
+    schema = context.config.get_main_option("target_schema")
+    return f'"{schema}".' if schema else ""
+
+
+def upgrade() -> None:
+    """Add config JSONB column to banks table with GIN index."""
+    schema = _get_schema_prefix()
+
+    # Add config column to banks table
+    op.execute(f"""
+        ALTER TABLE {schema}banks
+        ADD COLUMN config JSONB NOT NULL DEFAULT '{{}}'::jsonb
+    """)
+
+    # Add GIN index for efficient JSONB queries
+    op.execute(f"""
+        CREATE INDEX idx_banks_config
+        ON {schema}banks
+        USING gin(config)
+    """)
+
+
+def downgrade() -> None:
+    """Remove config column and index from banks table."""
+    schema = _get_schema_prefix()
+
+    # Drop index first
+    op.execute(f"DROP INDEX IF EXISTS {schema}idx_banks_config")
+
+    # Drop column
+    op.execute(f"""
+        ALTER TABLE {schema}banks
+        DROP COLUMN IF EXISTS config
+    """)
@@ -6,7 +6,6 @@ Provides both HTTP REST API and MCP (Model Context Protocol) server.

 import logging
 from contextlib import asynccontextmanager
-from typing import Optional

 from fastapi import FastAPI

@@ -46,14 +45,14 @@ def create_app(
        # Both HTTP and MCP
        app = create_app(memory, mcp_api_enabled=True)
    """
-    mcp_app = None
+    mcp_servers = None

-    # Create MCP app first if enabled (we need its lifespan for chaining)
+    # Create MCP servers first if enabled (we need their lifespans for chaining)
    if mcp_api_enabled:
        try:
-            from .mcp import create_mcp_app
+            from .mcp import MCPMiddleware, create_mcp_servers

-            mcp_app = create_mcp_app(memory=memory)
+            mcp_servers = create_mcp_servers(memory=memory)
        except ImportError as e:
            logger.error(f"MCP server requested but dependencies not available: {e}")
            logger.error("Install with: pip install hindsight-api[mcp]")
@@ -70,30 +69,41 @@ def create_app(
        app = FastAPI(title="Hindsight API", version="0.0.7")
        logger.info("HTTP REST API disabled")

-    # Mount MCP server and chain its lifespan if enabled
-    if mcp_app is not None:
-        # Get the MCP app's underlying Starlette app for lifespan access
-        mcp_starlette_app = mcp_app.mcp_app
+    # Add MCP middleware and chain its lifespan if enabled
+    if mcp_servers is not None:
+        multi_bank_server, single_bank_server, multi_bank_starlette_app, single_bank_starlette_app = mcp_servers

        # Store the original lifespan
        original_lifespan = app.router.lifespan_context

        @asynccontextmanager
        async def chained_lifespan(app_instance: FastAPI):
-            """Chain the MCP lifespan with the main app lifespan."""
-            # Start MCP lifespan first
-            async with mcp_starlette_app.router.lifespan_context(mcp_starlette_app):
-                logger.info("MCP lifespan started")
-                # Then start the original app lifespan
-                async with original_lifespan(app_instance):
-                    yield
-            logger.info("MCP lifespan stopped")
+            """Chain both MCP lifespans with the main app lifespan."""
+            # Start both MCP lifespans (multi-bank and single-bank)
+            async with multi_bank_starlette_app.router.lifespan_context(multi_bank_starlette_app):
+                async with single_bank_starlette_app.router.lifespan_context(single_bank_starlette_app):
+                    logger.info("MCP lifespans started (multi-bank and single-bank)")
+                    # Then start the original app lifespan
+                    async with original_lifespan(app_instance):
+                        yield
+                logger.info("MCP lifespans stopped")

        # Replace the app's lifespan with the chained version
        app.router.lifespan_context = chained_lifespan

-        # Mount the MCP middleware
-        app.mount(mcp_mount_path, mcp_app)
+        # Add MCP as a wrapping middleware — intercepts /mcp* requests directly,
+        # passes everything else through to the FastAPI app. No Starlette Mount
+        # means no 307 redirect for /mcp (no trailing slash).
+        app.add_middleware(
+            MCPMiddleware,
+            memory=memory,
+            prefix=mcp_mount_path,
+            multi_bank_app=multi_bank_starlette_app,
+            single_bank_app=single_bank_starlette_app,
+            multi_bank_server=multi_bank_server,
+            single_bank_server=single_bank_server,
+        )
+
        logger.info(f"MCP server enabled at {mcp_mount_path}/")

    return app
@@ -5,6 +5,7 @@ This module provides the create_app function to create and configure
 the FastAPI application with all API endpoints.
 """

+import asyncio
 import json
 import logging
 import uuid
@@ -31,11 +32,47 @@ def _parse_metadata(metadata: Any) -> dict[str, Any]:
    return {}


+from typing import Callable
+
 from pydantic import BaseModel, ConfigDict, Field, field_validator

 from hindsight_api import MemoryEngine
+
+
+def FieldWithDefault(default_factory: Callable, **kwargs) -> Any:
+    """
+    Field wrapper that ensures default_factory values appear in OpenAPI schema.
+
+    Pydantic doesn't include default_factory in OpenAPI schemas, causing OpenAPI
+    Generator to make fields Optional with default=None instead of non-optional
+    with the correct default value.
+
+    This wrapper adds json_schema_extra to include the default in the schema.
+    """
+    # Determine the default value for the schema based on the factory
+    if default_factory is list:
+        schema_default = []
+    elif default_factory is dict:
+        schema_default = {}
+    else:
+        # For custom factories (like IncludeOptions), use empty dict as placeholder
+        schema_default = {}
+
+    # Add or merge json_schema_extra
+    json_extra = kwargs.pop("json_schema_extra", {})
+    if isinstance(json_extra, dict):
+        json_extra["default"] = schema_default
+    else:
+        # If json_schema_extra was a function, we can't merge easily
+        # Fall back to just setting default
+        json_extra = {"default": schema_default}
+
+    return Field(default_factory=default_factory, json_schema_extra=json_extra, **kwargs)
+
+
+from hindsight_api.config import get_config
 from hindsight_api.engine.db_utils import acquire_with_retry
-from hindsight_api.engine.memory_engine import Budget, fq_table
+from hindsight_api.engine.memory_engine import Budget, _get_tiktoken_encoding, fq_table
 from hindsight_api.engine.reflect.observations import Observation
 from hindsight_api.engine.response_models import VALID_RECALL_FACT_TYPES, TokenUsage
 from hindsight_api.engine.search.tags import TagsMatch
@@ -45,6 +82,8 @@ from hindsight_api.models import RequestContext

 logger = logging.getLogger(__name__)

+MAX_QUERY_TOKENS = 500  # Maximum tokens allowed in recall query
+

 class EntityIncludeOptions(BaseModel):
    """Options for including entity observations in recall results."""
@@ -100,8 +139,8 @@ class RecallRequest(BaseModel):
    query_timestamp: str | None = Field(
        default=None, description="ISO format date string (e.g., '2023-05-30T23:40:00')"
    )
-    include: IncludeOptions = Field(
-        default_factory=IncludeOptions,
+    include: IncludeOptions = FieldWithDefault(
+        IncludeOptions,
        description="Options for including additional data (entities are included by default)",
    )
    tags: list[str] | None = Field(
@@ -520,7 +559,9 @@ class ReflectFact(BaseModel):
    )

    id: str | None = None
-    text: str
+    text: str = Field(
+        description="Fact text. When type='observation', this contains markdown-formatted consolidated knowledge"
+    )
    type: str | None = None  # fact type: world, experience, observation
    context: str | None = None
    occurred_start: str | None = None
@@ -565,18 +606,16 @@ class ReflectLLMCall(BaseModel):
 class ReflectBasedOn(BaseModel):
    """Evidence the response is based on: memories, mental models, and directives."""

-    memories: list[ReflectFact] = Field(default_factory=list, description="Memory facts used to generate the response")
-    mental_models: list[ReflectMentalModel] = Field(
-        default_factory=list, description="Mental models used during reflection"
-    )
-    directives: list[ReflectDirective] = Field(default_factory=list, description="Directives applied during reflection")
+    memories: list[ReflectFact] = FieldWithDefault(list, description="Memory facts used to generate the response")
+    mental_models: list[ReflectMentalModel] = FieldWithDefault(list, description="Mental models used during reflection")
+    directives: list[ReflectDirective] = FieldWithDefault(list, description="Directives applied during reflection")


 class ReflectTrace(BaseModel):
    """Execution trace of LLM and tool calls during reflection."""

-    tool_calls: list[ReflectToolCall] = Field(default_factory=list, description="Tool calls made during reflection")
-    llm_calls: list[ReflectLLMCall] = Field(default_factory=list, description="LLM calls made during reflection")
+    tool_calls: list[ReflectToolCall] = FieldWithDefault(list, description="Tool calls made during reflection")
+    llm_calls: list[ReflectLLMCall] = FieldWithDefault(list, description="LLM calls made during reflection")


 class ReflectResponse(BaseModel):
@@ -585,7 +624,7 @@ class ReflectResponse(BaseModel):
    model_config = ConfigDict(
        json_schema_extra={
            "example": {
-                "text": "Based on my understanding, AI is a transformative technology...",
+                "text": "## AI Overview\n\nBased on my understanding, AI is a **transformative technology**:\n\n- Used extensively in healthcare\n- Discussed in recent conversations\n- Continues to evolve rapidly",
                "based_on": {
                    "memories": [
                        {"id": "123", "text": "AI is used in healthcare", "type": "world"},
@@ -613,7 +652,9 @@ class ReflectResponse(BaseModel):
        }
    )

-    text: str
+    text: str = Field(
+        description="The reflect response as well-formatted markdown (headers, lists, bold/italic, code blocks, etc.)"
+    )
    based_on: ReflectBasedOn | None = Field(
        default=None,
        description="Evidence used to generate the response. Only present when include.facts is set.",
@@ -786,6 +827,55 @@ class CreateBankRequest(BaseModel):
    background: str | None = Field(default=None, description="Deprecated: use mission instead")


+class BankConfigUpdate(BaseModel):
+    """Request model for updating bank configuration."""
+
+    model_config = ConfigDict(
+        json_schema_extra={
+            "example": {
+                "updates": {
+                    "llm_model": "claude-sonnet-4-5",
+                    "retain_extraction_mode": "verbose",
+                    "retain_custom_instructions": "Extract technical details carefully",
+                }
+            }
+        }
+    )
+
+    updates: dict[str, Any] = Field(
+        description="Configuration overrides. Keys can be in Python field format (llm_provider) "
+        "or environment variable format (HINDSIGHT_API_LLM_PROVIDER). "
+        "Only hierarchical fields can be overridden per-bank."
+    )
+
+
+class BankConfigResponse(BaseModel):
+    """Response model for bank configuration."""
+
+    model_config = ConfigDict(
+        json_schema_extra={
+            "example": {
+                "bank_id": "my-bank",
+                "config": {
+                    "llm_provider": "openai",
+                    "llm_model": "gpt-4",
+                    "retain_extraction_mode": "verbose",
+                },
+                "overrides": {
+                    "llm_model": "gpt-4",
+                    "retain_extraction_mode": "verbose",
+                },
+            }
+        }
+    )
+
+    bank_id: str = Field(description="Bank identifier")
+    config: dict[str, Any] = Field(
+        description="Fully resolved configuration with all hierarchical overrides applied (Python field names)"
+    )
+    overrides: dict[str, Any] = Field(description="Bank-specific configuration overrides only (Python field names)")
+
+
 class GraphDataResponse(BaseModel):
    """Response model for graph data endpoint."""

@@ -863,6 +953,7 @@ class ListDocumentsResponse(BaseModel):
                        "updated_at": "2024-01-15T10:30:00Z",
                        "text_length": 5420,
                        "memory_unit_count": 15,
+                        "tags": ["user_a", "session_123"],
                    }
                ],
                "total": 50,
@@ -934,7 +1025,7 @@ class DocumentResponse(BaseModel):
    created_at: str
    updated_at: str
    memory_unit_count: int
-    tags: list[str] = Field(default_factory=list, description="Tags associated with this document")
+    tags: list[str] = FieldWithDefault(list, description="Tags associated with this document")


 class DeleteDocumentResponse(BaseModel):
@@ -1058,7 +1149,7 @@ class DirectiveResponse(BaseModel):
    content: str
    priority: int = 0
    is_active: bool = True
-    tags: list[str] = Field(default_factory=list)
+    tags: list[str] = FieldWithDefault(list)
    created_at: str | None = None
    updated_at: str | None = None

@@ -1076,7 +1167,7 @@ class CreateDirectiveRequest(BaseModel):
    content: str = Field(description="The directive text to inject into prompts")
    priority: int = Field(default=0, description="Higher priority directives are injected first")
    is_active: bool = Field(default=True, description="Whether this directive is active")
-    tags: list[str] = Field(default_factory=list, description="Tags for filtering")
+    tags: list[str] = FieldWithDefault(list, description="Tags for filtering")


 class UpdateDirectiveRequest(BaseModel):
@@ -1110,10 +1201,12 @@ class MentalModelResponse(BaseModel):
    bank_id: str
    name: str
    source_query: str
-    content: str
-    tags: list[str] = Field(default_factory=list)
+    content: str = Field(
+        description="The mental model content as well-formatted markdown (auto-generated from reflect endpoint)"
+    )
+    tags: list[str] = FieldWithDefault(list)
    max_tokens: int = Field(default=2048)
-    trigger: MentalModelTrigger = Field(default_factory=MentalModelTrigger)
+    trigger: MentalModelTrigger = FieldWithDefault(MentalModelTrigger)
    last_refreshed_at: str | None = None
    created_at: str | None = None
    reflect_response: dict | None = Field(
@@ -1134,6 +1227,7 @@ class CreateMentalModelRequest(BaseModel):
    model_config = ConfigDict(
        json_schema_extra={
            "example": {
+                "id": "team-communication",
                "name": "Team Communication Preferences",
                "source_query": "How does the team prefer to communicate?",
                "tags": ["team"],
@@ -1143,17 +1237,21 @@ class CreateMentalModelRequest(BaseModel):
        }
    )

+    id: str | None = Field(
+        None, description="Optional custom ID for the mental model (alphanumeric lowercase with hyphens)"
+    )
    name: str = Field(description="Human-readable name for the mental model")
    source_query: str = Field(description="The query to run to generate content")
-    tags: list[str] = Field(default_factory=list, description="Tags for scoped visibility")
+    tags: list[str] = FieldWithDefault(list, description="Tags for scoped visibility")
    max_tokens: int = Field(default=2048, ge=256, le=8192, description="Maximum tokens for generated content")
-    trigger: MentalModelTrigger = Field(default_factory=MentalModelTrigger, description="Trigger settings")
+    trigger: MentalModelTrigger = FieldWithDefault(MentalModelTrigger, description="Trigger settings")


 class CreateMentalModelResponse(BaseModel):
    """Response model for mental model creation."""

-    operation_id: str = Field(description="Operation ID to track progress")
+    mental_model_id: str | None = Field(None, description="ID of the created mental model")
+    operation_id: str = Field(description="Operation ID to track refresh progress")


 class UpdateMentalModelRequest(BaseModel):
@@ -1307,6 +1405,7 @@ class FeaturesInfo(BaseModel):
    observations: bool = Field(description="Whether observations (auto-consolidation) are enabled")
    mcp: bool = Field(description="Whether MCP (Model Context Protocol) server is enabled")
    worker: bool = Field(description="Whether the background worker is enabled")
+    bank_config_api: bool = Field(description="Whether per-bank configuration API is enabled")


 class VersionResponse(BaseModel):
@@ -1320,6 +1419,7 @@ class VersionResponse(BaseModel):
                    "observations": False,
                    "mcp": True,
                    "worker": True,
+                    "bank_config_api": False,
                },
            }
        }
@@ -1385,6 +1485,26 @@ def create_app(
            app.state.prometheus_reader = None
            # Metrics collector is already initialized as no-op by default

+        # Initialize OpenTelemetry tracing if enabled
+        if config.otel_traces_enabled:
+            if not config.otel_exporter_otlp_endpoint:
+                logging.warning("OTEL tracing enabled but no endpoint configured. Tracing disabled.")
+            else:
+                from hindsight_api.tracing import create_span_recorder, initialize_tracing
+
+                try:
+                    initialize_tracing(
+                        service_name=config.otel_service_name,
+                        endpoint=config.otel_exporter_otlp_endpoint,
+                        headers=config.otel_exporter_otlp_headers,
+                        deployment_environment=config.otel_deployment_environment,
+                    )
+                    create_span_recorder()
+                    logging.info("OpenTelemetry tracing enabled and configured")
+                except Exception as e:
+                    logging.error(f"Failed to initialize tracing: {e}")
+                    logging.warning("Continuing without tracing")
+
        # Startup: Initialize database and memory system (migrations run inside initialize if enabled)
        if initialize_memory:
            await memory.initialize()
@@ -1410,13 +1530,19 @@ def create_app(
                poll_interval_ms=config.worker_poll_interval_ms,
                max_retries=config.worker_max_retries,
                schema=schema,
-                tenant_extension=getattr(memory, "_tenant_extension", None),
+                tenant_extension=memory._tenant_extension,
                max_slots=config.worker_max_slots,
                consolidation_max_slots=config.worker_consolidation_max_slots,
            )
            poller_task = asyncio.create_task(poller.run())
            logging.info(f"Worker poller started (worker_id={worker_id})")

+        # Call tenant extension startup hook (e.g. JWKS fetch for Supabase)
+        tenant_extension = memory.tenant_extension
+        if tenant_extension:
+            await tenant_extension.on_startup()
+            logging.info("Tenant extension started")
+
        # Call HTTP extension startup hook
        if http_extension:
            await http_extension.on_startup()
@@ -1435,6 +1561,11 @@ def create_app(
                    pass
            logging.info("Worker poller stopped")

+        # Call tenant extension shutdown hook
+        if tenant_extension:
+            await tenant_extension.on_shutdown()
+            logging.info("Tenant extension stopped")
+
        # Call HTTP extension shutdown hook
        if http_extension:
            await http_extension.on_shutdown()
@@ -1445,6 +1576,9 @@ def create_app(
        logging.info("Memory system closed")

    from hindsight_api import __version__
+    from hindsight_api.config import get_config
+
+    config = get_config()

    app = FastAPI(
        title="Hindsight HTTP API",
@@ -1458,6 +1592,7 @@ def create_app(
            "url": "https://www.apache.org/licenses/LICENSE-2.0.html",
        },
        lifespan=lifespan,
+        root_path=config.base_path,
    )

    # IMPORTANT: Set memory on app.state immediately, don't wait for lifespan
@@ -1564,17 +1699,21 @@ def _register_routes(app: FastAPI):

        Returns version info and feature flags that can be used by clients
        to determine which capabilities are available.
+
+        Note: observations flag shows the global default. Individual banks
+        may override this setting via bank-specific configuration.
        """
        from hindsight_api import __version__
-        from hindsight_api.config import get_config
+        from hindsight_api.config import _get_raw_config

-        config = get_config()
+        config = _get_raw_config()
        return VersionResponse(
            api_version=__version__,
            features=FeaturesInfo(
                observations=config.enable_observations,
                mcp=config.mcp_enabled,
                worker=config.worker_enabled,
+                bank_config_api=config.enable_bank_config_api,
            ),
        )

@@ -1718,6 +1857,15 @@ def _register_routes(app: FastAPI):
        handler_start = time.time()
        metrics = get_metrics_collector()

+        # Validate query length to prevent expensive operations on oversized queries
+        encoding = _get_tiktoken_encoding()
+        query_tokens = len(encoding.encode(request.query))
+        if query_tokens > MAX_QUERY_TOKENS:
+            raise HTTPException(
+                status_code=400,
+                detail=f"Query too long: {query_tokens} tokens exceeds maximum of {MAX_QUERY_TOKENS}. Please shorten your query.",
+            )
+
        try:
            # Default to world and experience if not specified (exclude observation)
            fact_types = request.types if request.types else list(VALID_RECALL_FACT_TYPES)
@@ -1832,6 +1980,15 @@ def _register_routes(app: FastAPI):
            raise HTTPException(status_code=e.status_code, detail=e.reason)
        except (AuthenticationError, HTTPException):
            raise
+        except (asyncio.TimeoutError, TimeoutError):
+            handler_duration = time.time() - handler_start
+            logger.error(
+                f"[RECALL TIMEOUT] bank={bank_id} handler_duration={handler_duration:.3f}s - database query timed out"
+            )
+            raise HTTPException(
+                status_code=504,
+                detail="Request timed out while searching memories. Try a shorter or more specific query.",
+            )
        except Exception as e:
            import traceback

@@ -1889,17 +2046,17 @@ def _register_routes(app: FastAPI):
                directives = []
                for fact_type, facts in core_result.based_on.items():
                    if fact_type == "directives":
-                        # Directives have different structure (id, name, content)
+                        # Directives are dicts with id, name, content (not MemoryFact objects)
                        for directive in facts:
                            directives.append(
                                ReflectDirective(
-                                    id=directive.id,
-                                    name=directive.name,
-                                    content=directive.content,
+                                    id=directive["id"],
+                                    name=directive["name"],
+                                    content=directive["content"],
                                )
                            )
-                    elif fact_type == "mental_models":
-                        # Mental models are MemoryFact with type "mental_models"
+                    elif fact_type == "mental-models":
+                        # Mental models are MemoryFact with type "mental-models" (note: hyphen, not underscore)
                        for fact in facts:
                            mental_models.append(
                                ReflectMentalModel(
@@ -2297,9 +2454,12 @@ def _register_routes(app: FastAPI):
            )
            if mental_model is None:
                raise HTTPException(status_code=404, detail=f"Mental model '{mental_model_id}' not found")
+
            return MentalModelResponse(**mental_model)
        except (AuthenticationError, HTTPException):
            raise
+        except OperationValidationError as e:
+            raise HTTPException(status_code=e.status_code, detail=e.reason)
        except Exception as e:
            import traceback

@@ -2330,6 +2490,7 @@ def _register_routes(app: FastAPI):
                name=body.name,
                source_query=body.source_query,
                content="Generating content...",
+                mental_model_id=body.id if body.id else None,
                tags=body.tags if body.tags else None,
                max_tokens=body.max_tokens,
                trigger=body.trigger.model_dump() if body.trigger else None,
@@ -2341,11 +2502,13 @@ def _register_routes(app: FastAPI):
                mental_model_id=mental_model["id"],
                request_context=request_context,
            )
-            return CreateMentalModelResponse(operation_id=result["operation_id"])
+            return CreateMentalModelResponse(mental_model_id=mental_model["id"], operation_id=result["operation_id"])
        except ValueError as e:
            raise HTTPException(status_code=400, detail=str(e))
        except (AuthenticationError, HTTPException):
            raise
+        except OperationValidationError as e:
+            raise HTTPException(status_code=e.status_code, detail=e.reason)
        except Exception as e:
            import traceback

@@ -2378,6 +2541,8 @@ def _register_routes(app: FastAPI):
            raise HTTPException(status_code=404, detail=str(e))
        except (AuthenticationError, HTTPException):
            raise
+        except OperationValidationError as e:
+            raise HTTPException(status_code=e.status_code, detail=e.reason)
        except Exception as e:
            import traceback

@@ -3202,6 +3367,112 @@ def _register_routes(app: FastAPI):
            logger.error(f"Error in DELETE /v1/default/banks/{bank_id}/observations: {error_detail}")
            raise HTTPException(status_code=500, detail=str(e))

+    @app.get(
+        "/v1/default/banks/{bank_id}/config",
+        response_model=BankConfigResponse,
+        summary="Get bank configuration",
+        description="Get fully resolved configuration for a bank including all hierarchical overrides (global → tenant → bank). "
+        "The 'config' field contains all resolved config values. The 'overrides' field shows only bank-specific overrides.",
+        operation_id="get_bank_config",
+        tags=["Banks"],
+    )
+    async def api_get_bank_config(bank_id: str, request_context: RequestContext = Depends(get_request_context)):
+        """Get configuration for a bank with all hierarchical overrides applied."""
+        if not get_config().enable_bank_config_api:
+            raise HTTPException(
+                status_code=404,
+                detail="Bank configuration API is disabled. Set HINDSIGHT_API_ENABLE_BANK_CONFIG_API=true to enable.",
+            )
+        try:
+            # Get resolved config from config resolver
+            config_dict = await app.state.memory._config_resolver.get_bank_config(bank_id, request_context)
+
+            # Get bank-specific overrides only
+            bank_overrides = await app.state.memory._config_resolver._load_bank_config(bank_id)
+
+            return BankConfigResponse(bank_id=bank_id, config=config_dict, overrides=bank_overrides)
+        except (AuthenticationError, HTTPException):
+            raise
+        except Exception as e:
+            import traceback
+
+            error_detail = f"{str(e)}\n\nTraceback:\n{traceback.format_exc()}"
+            logger.error(f"Error in GET /v1/default/banks/{bank_id}/config: {error_detail}")
+            raise HTTPException(status_code=500, detail=str(e))
+
+    @app.patch(
+        "/v1/default/banks/{bank_id}/config",
+        response_model=BankConfigResponse,
+        summary="Update bank configuration",
+        description="Update configuration overrides for a bank. Only hierarchical fields can be overridden (LLM settings, retention parameters, etc.). "
+        "Keys can be provided in Python field format (llm_provider) or environment variable format (HINDSIGHT_API_LLM_PROVIDER).",
+        operation_id="update_bank_config",
+        tags=["Banks"],
+    )
+    async def api_update_bank_config(
+        bank_id: str, request: BankConfigUpdate, request_context: RequestContext = Depends(get_request_context)
+    ):
+        """Update configuration overrides for a bank."""
+        if not get_config().enable_bank_config_api:
+            raise HTTPException(
+                status_code=404,
+                detail="Bank configuration API is disabled. Set HINDSIGHT_API_ENABLE_BANK_CONFIG_API=true to enable.",
+            )
+        try:
+            # Update config via config resolver (validates configurable fields and permissions)
+            await app.state.memory._config_resolver.update_bank_config(bank_id, request.updates, request_context)
+
+            # Return updated config
+            config_dict = await app.state.memory._config_resolver.get_bank_config(bank_id, request_context)
+            bank_overrides = await app.state.memory._config_resolver._load_bank_config(bank_id)
+
+            return BankConfigResponse(bank_id=bank_id, config=config_dict, overrides=bank_overrides)
+        except ValueError as e:
+            # Validation error (e.g., trying to override static field)
+            raise HTTPException(status_code=400, detail=str(e))
+        except (AuthenticationError, HTTPException):
+            raise
+        except Exception as e:
+            import traceback
+
+            error_detail = f"{str(e)}\n\nTraceback:\n{traceback.format_exc()}"
+            logger.error(f"Error in PATCH /v1/default/banks/{bank_id}/config: {error_detail}")
+            raise HTTPException(status_code=500, detail=str(e))
+
+    @app.delete(
+        "/v1/default/banks/{bank_id}/config",
+        response_model=BankConfigResponse,
+        summary="Reset bank configuration",
+        description="Reset bank configuration to defaults by removing all bank-specific overrides. "
+        "The bank will then use global and tenant-level configuration only.",
+        operation_id="reset_bank_config",
+        tags=["Banks"],
+    )
+    async def api_reset_bank_config(bank_id: str, request_context: RequestContext = Depends(get_request_context)):
+        """Reset bank configuration to defaults (remove all overrides)."""
+        if not get_config().enable_bank_config_api:
+            raise HTTPException(
+                status_code=404,
+                detail="Bank configuration API is disabled. Set HINDSIGHT_API_ENABLE_BANK_CONFIG_API=true to enable.",
+            )
+        try:
+            # Reset config via config resolver
+            await app.state.memory._config_resolver.reset_bank_config(bank_id)
+
+            # Return updated config (should match defaults now)
+            config_dict = await app.state.memory._config_resolver.get_bank_config(bank_id, request_context)
+            bank_overrides = await app.state.memory._config_resolver._load_bank_config(bank_id)
+
+            return BankConfigResponse(bank_id=bank_id, config=config_dict, overrides=bank_overrides)
+        except (AuthenticationError, HTTPException):
+            raise
+        except Exception as e:
+            import traceback
+
+            error_detail = f"{str(e)}\n\nTraceback:\n{traceback.format_exc()}"
+            logger.error(f"Error in DELETE /v1/default/banks/{bank_id}/config: {error_detail}")
+            raise HTTPException(status_code=500, detail=str(e))
+
    @app.post(
        "/v1/default/banks/{bank_id}/consolidate",
        response_model=ConsolidationResponse,
@@ -8,7 +8,11 @@ from contextvars import ContextVar
 from fastmcp import FastMCP

 from hindsight_api import MemoryEngine
+from hindsight_api.engine.memory_engine import _current_schema
+from hindsight_api.extensions import MCPExtension, load_extension
+from hindsight_api.extensions.tenant import AuthenticationError
 from hindsight_api.mcp_tools import MCPToolsConfig, register_mcp_tools
+from hindsight_api.models import RequestContext

 # Configure logging from HINDSIGHT_API_LOG_LEVEL environment variable
 _log_level_str = os.environ.get("HINDSIGHT_API_LOG_LEVEL", "info").lower()
@@ -29,7 +33,8 @@ logger = logging.getLogger(__name__)
 # Default bank_id from environment variable
 DEFAULT_BANK_ID = os.environ.get("HINDSIGHT_MCP_BANK_ID", "default")

-# MCP authentication token (optional - if set, Bearer token auth is required)
+# Legacy MCP authentication token (for backwards compatibility)
+# If set, this token is checked first before TenantExtension auth
 MCP_AUTH_TOKEN = os.environ.get("HINDSIGHT_API_MCP_AUTH_TOKEN")

 # Context variable to hold the current bank_id
@@ -38,6 +43,10 @@ _current_bank_id: ContextVar[str | None] = ContextVar("current_bank_id", default
 # Context variable to hold the current API key (for tenant auth propagation)
 _current_api_key: ContextVar[str | None] = ContextVar("current_api_key", default=None)

+# Context variables for tenant_id and api_key_id (set by authenticate, used by usage metering)
+_current_tenant_id: ContextVar[str | None] = ContextVar("current_tenant_id", default=None)
+_current_api_key_id: ContextVar[str | None] = ContextVar("current_api_key_id", default=None)
+

 def get_current_bank_id() -> str | None:
    """Get the current bank_id from context."""
@@ -49,12 +58,24 @@ def get_current_api_key() -> str | None:
    return _current_api_key.get()


-def create_mcp_server(memory: MemoryEngine) -> FastMCP:
+def get_current_tenant_id() -> str | None:
+    """Get the current tenant_id from context."""
+    return _current_tenant_id.get()
+
+
+def get_current_api_key_id() -> str | None:
+    """Get the current api_key_id from context."""
+    return _current_api_key_id.get()
+
+
+def create_mcp_server(memory: MemoryEngine, multi_bank: bool = True) -> FastMCP:
    """
    Create and configure the Hindsight MCP server.

    Args:
        memory: MemoryEngine instance (required)
+        multi_bank: If True, expose all tools with bank_id parameters (default).
+                   If False, only expose bank-scoped tools without bank_id parameters.

    Returns:
        Configured FastMCP server instance with stateless_http enabled
@@ -66,40 +87,133 @@ def create_mcp_server(memory: MemoryEngine) -> FastMCP:
    config = MCPToolsConfig(
        bank_id_resolver=get_current_bank_id,
        api_key_resolver=get_current_api_key,  # Propagate API key for tenant auth
-        include_bank_id_param=True,  # HTTP MCP supports multi-bank via parameter
-        tools=None,  # All tools
+        tenant_id_resolver=get_current_tenant_id,  # Propagate tenant_id for usage metering
+        api_key_id_resolver=get_current_api_key_id,  # Propagate api_key_id for usage metering
+        include_bank_id_param=multi_bank,
+        tools=None
+        if multi_bank
+        else {
+            "retain",
+            "recall",
+            "reflect",
+            "list_mental_models",
+            "get_mental_model",
+            "create_mental_model",
+            "update_mental_model",
+            "delete_mental_model",
+            "refresh_mental_model",
+        },  # Scoped tools for single-bank mode (excludes bank management: list_banks, create_bank)
        retain_fire_and_forget=False,  # HTTP MCP supports sync/async modes
    )

    register_mcp_tools(mcp, memory, config)

+    # Load and register additional tools from MCP extension if configured
+    mcp_extension = load_extension("MCP", MCPExtension)
+    if mcp_extension:
+        logger.info(f"Loading MCP extension: {mcp_extension.__class__.__name__}")
+        mcp_extension.register_tools(mcp, memory)
+
+    # Make all tools tolerant of extra arguments from LLMs (e.g., "explanation")
+    _make_tools_tolerant(mcp)
+
    return mcp


+def _make_tools_tolerant(mcp: FastMCP) -> None:
+    """Wrap all tool run methods to strip unknown arguments before validation.
+
+    LLMs frequently add extra fields like "explanation" or "reasoning" to tool calls.
+    FastMCP's Pydantic TypeAdapter rejects these with "Unexpected keyword argument".
+    This wraps each tool's run() to filter arguments to only known parameters.
+    """
+    try:
+        for name, tool in mcp._tool_manager._tools.items():
+            if hasattr(tool, "parameters") and tool.parameters:
+                allowed = set(tool.parameters.get("properties", {}).keys())
+                original_run = tool.run
+
+                async def _tolerant_run(arguments, _allowed=allowed, _orig=original_run):
+                    extra_keys = set(arguments.keys()) - _allowed
+                    if extra_keys:
+                        logger.debug(f"Stripping unknown arguments from tool call: {extra_keys}")
+                        arguments = {k: v for k, v in arguments.items() if k in _allowed}
+                    return await _orig(arguments)
+
+                # FunctionTool is a Pydantic model with extra='forbid', so use
+                # object.__setattr__ to bypass Pydantic's setter validation.
+                object.__setattr__(tool, "run", _tolerant_run)
+    except (AttributeError, KeyError) as e:
+        logger.warning(f"Could not make tools tolerant of extra arguments: {e}")
+
+
 class MCPMiddleware:
-    """ASGI middleware that handles authentication and extracts bank_id from header or path.
+    """ASGI middleware that intercepts MCP requests and routes to appropriate MCP server.
+
+    This middleware wraps the main FastAPI app and intercepts requests matching the
+    configured prefix (default: /mcp). Non-MCP requests pass through to the inner app.

    Authentication:
-        If HINDSIGHT_API_MCP_AUTH_TOKEN is set, all requests must include a valid
-        Authorization header with Bearer token or direct token matching the configured value.
+        1. If HINDSIGHT_API_MCP_AUTH_TOKEN is set (legacy), validates against that token
+        2. Otherwise, uses TenantExtension.authenticate_mcp() from the MemoryEngine
+           - DefaultTenantExtension: no auth required (local dev)
+           - ApiKeyTenantExtension: validates against env var

-    Bank ID can be provided via:
-    1. X-Bank-Id header (recommended for Claude Code)
-    2. URL path: /mcp/{bank_id}/
-    3. Environment variable HINDSIGHT_MCP_BANK_ID (fallback default)
+    Two modes based on URL structure:

-    For Claude Code, configure with:
+    1. Multi-bank mode (for /mcp/ root endpoint):
+       - Exposes all tools: retain, recall, reflect, list_banks, create_bank
+       - All tools include optional bank_id parameter for cross-bank operations
+       - Bank ID from: X-Bank-Id header or HINDSIGHT_MCP_BANK_ID env var
+
+    2. Single-bank mode (for /mcp/{bank_id}/ endpoints):
+       - Exposes bank-scoped tools only: retain, recall, reflect
+       - No bank_id parameter (comes from URL)
+       - No bank management tools (list_banks, create_bank)
+       - Recommended for agent isolation
+
+    Bank ID resolution priority:
+        1. URL path (e.g., /mcp/{bank_id}/) → single-bank mode
+        2. X-Bank-Id header → multi-bank mode
+        3. HINDSIGHT_MCP_BANK_ID env var → multi-bank mode (default: "default")
+
+    Examples:
+        # Single-bank mode (recommended for agent isolation)
+        claude mcp add --transport http my-agent http://localhost:8888/mcp/my-agent-bank/ \\
+            --header "Authorization: Bearer <token>"
+
+        # Multi-bank mode (for cross-bank operations)
        claude mcp add --transport http hindsight http://localhost:8888/mcp \\
            --header "X-Bank-Id: my-bank" --header "Authorization: Bearer <token>"
    """

-    def __init__(self, app, memory: MemoryEngine):
+    def __init__(
+        self,
+        app,
+        memory: MemoryEngine,
+        prefix: str = "/mcp",
+        multi_bank_app=None,
+        single_bank_app=None,
+        multi_bank_server=None,
+        single_bank_server=None,
+    ):
        self.app = app
+        self.prefix = prefix
        self.memory = memory
-        self.mcp_server = create_mcp_server(memory)
-        self.mcp_app = self.mcp_server.http_app(path="/")
-        # Expose the lifespan for the parent app to chain
-        self.lifespan = self.mcp_app.lifespan_handler if hasattr(self.mcp_app, "lifespan_handler") else None
+        self.tenant_extension = memory._tenant_extension
+
+        if multi_bank_app and single_bank_app:
+            # Pre-created servers (used when called via add_middleware from create_app)
+            self.multi_bank_app = multi_bank_app
+            self.single_bank_app = single_bank_app
+            self.multi_bank_server = multi_bank_server
+            self.single_bank_server = single_bank_server
+        else:
+            # Create servers internally (for direct construction / tests)
+            self.multi_bank_server = create_mcp_server(memory, multi_bank=True)
+            self.multi_bank_app = self.multi_bank_server.http_app(path="/")
+            self.single_bank_server = create_mcp_server(memory, multi_bank=False)
+            self.single_bank_app = self.single_bank_server.http_app(path="/")

    def _get_header(self, scope: dict, name: str) -> str | None:
        """Extract a header value from ASGI scope."""
@@ -111,9 +225,20 @@ class MCPMiddleware:

    async def __call__(self, scope, receive, send):
        if scope["type"] != "http":
-            await self.mcp_app(scope, receive, send)
+            await self.app(scope, receive, send)
            return

+        path = scope.get("path", "")
+
+        # Check if this is an MCP request (matches prefix)
+        if not (path == self.prefix or path.startswith(self.prefix + "/")):
+            # Not an MCP request — pass through to the inner app
+            await self.app(scope, receive, send)
+            return
+
+        # Strip prefix from path
+        path = path[len(self.prefix) :] or "/"
+
        # Extract auth token from header (for tenant auth propagation)
        auth_header = self._get_header(scope, "Authorization")
        auth_token: str | None = None
@@ -121,62 +246,92 @@ class MCPMiddleware:
            # Support both "Bearer <token>" and direct token
            auth_token = auth_header[7:].strip() if auth_header.startswith("Bearer ") else auth_header.strip()

-        # Authenticate if MCP_AUTH_TOKEN is configured
+        # Authenticate: check legacy MCP_AUTH_TOKEN first, then TenantExtension
+        tenant_context = None
+        auth_tenant_id: str | None = None
+        auth_api_key_id: str | None = None
        if MCP_AUTH_TOKEN:
+            # Legacy authentication mode - validate against static token
            if not auth_token:
                await self._send_error(send, 401, "Authorization header required")
                return
            if auth_token != MCP_AUTH_TOKEN:
                await self._send_error(send, 401, "Invalid authentication token")
                return
+            # Legacy mode doesn't use tenant schemas
+            tenant_context = None
+        else:
+            # Use TenantExtension.authenticate_mcp() for auth
+            try:
+                auth_context = RequestContext(api_key=auth_token)
+                tenant_context = await self.tenant_extension.authenticate_mcp(auth_context)
+                # Capture tenant_id and api_key_id set by authenticate() for usage metering
+                auth_tenant_id = auth_context.tenant_id
+                auth_api_key_id = auth_context.api_key_id
+            except AuthenticationError as e:
+                await self._send_error(send, 401, str(e))
+                return

-        path = scope.get("path", "")
+        # Set schema from tenant context so downstream DB queries use the correct schema
+        schema_token = (
+            _current_schema.set(tenant_context.schema_name) if tenant_context and tenant_context.schema_name else None
+        )

-        # Strip any mount prefix (e.g., /mcp) that FastAPI might not have stripped
-        root_path = scope.get("root_path", "")
-        if root_path and path.startswith(root_path):
-            path = path[len(root_path) :] or "/"
-
-        # Also handle case where mount path wasn't stripped (e.g., /mcp/...)
-        if path.startswith("/mcp/"):
-            path = path[4:]  # Remove /mcp prefix
-        elif path == "/mcp":
-            path = "/"
-
-        # Try to get bank_id from header first (for Claude Code compatibility)
-        bank_id = self._get_header(scope, "X-Bank-Id")
-
-        # MCP endpoint paths that should not be treated as bank_ids
-        MCP_ENDPOINTS = {"sse", "messages"}
-
-        # If no header, try to extract from path: /{bank_id}/...
+        # Resolve bank_id: path takes priority over header.
+        # Path = user's explicit connection endpoint (e.g., /mcp/my-bank/).
+        # X-Bank-Id header = per-request override for multi-bank mode only.
+        bank_id = None
+        bank_id_from_path = False
        new_path = path
-        if not bank_id and path.startswith("/") and len(path) > 1:
+
+        # First, try to extract from path: /{bank_id}/...
+        if path.startswith("/") and len(path) > 1:
            parts = path[1:].split("/", 1)
-            # Don't treat MCP endpoints as bank_ids
-            if parts[0] and parts[0] not in MCP_ENDPOINTS:
-                # First segment looks like a bank_id
+            if parts[0]:
                bank_id = parts[0]
+                bank_id_from_path = True
                new_path = "/" + parts[1] if len(parts) > 1 else "/"

+        # If no path-based bank_id, try X-Bank-Id header (multi-bank mode)
+        if not bank_id:
+            bank_id = self._get_header(scope, "X-Bank-Id")
+
        # Fall back to default bank_id
        if not bank_id:
            bank_id = DEFAULT_BANK_ID
            logger.debug(f"Using default bank_id: {bank_id}")

-        # Set bank_id and api_key context
+        # Select the appropriate MCP app based on how bank_id was provided:
+        # - Path-based bank_id → single-bank app (no bank_id param, scoped tools)
+        # - Header/env bank_id → multi-bank app (bank_id param, all tools)
+        target_app = self.single_bank_app if bank_id_from_path else self.multi_bank_app
+
+        # Set bank_id, api_key, tenant_id, and api_key_id context
        bank_id_token = _current_bank_id.set(bank_id)
        # Store the auth token for tenant extension to validate
        api_key_token = _current_api_key.set(auth_token) if auth_token else None
+        # Store tenant_id and api_key_id from authentication for usage metering
+        tenant_id_token = _current_tenant_id.set(auth_tenant_id) if auth_tenant_id else None
+        api_key_id_token = _current_api_key_id.set(auth_api_key_id) if auth_api_key_id else None
        try:
            new_scope = scope.copy()
            new_scope["path"] = new_path
            # Clear root_path since we're passing directly to the app
            new_scope["root_path"] = ""

-            # Wrap send to rewrite the SSE endpoint URL to include bank_id if using path-based routing
+            # Wrap send to rewrite the SSE endpoint URL to include bank_id if using path-based routing.
+            # Only rewrite SSE (text/event-stream) responses to avoid corrupting tool results
+            # that might contain the literal string "data: /messages".
+            is_sse_response = False
+
            async def send_wrapper(message):
-                if message["type"] == "http.response.body":
+                nonlocal is_sse_response
+                if message["type"] == "http.response.start":
+                    for header_name, header_value in message.get("headers", []):
+                        if header_name == b"content-type" and b"text/event-stream" in header_value:
+                            is_sse_response = True
+                            break
+                if message["type"] == "http.response.body" and bank_id_from_path and is_sse_response:
                    body = message.get("body", b"")
                    if body and b"/messages" in body:
                        # Rewrite /messages to /{bank_id}/messages in SSE endpoint event
@@ -184,11 +339,17 @@ class MCPMiddleware:
                        message = {**message, "body": body}
                await send(message)

-            await self.mcp_app(new_scope, receive, send_wrapper)
+            await target_app(new_scope, receive, send_wrapper)
        finally:
            _current_bank_id.reset(bank_id_token)
            if api_key_token is not None:
                _current_api_key.reset(api_key_token)
+            if tenant_id_token is not None:
+                _current_tenant_id.reset(tenant_id_token)
+            if api_key_id_token is not None:
+                _current_api_key_id.reset(api_key_id_token)
+            if schema_token is not None:
+                _current_schema.reset(schema_token)

    async def _send_error(self, send, status: int, message: str):
        """Send an error response."""
@@ -208,23 +369,19 @@ class MCPMiddleware:
        )


-def create_mcp_app(memory: MemoryEngine):
-    """
-    Create an ASGI app that handles MCP requests.
+def create_mcp_servers(memory: MemoryEngine):
+    """Create multi-bank and single-bank MCP servers and their Starlette apps.

-    Authentication:
-        Set HINDSIGHT_API_MCP_AUTH_TOKEN to require Bearer token authentication.
-        If not set, MCP endpoint is open (for local development).
-
-    Bank ID can be provided via:
-    1. X-Bank-Id header: claude mcp add --transport http hindsight http://localhost:8888/mcp --header "X-Bank-Id: my-bank"
-    2. URL path: /mcp/{bank_id}/
-    3. Environment variable HINDSIGHT_MCP_BANK_ID (fallback, default: "default")
-
-    Args:
-        memory: MemoryEngine instance
+    Returns the servers and apps separately so lifespans can be chained before
+    the middleware wraps the main app.

    Returns:
-        ASGI application
+        Tuple of (multi_bank_server, single_bank_server, multi_bank_app, single_bank_app)
    """
-    return MCPMiddleware(None, memory)
+    multi_bank_server = create_mcp_server(memory, multi_bank=True)
+    multi_bank_app = multi_bank_server.http_app(path="/")
+
+    single_bank_server = create_mcp_server(memory, multi_bank=False)
+    single_bank_app = single_bank_server.http_app(path="/")
+
+    return multi_bank_server, single_bank_server, multi_bank_app, single_bank_app
@@ -4,6 +4,8 @@ Banner display for Hindsight API startup.
 Shows the logo and tagline with gradient colors.
 """

+from .utils import mask_network_location
+
 # Gradient colors: #0074d9 -> #009296
 GRADIENT_START = (0, 116, 217)  # #0074d9
 GRADIENT_END = (0, 146, 150)  # #009296
@@ -83,11 +85,14 @@ def print_startup_info(
    embeddings_provider: str,
    reranker_provider: str,
    mcp_enabled: bool = False,
+    version: str | None = None,
 ):
    """Print styled startup information."""
    print(color_start("Starting Hindsight API..."))
+    if version:
+        print(f"  {dim('Version:')} {color(f'v{version}', 0.1)}")
    print(f"  {dim('URL:')} {color(f'http://{host}:{port}', 0.2)}")
-    print(f"  {dim('Database:')} {color(database_url, 0.4)}")
+    print(f"  {dim('Database:')} {color(mask_network_location(database_url), 0.4)}")
    print(f"  {dim('LLM:')} {color(f'{llm_provider} / {llm_model}', 0.6)}")
    print(f"  {dim('Embeddings:')} {color(embeddings_provider, 0.8)}")
    print(f"  {dim('Reranker:')} {color(reranker_provider, 1.0)}")
@@ -8,8 +8,9 @@ import json
 import logging
 import os
 import sys
-from dataclasses import dataclass
+from dataclasses import dataclass, field, fields
 from datetime import datetime, timezone
+from typing import Any

 from dotenv import find_dotenv, load_dotenv

@@ -18,6 +19,103 @@ load_dotenv(find_dotenv(usecwd=True), override=True)

 logger = logging.getLogger(__name__)

+
+class ConfigFieldAccessError(AttributeError):
+    """Raised when trying to access a bank-configurable field from global config."""
+
+    pass
+
+
+class StaticConfigProxy:
+    """
+    Proxy that wraps HindsightConfig and only allows access to static (non-configurable) fields.
+
+    Raises ConfigFieldAccessError when trying to access configurable fields that vary per-bank.
+    Forces developers to use get_resolved_config(bank_id, context) for bank-specific settings.
+    """
+
+    def __init__(self, config: "HindsightConfig"):
+        object.__setattr__(self, "_config", config)
+        object.__setattr__(self, "_configurable_fields", HindsightConfig.get_configurable_fields())
+
+    def __getattribute__(self, name: str):
+        if name.startswith("_"):
+            return object.__getattribute__(self, name)
+
+        configurable_fields = object.__getattribute__(self, "_configurable_fields")
+        if name in configurable_fields:
+            raise ConfigFieldAccessError(
+                f"Field '{name}' is bank-configurable and cannot be accessed from global config. "
+                f"Use ConfigResolver.resolve_full_config(bank_id, context) to get bank-specific config. "
+                f"This prevents accidentally using global defaults when bank-specific overrides exist."
+            )
+
+        config = object.__getattribute__(self, "_config")
+        return getattr(config, name)
+
+    def __setattr__(self, name: str, value):
+        raise AttributeError("Config is read-only. Modifications must go through ConfigResolver.")
+
+
+# Configuration field markers for hierarchical configuration
+def hierarchical(default_value):
+    """
+    Mark a config field as hierarchical (can be overridden per-tenant/bank).
+
+    Hierarchical fields can be customized at the tenant or bank level via database
+    configuration. Examples: LLM settings, retention parameters, retrieval settings.
+    """
+    return field(default=default_value, metadata={"hierarchical": True})
+
+
+def static(default_value):
+    """
+    Mark a config field as static (server-level only, cannot be overridden).
+
+    Static fields are infrastructure-level settings that affect the entire server
+    and cannot vary per tenant or bank. Examples: database URL, API port, worker settings.
+    """
+    return field(default=default_value, metadata={"hierarchical": False})
+
+
+# Configuration key normalization utilities
+def normalize_config_key(key: str) -> str:
+    """
+    Convert environment variable format to Python field name format.
+
+    Examples:
+        HINDSIGHT_API_LLM_PROVIDER -> llm_provider
+        LLM_MODEL -> llm_model
+        llm_model -> llm_model (already normalized)
+
+    Args:
+        key: Environment variable name or Python field name
+
+    Returns:
+        Normalized Python field name (lowercase snake_case)
+    """
+    if key.startswith("HINDSIGHT_API_"):
+        key = key[len("HINDSIGHT_API_") :]
+    return key.lower()
+
+
+def normalize_config_dict(config: dict[str, Any]) -> dict[str, Any]:
+    """
+    Normalize all keys in a config dict to Python field names.
+
+    Allows users to provide config overrides in either format:
+    - Python field format: {"llm_provider": "openai"}
+    - Env var format: {"HINDSIGHT_API_LLM_PROVIDER": "openai"}
+
+    Args:
+        config: Dict with env var or Python field names as keys
+
+    Returns:
+        Dict with all keys normalized to Python field names
+    """
+    return {normalize_config_key(k): v for k, v in config.items()}
+
+
 # Environment variable names
 ENV_DATABASE_URL = "HINDSIGHT_API_DATABASE_URL"
 ENV_DATABASE_SCHEMA = "HINDSIGHT_API_DATABASE_SCHEMA"
@@ -66,27 +164,40 @@ ENV_CONSOLIDATION_LLM_TIMEOUT = "HINDSIGHT_API_CONSOLIDATION_LLM_TIMEOUT"
 ENV_EMBEDDINGS_PROVIDER = "HINDSIGHT_API_EMBEDDINGS_PROVIDER"
 ENV_EMBEDDINGS_LOCAL_MODEL = "HINDSIGHT_API_EMBEDDINGS_LOCAL_MODEL"
 ENV_EMBEDDINGS_LOCAL_FORCE_CPU = "HINDSIGHT_API_EMBEDDINGS_LOCAL_FORCE_CPU"
+ENV_EMBEDDINGS_LOCAL_TRUST_REMOTE_CODE = "HINDSIGHT_API_EMBEDDINGS_LOCAL_TRUST_REMOTE_CODE"
 ENV_EMBEDDINGS_TEI_URL = "HINDSIGHT_API_EMBEDDINGS_TEI_URL"
 ENV_EMBEDDINGS_OPENAI_API_KEY = "HINDSIGHT_API_EMBEDDINGS_OPENAI_API_KEY"
 ENV_EMBEDDINGS_OPENAI_MODEL = "HINDSIGHT_API_EMBEDDINGS_OPENAI_MODEL"
 ENV_EMBEDDINGS_OPENAI_BASE_URL = "HINDSIGHT_API_EMBEDDINGS_OPENAI_BASE_URL"

-ENV_COHERE_API_KEY = "HINDSIGHT_API_COHERE_API_KEY"
+# Cohere configuration (separate for embeddings and reranker)
+ENV_EMBEDDINGS_COHERE_API_KEY = "HINDSIGHT_API_EMBEDDINGS_COHERE_API_KEY"
 ENV_EMBEDDINGS_COHERE_MODEL = "HINDSIGHT_API_EMBEDDINGS_COHERE_MODEL"
 ENV_EMBEDDINGS_COHERE_BASE_URL = "HINDSIGHT_API_EMBEDDINGS_COHERE_BASE_URL"
+ENV_RERANKER_COHERE_API_KEY = "HINDSIGHT_API_RERANKER_COHERE_API_KEY"
 ENV_RERANKER_COHERE_MODEL = "HINDSIGHT_API_RERANKER_COHERE_MODEL"
 ENV_RERANKER_COHERE_BASE_URL = "HINDSIGHT_API_RERANKER_COHERE_BASE_URL"

-# LiteLLM gateway configuration (for embeddings and reranker via LiteLLM proxy)
+# Deprecated: Legacy shared Cohere API key (for backward compatibility)
+ENV_COHERE_API_KEY = "HINDSIGHT_API_COHERE_API_KEY"
+
+# LiteLLM configuration (separate for embeddings and reranker)
+ENV_EMBEDDINGS_LITELLM_API_BASE = "HINDSIGHT_API_EMBEDDINGS_LITELLM_API_BASE"
+ENV_EMBEDDINGS_LITELLM_API_KEY = "HINDSIGHT_API_EMBEDDINGS_LITELLM_API_KEY"
+ENV_EMBEDDINGS_LITELLM_MODEL = "HINDSIGHT_API_EMBEDDINGS_LITELLM_MODEL"
+ENV_RERANKER_LITELLM_API_BASE = "HINDSIGHT_API_RERANKER_LITELLM_API_BASE"
+ENV_RERANKER_LITELLM_API_KEY = "HINDSIGHT_API_RERANKER_LITELLM_API_KEY"
+ENV_RERANKER_LITELLM_MODEL = "HINDSIGHT_API_RERANKER_LITELLM_MODEL"
+
+# Deprecated: Legacy shared LiteLLM config (for backward compatibility)
 ENV_LITELLM_API_BASE = "HINDSIGHT_API_LITELLM_API_BASE"
 ENV_LITELLM_API_KEY = "HINDSIGHT_API_LITELLM_API_KEY"
-ENV_EMBEDDINGS_LITELLM_MODEL = "HINDSIGHT_API_EMBEDDINGS_LITELLM_MODEL"
-ENV_RERANKER_LITELLM_MODEL = "HINDSIGHT_API_RERANKER_LITELLM_MODEL"

 ENV_RERANKER_PROVIDER = "HINDSIGHT_API_RERANKER_PROVIDER"
 ENV_RERANKER_LOCAL_MODEL = "HINDSIGHT_API_RERANKER_LOCAL_MODEL"
 ENV_RERANKER_LOCAL_FORCE_CPU = "HINDSIGHT_API_RERANKER_LOCAL_FORCE_CPU"
 ENV_RERANKER_LOCAL_MAX_CONCURRENT = "HINDSIGHT_API_RERANKER_LOCAL_MAX_CONCURRENT"
+ENV_RERANKER_LOCAL_TRUST_REMOTE_CODE = "HINDSIGHT_API_RERANKER_LOCAL_TRUST_REMOTE_CODE"
 ENV_RERANKER_TEI_URL = "HINDSIGHT_API_RERANKER_TEI_URL"
 ENV_RERANKER_TEI_BATCH_SIZE = "HINDSIGHT_API_RERANKER_TEI_BATCH_SIZE"
 ENV_RERANKER_TEI_MAX_CONCURRENT = "HINDSIGHT_API_RERANKER_TEI_MAX_CONCURRENT"
@@ -94,12 +205,17 @@ ENV_RERANKER_MAX_CANDIDATES = "HINDSIGHT_API_RERANKER_MAX_CANDIDATES"
 ENV_RERANKER_FLASHRANK_MODEL = "HINDSIGHT_API_RERANKER_FLASHRANK_MODEL"
 ENV_RERANKER_FLASHRANK_CACHE_DIR = "HINDSIGHT_API_RERANKER_FLASHRANK_CACHE_DIR"

+ENV_VECTOR_EXTENSION = "HINDSIGHT_API_VECTOR_EXTENSION"
+ENV_TEXT_SEARCH_EXTENSION = "HINDSIGHT_API_TEXT_SEARCH_EXTENSION"
+
 ENV_HOST = "HINDSIGHT_API_HOST"
 ENV_PORT = "HINDSIGHT_API_PORT"
+ENV_BASE_PATH = "HINDSIGHT_API_BASE_PATH"
 ENV_LOG_LEVEL = "HINDSIGHT_API_LOG_LEVEL"
 ENV_LOG_FORMAT = "HINDSIGHT_API_LOG_FORMAT"
 ENV_WORKERS = "HINDSIGHT_API_WORKERS"
 ENV_MCP_ENABLED = "HINDSIGHT_API_MCP_ENABLED"
+ENV_ENABLE_BANK_CONFIG_API = "HINDSIGHT_API_ENABLE_BANK_CONFIG_API"
 ENV_GRAPH_RETRIEVER = "HINDSIGHT_API_GRAPH_RETRIEVER"
 ENV_MPFP_TOP_K_NEIGHBORS = "HINDSIGHT_API_MPFP_TOP_K_NEIGHBORS"
 ENV_RECALL_MAX_CONCURRENT = "HINDSIGHT_API_RECALL_MAX_CONCURRENT"
@@ -108,6 +224,13 @@ ENV_MCP_LOCAL_BANK_ID = "HINDSIGHT_API_MCP_LOCAL_BANK_ID"
 ENV_MCP_INSTRUCTIONS = "HINDSIGHT_API_MCP_INSTRUCTIONS"
 ENV_MENTAL_MODEL_REFRESH_CONCURRENCY = "HINDSIGHT_API_MENTAL_MODEL_REFRESH_CONCURRENCY"

+# OpenTelemetry tracing configuration
+ENV_OTEL_TRACES_ENABLED = "HINDSIGHT_API_OTEL_TRACES_ENABLED"
+ENV_OTEL_EXPORTER_OTLP_ENDPOINT = "HINDSIGHT_API_OTEL_EXPORTER_OTLP_ENDPOINT"
+ENV_OTEL_EXPORTER_OTLP_HEADERS = "HINDSIGHT_API_OTEL_EXPORTER_OTLP_HEADERS"
+ENV_OTEL_SERVICE_NAME = "HINDSIGHT_API_OTEL_SERVICE_NAME"
+ENV_OTEL_DEPLOYMENT_ENVIRONMENT = "HINDSIGHT_API_OTEL_DEPLOYMENT_ENVIRONMENT"
+
 # Vertex AI configuration
 ENV_LLM_VERTEXAI_PROJECT_ID = "HINDSIGHT_API_LLM_VERTEXAI_PROJECT_ID"
 ENV_LLM_VERTEXAI_REGION = "HINDSIGHT_API_LLM_VERTEXAI_REGION"
@@ -154,7 +277,21 @@ ENV_REFLECT_MAX_ITERATIONS = "HINDSIGHT_API_REFLECT_MAX_ITERATIONS"
 DEFAULT_DATABASE_URL = "pg0"
 DEFAULT_DATABASE_SCHEMA = "public"
 DEFAULT_LLM_PROVIDER = "openai"
-DEFAULT_LLM_MODEL = "gpt-5-mini"
+
+# Provider-specific default models
+PROVIDER_DEFAULT_MODELS = {
+    "openai": "o3-mini",
+    "anthropic": "claude-haiku-4-5-20251001",
+    "gemini": "gemini-2.5-flash",
+    "groq": "openai/gpt-oss-120b",
+    "ollama": "gemma3:12b",
+    "lmstudio": "local-model",
+    "vertexai": "gemini-2.0-flash-001",
+    "openai-codex": "gpt-5.2-codex",
+    "claude-code": "claude-sonnet-4-5-20250929",
+    "mock": "mock-model",
+}
+DEFAULT_LLM_MODEL = "o3-mini"  # Fallback if provider not in table
 DEFAULT_LLM_MAX_CONCURRENT = 32
 DEFAULT_LLM_MAX_RETRIES = 10  # Max retry attempts for LLM API calls
 DEFAULT_LLM_INITIAL_BACKOFF = 1.0  # Initial backoff in seconds for retry exponential backoff
@@ -169,6 +306,7 @@ DEFAULT_LLM_VERTEXAI_SERVICE_ACCOUNT_KEY = None  # Optional, uses ADC if not set
 DEFAULT_EMBEDDINGS_PROVIDER = "local"
 DEFAULT_EMBEDDINGS_LOCAL_MODEL = "BAAI/bge-small-en-v1.5"
 DEFAULT_EMBEDDINGS_LOCAL_FORCE_CPU = False  # Force CPU mode for local embeddings (avoids MPS/XPC issues on macOS)
+DEFAULT_EMBEDDINGS_LOCAL_TRUST_REMOTE_CODE = False  # Security: disabled by default, required for some models
 DEFAULT_EMBEDDINGS_OPENAI_MODEL = "text-embedding-3-small"
 DEFAULT_EMBEDDING_DIMENSION = 384

@@ -176,6 +314,9 @@ DEFAULT_RERANKER_PROVIDER = "local"
 DEFAULT_RERANKER_LOCAL_MODEL = "cross-encoder/ms-marco-MiniLM-L-6-v2"
 DEFAULT_RERANKER_LOCAL_FORCE_CPU = False  # Force CPU mode for local reranker (avoids MPS/XPC issues on macOS)
 DEFAULT_RERANKER_LOCAL_MAX_CONCURRENT = 4  # Limit concurrent CPU-bound reranking to prevent thrashing
+DEFAULT_RERANKER_LOCAL_TRUST_REMOTE_CODE = (
+    False  # Security: disabled by default, required for some models like jina-reranker-v2
+)
 DEFAULT_RERANKER_TEI_BATCH_SIZE = 128
 DEFAULT_RERANKER_TEI_MAX_CONCURRENT = 8
 DEFAULT_RERANKER_MAX_CANDIDATES = 300
@@ -185,6 +326,12 @@ DEFAULT_RERANKER_FLASHRANK_CACHE_DIR = None  # Use default cache directory
 DEFAULT_EMBEDDINGS_COHERE_MODEL = "embed-english-v3.0"
 DEFAULT_RERANKER_COHERE_MODEL = "rerank-english-v3.0"

+# Vector extension (pgvector vs vchord)
+DEFAULT_VECTOR_EXTENSION = "pgvector"  # Options: "pgvector", "vchord"
+
+# Text search extension (native PostgreSQL vs vchord BM25)
+DEFAULT_TEXT_SEARCH_EXTENSION = "native"  # Options: "native", "vchord"
+
 # LiteLLM defaults
 DEFAULT_LITELLM_API_BASE = "http://localhost:4000"
 DEFAULT_EMBEDDINGS_LITELLM_MODEL = "text-embedding-3-small"
@@ -192,10 +339,12 @@ DEFAULT_RERANKER_LITELLM_MODEL = "cohere/rerank-english-v3.0"

 DEFAULT_HOST = "0.0.0.0"
 DEFAULT_PORT = 8888
+DEFAULT_BASE_PATH = ""  # Empty string = root path
 DEFAULT_LOG_LEVEL = "info"
 DEFAULT_LOG_FORMAT = "text"  # Options: "text", "json"
 DEFAULT_WORKERS = 1
 DEFAULT_MCP_ENABLED = True
+DEFAULT_ENABLE_BANK_CONFIG_API = False  # Disabled by default for security
 DEFAULT_GRAPH_RETRIEVER = "link_expansion"  # Options: "link_expansion", "mpfp", "bfs"
 DEFAULT_MPFP_TOP_K_NEIGHBORS = 20  # Fan-out limit per node in MPFP graph traversal
 DEFAULT_RECALL_MAX_CONCURRENT = 32  # Max concurrent recall operations per worker
@@ -237,6 +386,11 @@ DEFAULT_WORKER_CONSOLIDATION_MAX_SLOTS = 2  # Max concurrent consolidation tasks
 # Reflect agent settings
 DEFAULT_REFLECT_MAX_ITERATIONS = 10  # Max tool call iterations before forcing response

+# OpenTelemetry tracing configuration
+DEFAULT_OTEL_TRACES_ENABLED = False  # Disabled by default for backward compatibility
+DEFAULT_OTEL_SERVICE_NAME = "hindsight-api"
+DEFAULT_OTEL_DEPLOYMENT_ENVIRONMENT = "development"
+
 # Default MCP tool descriptions (can be customized via env vars)
 DEFAULT_MCP_RETAIN_DESCRIPTION = """Store important information to long-term memory.

@@ -303,6 +457,11 @@ def _validate_extraction_mode(mode: str) -> str:
    return mode_lower


+def _get_default_model_for_provider(provider: str) -> str:
+    """Get the default model for a given provider."""
+    return PROVIDER_DEFAULT_MODELS.get(provider.lower(), DEFAULT_LLM_MODEL)
+
+
@dataclass
 class HindsightConfig:
    """Configuration container for Hindsight API."""
@@ -310,6 +469,8 @@ class HindsightConfig:
    # Database
    database_url: str
    database_schema: str
+    vector_extension: str  # "pgvector" or "vchord"
+    text_search_extension: str  # "native" or "vchord"

    # LLM (default, used as fallback for per-operation config)
    llm_provider: str
@@ -362,27 +523,41 @@ class HindsightConfig:
    embeddings_provider: str
    embeddings_local_model: str
    embeddings_local_force_cpu: bool
+    embeddings_local_trust_remote_code: bool
    embeddings_tei_url: str | None
    embeddings_openai_base_url: str | None
+    embeddings_cohere_api_key: str | None
+    embeddings_cohere_model: str
    embeddings_cohere_base_url: str | None
+    embeddings_litellm_api_base: str
+    embeddings_litellm_api_key: str | None
+    embeddings_litellm_model: str

    # Reranker
    reranker_provider: str
    reranker_local_model: str
    reranker_local_force_cpu: bool
    reranker_local_max_concurrent: int
+    reranker_local_trust_remote_code: bool
    reranker_tei_url: str | None
    reranker_tei_batch_size: int
    reranker_tei_max_concurrent: int
    reranker_max_candidates: int
+    reranker_cohere_api_key: str | None
+    reranker_cohere_model: str
    reranker_cohere_base_url: str | None
+    reranker_litellm_api_base: str
+    reranker_litellm_api_key: str | None
+    reranker_litellm_model: str

    # Server
    host: str
    port: int
+    base_path: str
    log_level: str
    log_format: str
    mcp_enabled: bool
+    enable_bank_config_api: bool

    # Recall
    graph_retriever: str
@@ -428,17 +603,146 @@ class HindsightConfig:
    # Reflect agent settings
    reflect_max_iterations: int

+    # OpenTelemetry tracing configuration
+    otel_traces_enabled: bool
+    otel_exporter_otlp_endpoint: str | None
+    otel_exporter_otlp_headers: str | None
+    otel_service_name: str
+    otel_deployment_environment: str
+
+    # Class-level sets for configuration categorization
+
+    # CREDENTIAL_FIELDS: Never exposed via API, never configurable per-tenant/bank
+    _CREDENTIAL_FIELDS = {
+        # API Keys
+        "llm_api_key",
+        "retain_llm_api_key",
+        "reflect_llm_api_key",
+        "consolidation_llm_api_key",
+        # Base URLs (could expose infrastructure)
+        "llm_base_url",
+        "retain_llm_base_url",
+        "reflect_llm_base_url",
+        "consolidation_llm_base_url",
+        "embeddings_tei_base_url",
+        "reranker_tei_base_url",
+        "reranker_cohere_base_url",
+        # Service Account Keys
+        "llm_vertexai_service_account_key",
+    }
+
+    # CONFIGURABLE_FIELDS: Safe behavioral settings that can be customized per-tenant/bank
+    # These fields are manually tagged as safe to expose and modify.
+    # Excludes credentials, infrastructure config, provider/model selection, and performance tuning.
+    _CONFIGURABLE_FIELDS = {
+        # Retention settings (behavioral)
+        "retain_chunk_size",
+        "retain_extraction_mode",
+        "retain_custom_instructions",
+        # Consolidation settings
+        "enable_observations",
+    }
+
+    @classmethod
+    def get_configurable_fields(cls) -> set[str]:
+        """
+        Get set of field names that are configurable per-tenant/bank via API.
+
+        Configurable fields are manually tagged behavioral settings that are safe
+        to expose and modify (e.g., retain_chunk_size, custom_instructions).
+        Excludes credentials, infrastructure config, and provider/model selection.
+
+        Returns:
+            Set of configurable field names
+        """
+        return cls._CONFIGURABLE_FIELDS.copy()
+
+    @classmethod
+    def get_credential_fields(cls) -> set[str]:
+        """
+        Get set of field names that are credentials (NEVER exposed via API).
+
+        Credential fields include API keys, base URLs, and service account keys.
+        These must never be returned in API responses or accepted in updates.
+
+        Returns:
+            Set of credential field names
+        """
+        return cls._CREDENTIAL_FIELDS.copy()
+
+    @classmethod
+    def get_hierarchical_fields(cls) -> set[str]:
+        """
+        DEPRECATED: Use get_configurable_fields() instead.
+
+        Kept for backward compatibility during migration.
+        """
+        return cls.get_configurable_fields()
+
+    @classmethod
+    def get_static_fields(cls) -> set[str]:
+        """
+        Get set of field names that are static (server-level only).
+
+        Static fields are infrastructure-level settings that cannot vary
+        per tenant or bank. These include database config, API port, worker settings, etc.
+        Also includes credential fields which are never configurable.
+
+        Returns:
+            Set of static field names
+        """
+        # Get all field names from dataclass
+        all_fields = {f.name for f in fields(cls)}
+        # Static fields = all fields - configurable fields
+        return all_fields - cls._CONFIGURABLE_FIELDS
+
+    def validate(self) -> None:
+        """Validate configuration values and raise errors for invalid combinations."""
+        # Validate vector_extension
+        valid_extensions = ("pgvector", "vchord")
+        if self.vector_extension not in valid_extensions:
+            raise ValueError(
+                f"Invalid vector_extension: {self.vector_extension}. Must be one of: {', '.join(valid_extensions)}"
+            )
+
+        # Validate text_search_extension
+        valid_text_search = ("native", "vchord")
+        if self.text_search_extension not in valid_text_search:
+            raise ValueError(
+                f"Invalid text_search_extension: {self.text_search_extension}. Must be one of: {', '.join(valid_text_search)}"
+            )
+
+        # RETAIN_MAX_COMPLETION_TOKENS must be greater than RETAIN_CHUNK_SIZE
+        # to ensure the LLM has enough output capacity to extract facts from chunks
+        if self.retain_max_completion_tokens <= self.retain_chunk_size:
+            raise ValueError(
+                f"Invalid configuration: HINDSIGHT_API_RETAIN_MAX_COMPLETION_TOKENS "
+                f"({self.retain_max_completion_tokens}) must be greater than "
+                f"HINDSIGHT_API_RETAIN_CHUNK_SIZE ({self.retain_chunk_size}). "
+                f"\n\nYou have two options to fix this:"
+                f"\n  1. Increase HINDSIGHT_API_RETAIN_MAX_COMPLETION_TOKENS to a value > {self.retain_chunk_size}"
+                f"\n  2. Use a model that supports at least {self.retain_max_completion_tokens} output tokens"
+                f"\n     (current model: {self.retain_llm_model or self.llm_model}, "
+                f"provider: {self.retain_llm_provider or self.llm_provider})"
+            )
+
    @classmethod
    def from_env(cls) -> "HindsightConfig":
        """Create configuration from environment variables."""
-        return cls(
+        # Get provider first to determine default model
+        llm_provider = os.getenv(ENV_LLM_PROVIDER, DEFAULT_LLM_PROVIDER)
+        llm_model = os.getenv(ENV_LLM_MODEL) or _get_default_model_for_provider(llm_provider)
+
+        config = cls(
            # Database
            database_url=os.getenv(ENV_DATABASE_URL, DEFAULT_DATABASE_URL),
            database_schema=os.getenv(ENV_DATABASE_SCHEMA, DEFAULT_DATABASE_SCHEMA),
+            vector_extension=os.getenv(ENV_VECTOR_EXTENSION, DEFAULT_VECTOR_EXTENSION).lower(),
+            text_search_extension=os.getenv(ENV_TEXT_SEARCH_EXTENSION, DEFAULT_TEXT_SEARCH_EXTENSION).lower(),
            # LLM
-            llm_provider=os.getenv(ENV_LLM_PROVIDER, DEFAULT_LLM_PROVIDER),
+            llm_provider=llm_provider,
            llm_api_key=os.getenv(ENV_LLM_API_KEY),
-            llm_model=os.getenv(ENV_LLM_MODEL, DEFAULT_LLM_MODEL),
+            llm_model=llm_model,
            llm_base_url=os.getenv(ENV_LLM_BASE_URL) or None,
            llm_max_concurrent=int(os.getenv(ENV_LLM_MAX_CONCURRENT, str(DEFAULT_LLM_MAX_CONCURRENT))),
            llm_max_retries=int(os.getenv(ENV_LLM_MAX_RETRIES, str(DEFAULT_LLM_MAX_RETRIES))),
@@ -453,7 +757,12 @@ class HindsightConfig:
            # Per-operation LLM config (None = use default)
            retain_llm_provider=os.getenv(ENV_RETAIN_LLM_PROVIDER) or None,
            retain_llm_api_key=os.getenv(ENV_RETAIN_LLM_API_KEY) or None,
-            retain_llm_model=os.getenv(ENV_RETAIN_LLM_MODEL) or None,
+            retain_llm_model=os.getenv(ENV_RETAIN_LLM_MODEL)
+            or (
+                _get_default_model_for_provider(os.getenv(ENV_RETAIN_LLM_PROVIDER))
+                if os.getenv(ENV_RETAIN_LLM_PROVIDER)
+                else None
+            ),
            retain_llm_base_url=os.getenv(ENV_RETAIN_LLM_BASE_URL) or None,
            retain_llm_max_concurrent=int(os.getenv(ENV_RETAIN_LLM_MAX_CONCURRENT))
            if os.getenv(ENV_RETAIN_LLM_MAX_CONCURRENT)
@@ -470,7 +779,12 @@ class HindsightConfig:
            retain_llm_timeout=float(os.getenv(ENV_RETAIN_LLM_TIMEOUT)) if os.getenv(ENV_RETAIN_LLM_TIMEOUT) else None,
            reflect_llm_provider=os.getenv(ENV_REFLECT_LLM_PROVIDER) or None,
            reflect_llm_api_key=os.getenv(ENV_REFLECT_LLM_API_KEY) or None,
-            reflect_llm_model=os.getenv(ENV_REFLECT_LLM_MODEL) or None,
+            reflect_llm_model=os.getenv(ENV_REFLECT_LLM_MODEL)
+            or (
+                _get_default_model_for_provider(os.getenv(ENV_REFLECT_LLM_PROVIDER))
+                if os.getenv(ENV_REFLECT_LLM_PROVIDER)
+                else None
+            ),
            reflect_llm_base_url=os.getenv(ENV_REFLECT_LLM_BASE_URL) or None,
            reflect_llm_max_concurrent=int(os.getenv(ENV_REFLECT_LLM_MAX_CONCURRENT))
            if os.getenv(ENV_REFLECT_LLM_MAX_CONCURRENT)
@@ -489,7 +803,12 @@ class HindsightConfig:
            else None,
            consolidation_llm_provider=os.getenv(ENV_CONSOLIDATION_LLM_PROVIDER) or None,
            consolidation_llm_api_key=os.getenv(ENV_CONSOLIDATION_LLM_API_KEY) or None,
-            consolidation_llm_model=os.getenv(ENV_CONSOLIDATION_LLM_MODEL) or None,
+            consolidation_llm_model=os.getenv(ENV_CONSOLIDATION_LLM_MODEL)
+            or (
+                _get_default_model_for_provider(os.getenv(ENV_CONSOLIDATION_LLM_PROVIDER))
+                if os.getenv(ENV_CONSOLIDATION_LLM_PROVIDER)
+                else None
+            ),
            consolidation_llm_base_url=os.getenv(ENV_CONSOLIDATION_LLM_BASE_URL) or None,
            consolidation_llm_max_concurrent=int(os.getenv(ENV_CONSOLIDATION_LLM_MAX_CONCURRENT))
            if os.getenv(ENV_CONSOLIDATION_LLM_MAX_CONCURRENT)
@@ -513,9 +832,21 @@ class HindsightConfig:
                ENV_EMBEDDINGS_LOCAL_FORCE_CPU, str(DEFAULT_EMBEDDINGS_LOCAL_FORCE_CPU)
            ).lower()
            in ("true", "1"),
+            embeddings_local_trust_remote_code=os.getenv(
+                ENV_EMBEDDINGS_LOCAL_TRUST_REMOTE_CODE, str(DEFAULT_EMBEDDINGS_LOCAL_TRUST_REMOTE_CODE)
+            ).lower()
+            in ("true", "1"),
            embeddings_tei_url=os.getenv(ENV_EMBEDDINGS_TEI_URL),
            embeddings_openai_base_url=os.getenv(ENV_EMBEDDINGS_OPENAI_BASE_URL) or None,
+            # Cohere embeddings (with backward-compatible fallback to shared API key)
+            embeddings_cohere_api_key=os.getenv(ENV_EMBEDDINGS_COHERE_API_KEY) or os.getenv(ENV_COHERE_API_KEY),
+            embeddings_cohere_model=os.getenv(ENV_EMBEDDINGS_COHERE_MODEL, DEFAULT_EMBEDDINGS_COHERE_MODEL),
            embeddings_cohere_base_url=os.getenv(ENV_EMBEDDINGS_COHERE_BASE_URL) or None,
+            # LiteLLM embeddings (with backward-compatible fallback to shared config)
+            embeddings_litellm_api_base=os.getenv(ENV_EMBEDDINGS_LITELLM_API_BASE)
+            or os.getenv(ENV_LITELLM_API_BASE, DEFAULT_LITELLM_API_BASE),
+            embeddings_litellm_api_key=os.getenv(ENV_EMBEDDINGS_LITELLM_API_KEY) or os.getenv(ENV_LITELLM_API_KEY),
+            embeddings_litellm_model=os.getenv(ENV_EMBEDDINGS_LITELLM_MODEL, DEFAULT_EMBEDDINGS_LITELLM_MODEL),
            # Reranker
            reranker_provider=os.getenv(ENV_RERANKER_PROVIDER, DEFAULT_RERANKER_PROVIDER),
            reranker_local_model=os.getenv(ENV_RERANKER_LOCAL_MODEL, DEFAULT_RERANKER_LOCAL_MODEL),
@@ -526,19 +857,34 @@ class HindsightConfig:
            reranker_local_max_concurrent=int(
                os.getenv(ENV_RERANKER_LOCAL_MAX_CONCURRENT, str(DEFAULT_RERANKER_LOCAL_MAX_CONCURRENT))
            ),
+            reranker_local_trust_remote_code=os.getenv(
+                ENV_RERANKER_LOCAL_TRUST_REMOTE_CODE, str(DEFAULT_RERANKER_LOCAL_TRUST_REMOTE_CODE)
+            ).lower()
+            in ("true", "1"),
            reranker_tei_url=os.getenv(ENV_RERANKER_TEI_URL),
            reranker_tei_batch_size=int(os.getenv(ENV_RERANKER_TEI_BATCH_SIZE, str(DEFAULT_RERANKER_TEI_BATCH_SIZE))),
            reranker_tei_max_concurrent=int(
                os.getenv(ENV_RERANKER_TEI_MAX_CONCURRENT, str(DEFAULT_RERANKER_TEI_MAX_CONCURRENT))
            ),
            reranker_max_candidates=int(os.getenv(ENV_RERANKER_MAX_CANDIDATES, str(DEFAULT_RERANKER_MAX_CANDIDATES))),
+            # Cohere reranker (with backward-compatible fallback to shared API key)
+            reranker_cohere_api_key=os.getenv(ENV_RERANKER_COHERE_API_KEY) or os.getenv(ENV_COHERE_API_KEY),
+            reranker_cohere_model=os.getenv(ENV_RERANKER_COHERE_MODEL, DEFAULT_RERANKER_COHERE_MODEL),
            reranker_cohere_base_url=os.getenv(ENV_RERANKER_COHERE_BASE_URL) or None,
+            # LiteLLM reranker (with backward-compatible fallback to shared config)
+            reranker_litellm_api_base=os.getenv(ENV_RERANKER_LITELLM_API_BASE)
+            or os.getenv(ENV_LITELLM_API_BASE, DEFAULT_LITELLM_API_BASE),
+            reranker_litellm_api_key=os.getenv(ENV_RERANKER_LITELLM_API_KEY) or os.getenv(ENV_LITELLM_API_KEY),
+            reranker_litellm_model=os.getenv(ENV_RERANKER_LITELLM_MODEL, DEFAULT_RERANKER_LITELLM_MODEL),
            # Server
            host=os.getenv(ENV_HOST, DEFAULT_HOST),
            port=int(os.getenv(ENV_PORT, DEFAULT_PORT)),
+            base_path=os.getenv(ENV_BASE_PATH, DEFAULT_BASE_PATH),
            log_level=os.getenv(ENV_LOG_LEVEL, DEFAULT_LOG_LEVEL),
            log_format=os.getenv(ENV_LOG_FORMAT, DEFAULT_LOG_FORMAT).lower(),
            mcp_enabled=os.getenv(ENV_MCP_ENABLED, str(DEFAULT_MCP_ENABLED)).lower() == "true",
+            enable_bank_config_api=os.getenv(ENV_ENABLE_BANK_CONFIG_API, str(DEFAULT_ENABLE_BANK_CONFIG_API)).lower()
+            == "true",
            # Recall
            graph_retriever=os.getenv(ENV_GRAPH_RETRIEVER, DEFAULT_GRAPH_RETRIEVER),
            mpfp_top_k_neighbors=int(os.getenv(ENV_MPFP_TOP_K_NEIGHBORS, str(DEFAULT_MPFP_TOP_K_NEIGHBORS))),
@@ -592,7 +938,16 @@ class HindsightConfig:
            ),
            # Reflect agent settings
            reflect_max_iterations=int(os.getenv(ENV_REFLECT_MAX_ITERATIONS, str(DEFAULT_REFLECT_MAX_ITERATIONS))),
+            # OpenTelemetry tracing configuration
+            otel_traces_enabled=os.getenv(ENV_OTEL_TRACES_ENABLED, str(DEFAULT_OTEL_TRACES_ENABLED)).lower()
+            in ("true", "1", "yes"),
+            otel_exporter_otlp_endpoint=os.getenv(ENV_OTEL_EXPORTER_OTLP_ENDPOINT) or None,
+            otel_exporter_otlp_headers=os.getenv(ENV_OTEL_EXPORTER_OTLP_HEADERS) or None,
+            otel_service_name=os.getenv(ENV_OTEL_SERVICE_NAME, DEFAULT_OTEL_SERVICE_NAME),
+            otel_deployment_environment=os.getenv(ENV_OTEL_DEPLOYMENT_ENVIRONMENT, DEFAULT_OTEL_DEPLOYMENT_ENVIRONMENT),
        )
+        config.validate()
+        return config

    def get_llm_base_url(self) -> str:
        """Get the LLM base URL, with provider-specific defaults."""
@@ -670,8 +1025,35 @@ class HindsightConfig:
 _config_cache: HindsightConfig | None = None


-def get_config() -> HindsightConfig:
-    """Get the cached configuration, loading from environment on first call."""
+def get_config() -> StaticConfigProxy:
+    """
+    Get global configuration with ONLY static (non-configurable) fields accessible.
+
+    This returns a proxy that prevents access to bank-configurable fields
+    (like enable_observations, retain_chunk_size, etc.).
+
+    For bank-specific configuration, use:
+        config_resolver.resolve_full_config(bank_id, context)
+
+    This design prevents accidentally using global defaults when bank-specific
+    overrides exist.
+
+    Returns:
+        StaticConfigProxy that only exposes static infrastructure fields
+
+    Raises:
+        ConfigFieldAccessError: If you try to access a bank-configurable field
+    """
+    return StaticConfigProxy(_get_raw_config())
+
+
+def _get_raw_config() -> HindsightConfig:
+    """
+    Get raw config (internal use only).
+
+    INTERNAL USE ONLY. Do not use this directly in application code.
+    Use get_config() for static fields or ConfigResolver.resolve_full_config() for bank-specific config.
+    """
    global _config_cache
    if _config_cache is None:
        _config_cache = HindsightConfig.from_env()
@@ -0,0 +1,274 @@
+"""
+Configuration resolution with hierarchical overrides.
+
+Resolves config values through the hierarchy:
+  Global (env vars) → Tenant config (via extension) → Bank config (database)
+
+Config values are resolved on every request to ensure consistency across
+multiple API servers.
+"""
+
+import json
+import logging
+from dataclasses import asdict
+from typing import Any
+
+import asyncpg
+
+from hindsight_api.config import HindsightConfig, _get_raw_config, normalize_config_dict
+from hindsight_api.extensions.tenant import TenantExtension
+from hindsight_api.models import RequestContext
+
+logger = logging.getLogger(__name__)
+
+
+class ConfigResolver:
+    """Resolves hierarchical configuration with tenant/bank overrides."""
+
+    def __init__(self, pool: asyncpg.Pool, tenant_extension: TenantExtension | None = None):
+        """
+        Initialize config resolver.
+
+        Args:
+            pool: Database connection pool
+            tenant_extension: Optional tenant extension for tenant-level config and permissions
+        """
+        self.pool = pool
+        self.tenant_extension = tenant_extension
+        self._global_config = _get_raw_config()
+        self._configurable_fields = HindsightConfig.get_configurable_fields()
+        self._credential_fields = HindsightConfig.get_credential_fields()
+
+    async def resolve_full_config(self, bank_id: str, context: RequestContext | None = None) -> HindsightConfig:
+        """
+        Resolve full HindsightConfig for a bank with hierarchical overrides applied.
+
+        This is for INTERNAL USE ONLY. Returns the complete config object with all fields
+        including credentials and static fields. Use get_bank_config() for API responses.
+
+        Resolution order:
+        1. Global config (from environment variables)
+        2. Tenant config overrides (from TenantExtension.get_tenant_config())
+        3. Bank config overrides (from banks.config JSONB)
+
+        Args:
+            bank_id: Bank identifier
+            context: Request context for tenant config resolution
+
+        Returns:
+            Complete HindsightConfig with hierarchical overrides applied
+        """
+        # Start with global config (all fields)
+        config_dict = asdict(self._global_config)
+
+        # Load tenant config overrides (if tenant extension available)
+        if self.tenant_extension and context:
+            try:
+                tenant_overrides = await self.tenant_extension.get_tenant_config(context)
+                if tenant_overrides:
+                    # Normalize keys and filter to configurable fields only
+                    normalized_tenant = normalize_config_dict(tenant_overrides)
+                    configurable_tenant = {k: v for k, v in normalized_tenant.items() if k in self._configurable_fields}
+                    config_dict.update(configurable_tenant)
+                    logger.debug(
+                        f"Applied tenant config overrides for bank {bank_id}: {list(configurable_tenant.keys())}"
+                    )
+            except Exception as e:
+                logger.warning(f"Failed to load tenant config for bank {bank_id}: {e}")
+
+        # Load bank config overrides
+        bank_overrides = await self._load_bank_config(bank_id)
+        if bank_overrides:
+            config_dict.update(bank_overrides)
+            logger.debug(f"Applied bank config overrides for bank {bank_id}: {list(bank_overrides.keys())}")
+
+        # Return full config object (dataclass doesn't have __init__ that accepts kwargs, so we update the object)
+        # Create a new config instance by copying the global config and updating fields
+        resolved_config = HindsightConfig(**config_dict)
+        return resolved_config
+
+    async def get_bank_config(self, bank_id: str, context: RequestContext | None = None) -> dict[str, Any]:
+        """
+        Get fully resolved config for a bank (filtered by permissions).
+
+        Resolution order:
+        1. Global config (from environment variables)
+        2. Tenant config overrides (from TenantExtension.get_tenant_config())
+        3. Bank config overrides (from banks.config JSONB)
+
+        Note: Config is resolved on every call (not cached) to ensure consistency
+        across multiple API servers.
+
+        SECURITY:
+        - Only returns configurable fields (excludes static/infrastructure fields)
+        - Filters out ALL credential fields (API keys, base URLs, etc.)
+        - Further filtered by tenant/bank permissions if extension provides them
+
+        Args:
+            bank_id: Bank identifier
+            context: Request context for tenant config resolution and permissions
+
+        Returns:
+            Dict of allowed configurable fields only (never includes credentials or static fields)
+        """
+        # Resolve full config with all hierarchical overrides
+        resolved_config = await self.resolve_full_config(bank_id, context)
+        config_dict = asdict(resolved_config)
+
+        # SECURITY: Filter to only configurable fields (exclude static/infrastructure)
+        filtered = {k: v for k, v in config_dict.items() if k in self._configurable_fields}
+
+        # SECURITY: Remove ALL credential fields (API keys, base URLs, etc.)
+        filtered = {k: v for k, v in filtered.items() if k not in self._credential_fields}
+
+        # PERMISSIONS: Further filter based on tenant/bank permissions
+        if self.tenant_extension and context:
+            try:
+                allowed_fields = await self.tenant_extension.get_allowed_config_fields(context, bank_id)
+                if allowed_fields is not None:  # None means "allow all"
+                    filtered = {k: v for k, v in filtered.items() if k in allowed_fields}
+                    logger.debug(
+                        f"Applied permission filter for bank {bank_id}: allowed={len(allowed_fields)} fields, "
+                        f"returned={len(filtered)} fields"
+                    )
+            except Exception as e:
+                logger.warning(f"Failed to load permissions for bank {bank_id}: {e}")
+
+        return filtered
+
+    async def _load_bank_config(self, bank_id: str) -> dict[str, Any]:
+        """
+        Load bank config overrides from banks.config JSONB column.
+
+        Args:
+            bank_id: Bank identifier
+
+        Returns:
+            Dict of config overrides (only configurable fields, normalized keys)
+        """
+        try:
+            async with self.pool.acquire() as conn:
+                row = await conn.fetchrow(
+                    """
+                    SELECT config FROM banks WHERE bank_id = $1
+                    """,
+                    bank_id,
+                )
+
+                if row and row["config"]:
+                    config_data = row["config"]
+
+                    # Handle case where JSONB is returned as JSON string
+                    if isinstance(config_data, str):
+                        config_data = json.loads(config_data)
+
+                    # Normalize keys (handle both env var format and Python field format)
+                    normalized = normalize_config_dict(config_data)
+
+                    # Only return overrides for configurable fields
+                    return {k: v for k, v in normalized.items() if k in self._configurable_fields}
+        except Exception as e:
+            logger.error(f"Failed to load bank config for {bank_id}: {e}")
+
+        return {}
+
+    async def update_bank_config(
+        self, bank_id: str, updates: dict[str, Any], context: RequestContext | None = None
+    ) -> None:
+        """
+        Update bank configuration overrides (with permission checking).
+
+        Args:
+            bank_id: Bank identifier
+            updates: Dict of config field names to new values.
+                    Keys can be in env var format (HINDSIGHT_API_LLM_PROVIDER)
+                    or Python field format (llm_provider).
+                    Only configurable fields are allowed.
+            context: Request context for permission checking
+
+        Raises:
+            ValueError: If attempting to override invalid/disallowed fields
+        """
+        # Normalize keys
+        normalized_updates = normalize_config_dict(updates)
+
+        # SECURITY: Reject credential fields explicitly
+        credential_attempts = set(normalized_updates.keys()) & self._credential_fields
+        if credential_attempts:
+            raise ValueError(
+                f"Cannot set credential fields via API: {sorted(credential_attempts)}. "
+                f"Credentials (API keys, base URLs) must be set at server level only."
+            )
+
+        # Validate all fields are configurable
+        invalid_fields = set(normalized_updates.keys()) - self._configurable_fields
+        if invalid_fields:
+            static_fields = HindsightConfig.get_static_fields()
+            invalid_static = invalid_fields & static_fields
+            if invalid_static:
+                raise ValueError(
+                    f"Cannot override static (server-level) fields: {sorted(invalid_static)}. "
+                    f"Only configurable fields can be overridden per-bank. "
+                    f"Configurable fields include: {sorted(list(self._configurable_fields)[:10])}... "
+                    f"(total: {len(self._configurable_fields)} fields)"
+                )
+            else:
+                raise ValueError(
+                    f"Unknown configuration fields: {sorted(invalid_fields)}. "
+                    f"Valid configurable fields: {sorted(list(self._configurable_fields)[:10])}..."
+                )
+
+        # PERMISSIONS: Check tenant/bank permissions
+        if self.tenant_extension and context:
+            try:
+                allowed_fields = await self.tenant_extension.get_allowed_config_fields(context, bank_id)
+                if allowed_fields is not None:  # None means "allow all"
+                    disallowed = set(normalized_updates.keys()) - allowed_fields
+                    if disallowed:
+                        raise ValueError(
+                            f"Not allowed to modify fields: {sorted(disallowed)}. "
+                            f"Your permissions allow: {sorted(list(allowed_fields)[:10])}..."
+                            if allowed_fields
+                            else "Not allowed to modify fields: {sorted(disallowed)}. "
+                            "Your permissions do not allow any config modifications."
+                        )
+            except ValueError:
+                raise  # Re-raise permission errors
+            except Exception as e:
+                logger.warning(f"Failed to check permissions for bank {bank_id}: {e}")
+                # Continue without permission check (fail open for backward compatibility)
+
+        # Merge with existing config (JSONB || operator)
+        async with self.pool.acquire() as conn:
+            await conn.execute(
+                """
+                UPDATE banks
+                SET config = config || $1::jsonb,
+                    updated_at = now()
+                WHERE bank_id = $2
+                """,
+                json.dumps(normalized_updates),
+                bank_id,
+            )
+
+        logger.info(f"Updated bank config for {bank_id}: {list(normalized_updates.keys())}")
+
+    async def reset_bank_config(self, bank_id: str) -> None:
+        """
+        Reset bank configuration to defaults (remove all overrides).
+
+        Args:
+            bank_id: Bank identifier
+        """
+        async with self.pool.acquire() as conn:
+            await conn.execute(
+                """
+                UPDATE banks
+                SET config = '{}'::jsonb,
+                    updated_at = now()
+                WHERE bank_id = $1
+                """,
+                bank_id,
+            )
+
+        logger.info(f"Reset bank config for {bank_id} to defaults")
@@ -1,11 +1,10 @@
 """
 Daemon mode support for Hindsight API.

-Provides idle timeout and lockfile management for running as a background daemon.
+Provides idle timeout for running as a background daemon.
 """

 import asyncio
-import fcntl
 import logging
 import os
 import sys
@@ -15,10 +14,11 @@ from pathlib import Path
 logger = logging.getLogger(__name__)

 # Default daemon configuration
-DEFAULT_DAEMON_PORT = 8889
+DEFAULT_DAEMON_PORT = 8888
 DEFAULT_IDLE_TIMEOUT = 0  # 0 = no auto-exit (hindsight-embed passes its own timeout)
-LOCKFILE_PATH = Path.home() / ".hindsight" / "daemon.lock"
-DAEMON_LOG_PATH = Path.home() / ".hindsight" / "daemon.log"
+
+# Allow override via environment variable for profile-specific logs
+DAEMON_LOG_PATH = Path(os.getenv("HINDSIGHT_API_DAEMON_LOG", str(Path.home() / ".hindsight" / "daemon.log")))


 class IdleTimeoutMiddleware:
@@ -58,97 +58,27 @@ class IdleTimeoutMiddleware:
                os.kill(os.getpid(), signal.SIGTERM)


-class DaemonLock:
-    """
-    File-based lock to prevent multiple daemon instances.
-
-    Uses fcntl.flock for atomic locking on Unix systems.
-    """
-
-    def __init__(self, lockfile: Path = LOCKFILE_PATH):
-        self.lockfile = lockfile
-        self._fd = None
-
-    def acquire(self) -> bool:
-        """
-        Try to acquire the daemon lock.
-
-        Returns True if lock acquired, False if another daemon is running.
-        """
-        self.lockfile.parent.mkdir(parents=True, exist_ok=True)
-
-        try:
-            self._fd = open(self.lockfile, "w")
-            fcntl.flock(self._fd.fileno(), fcntl.LOCK_EX | fcntl.LOCK_NB)
-            # Write PID for debugging
-            self._fd.write(str(os.getpid()))
-            self._fd.flush()
-            return True
-        except (IOError, OSError):
-            # Lock is held by another process
-            if self._fd:
-                self._fd.close()
-                self._fd = None
-            return False
-
-    def release(self):
-        """Release the daemon lock."""
-        if self._fd:
-            try:
-                fcntl.flock(self._fd.fileno(), fcntl.LOCK_UN)
-                self._fd.close()
-            except Exception:
-                pass
-            finally:
-                self._fd = None
-                # Remove lockfile
-                try:
-                    self.lockfile.unlink()
-                except Exception:
-                    pass
-
-    def is_locked(self) -> bool:
-        """Check if the lock is held by another process."""
-        if not self.lockfile.exists():
-            return False
-
-        try:
-            fd = open(self.lockfile, "r")
-            fcntl.flock(fd.fileno(), fcntl.LOCK_EX | fcntl.LOCK_NB)
-            # We got the lock, so no one else has it
-            fcntl.flock(fd.fileno(), fcntl.LOCK_UN)
-            fd.close()
-            return False
-        except (IOError, OSError):
-            return True
-
-    def get_pid(self) -> int | None:
-        """Get the PID of the daemon holding the lock."""
-        if not self.lockfile.exists():
-            return None
-        try:
-            with open(self.lockfile, "r") as f:
-                return int(f.read().strip())
-        except (ValueError, IOError):
-            return None
-
-
 def daemonize():
    """
    Fork the current process into a background daemon.

    Uses double-fork technique to properly detach from terminal.
    """
-    # First fork
-    pid = os.fork()
-    if pid > 0:
-        # Parent exits
-        sys.exit(0)
+    # First fork - detach from parent
+    try:
+        pid = os.fork()
+        if pid > 0:
+            sys.exit(0)
+    except OSError as e:
+        sys.stderr.write(f"fork #1 failed: {e}\n")
+        sys.exit(1)

-    # Create new session
+    # Decouple from parent environment
+    os.chdir("/")
    os.setsid()
+    os.umask(0)

-    # Second fork to prevent zombie processes
+    # Second fork - prevent zombie
    pid = os.fork()
    if pid > 0:
        sys.exit(0)
@@ -181,27 +111,3 @@ def check_daemon_running(port: int = DEFAULT_DAEMON_PORT) -> bool:
        return result == 0
    except Exception:
        return False
-
-
-def stop_daemon(port: int = DEFAULT_DAEMON_PORT) -> bool:
-    """Stop a running daemon by sending SIGTERM to the process."""
-    lock = DaemonLock()
-    pid = lock.get_pid()
-
-    if pid is None:
-        return False
-
-    try:
-        import signal
-
-        os.kill(pid, signal.SIGTERM)
-        # Wait for process to exit
-        for _ in range(50):  # Wait up to 5 seconds
-            time.sleep(0.1)
-            try:
-                os.kill(pid, 0)  # Check if process exists
-            except OSError:
-                return True  # Process exited
-        return False
-    except OSError:
-        return False
@@ -18,6 +18,7 @@ import uuid
 from datetime import datetime, timezone
 from typing import TYPE_CHECKING, Any

+from ...config import get_config
 from ..memory_engine import fq_table
 from ..retain import embedding_utils
 from .prompts import (
@@ -82,9 +83,8 @@ async def run_consolidation_job(
    Returns:
        Dict with consolidation results
    """
-    from ...config import get_config
-
-    config = get_config()
+    # Resolve bank-specific config with hierarchical overrides
+    config = await memory_engine._config_resolver.resolve_full_config(bank_id, request_context)
    perf = ConsolidationPerfLog(bank_id)
    max_memories_per_batch = config.consolidation_batch_size

@@ -143,6 +143,9 @@ async def run_consolidation_job(
        "skipped": 0,
    }

+    # Track all unique tags from consolidated memories for mental model refresh filtering
+    consolidated_tags: set[str] = set()
+
    batch_num = 0
    last_progress_timings = {}  # Track timings at last progress log
    while True:
@@ -176,6 +179,11 @@ async def run_consolidation_job(
        for memory in memories:
            mem_start = time.time()

+            # Track tags from this memory for mental model refresh filtering
+            memory_tags = memory.get("tags") or []
+            if memory_tags:
+                consolidated_tags.update(memory_tags)
+
            # Process the memory (uses its own connection internally)
            async with pool.acquire() as conn:
                result = await _process_memory(
@@ -284,10 +292,12 @@ async def run_consolidation_job(
        perf.log(f"[4] Timing breakdown: {', '.join(timing_parts)}")

    # Trigger mental model refreshes for models with refresh_after_consolidation=true
+    # SECURITY: Only refresh mental models with matching tags (or all if no tags were consolidated)
    mental_models_refreshed = await _trigger_mental_model_refreshes(
        memory_engine=memory_engine,
        bank_id=bank_id,
        request_context=request_context,
+        consolidated_tags=list(consolidated_tags) if consolidated_tags else None,
        perf=perf,
    )
    stats["mental_models_refreshed"] = mental_models_refreshed
@@ -301,15 +311,20 @@ async def _trigger_mental_model_refreshes(
    memory_engine: "MemoryEngine",
    bank_id: str,
    request_context: "RequestContext",
+    consolidated_tags: list[str] | None = None,
    perf: ConsolidationPerfLog | None = None,
 ) -> int:
    """
    Trigger refreshes for mental models with refresh_after_consolidation=true.

+    SECURITY: Only triggers refresh for mental models whose tags overlap with the
+    consolidated memory tags, preventing unnecessary refreshes across security boundaries.
+
    Args:
        memory_engine: MemoryEngine instance
        bank_id: Bank identifier
        request_context: Request context for authentication
+        consolidated_tags: Tags from memories that were consolidated (None = refresh all)
        perf: Performance logging

    Returns:
@@ -318,22 +333,52 @@ async def _trigger_mental_model_refreshes(
    pool = memory_engine._pool

    # Find mental models with refresh_after_consolidation=true
+    # SECURITY: Control which mental models get refreshed based on tags
    async with pool.acquire() as conn:
-        rows = await conn.fetch(
-            f"""
-            SELECT id, name
-            FROM {fq_table("mental_models")}
-            WHERE bank_id = $1
-              AND (trigger->>'refresh_after_consolidation')::boolean = true
-            """,
-            bank_id,
-        )
+        if consolidated_tags:
+            # Tagged memories were consolidated - refresh:
+            # 1. Mental models with overlapping tags (security boundary)
+            # 2. Untagged mental models (they're "global" and available to all contexts)
+            # DO NOT refresh mental models with different tags
+            rows = await conn.fetch(
+                f"""
+                SELECT id, name, tags
+                FROM {fq_table("mental_models")}
+                WHERE bank_id = $1
+                  AND (trigger->>'refresh_after_consolidation')::boolean = true
+                  AND (
+                    (tags IS NOT NULL AND tags != '{{}}' AND tags && $2::varchar[])
+                    OR (tags IS NULL OR tags = '{{}}')
+                  )
+                """,
+                bank_id,
+                consolidated_tags,
+            )
+        else:
+            # Untagged memories were consolidated - only refresh untagged mental models
+            # SECURITY: Tagged mental models are NOT refreshed when untagged memories are consolidated
+            rows = await conn.fetch(
+                f"""
+                SELECT id, name, tags
+                FROM {fq_table("mental_models")}
+                WHERE bank_id = $1
+                  AND (trigger->>'refresh_after_consolidation')::boolean = true
+                  AND (tags IS NULL OR tags = '{{}}')
+                """,
+                bank_id,
+            )

    if not rows:
        return 0

    if perf:
-        perf.log(f"[5] Triggering refresh for {len(rows)} mental models with refresh_after_consolidation=true")
+        if consolidated_tags:
+            perf.log(
+                f"[5] Triggering refresh for {len(rows)} mental models with refresh_after_consolidation=true "
+                f"(filtered by tags: {consolidated_tags})"
+            )
+        else:
+            perf.log(f"[5] Triggering refresh for {len(rows)} mental models with refresh_after_consolidation=true")

    # Submit refresh tasks for each mental model
    refreshed_count = 0
@@ -381,92 +426,109 @@ async def _process_memory(
    Returns:
        Dict with action summary: created/updated/merged counts
    """
+    from ...tracing import get_tracer, is_tracing_enabled
+
    fact_text = memory["text"]
    memory_id = memory["id"]
    fact_tags = memory.get("tags") or []

-    # Find related observations using the full recall system (NO tag filtering)
-    t0 = time.time()
-    related_observations = await _find_related_observations(
-        conn=conn,
-        memory_engine=memory_engine,
-        bank_id=bank_id,
-        query=fact_text,
-        request_context=request_context,
-    )
-    if perf:
-        perf.record_timing("recall", time.time() - t0)
+    # Create parent span for this memory's consolidation
+    tracer = get_tracer()
+    if is_tracing_enabled():
+        consolidation_span = tracer.start_span("hindsight.consolidation")
+        consolidation_span.set_attribute("hindsight.memory_id", str(memory_id))
+        consolidation_span.set_attribute("hindsight.bank_id", bank_id)
+    else:
+        consolidation_span = None

-    # Single LLM call handles ALL cases (with or without existing observations)
-    # Note: Tags are NOT passed to LLM - they are handled algorithmically
-    t0 = time.time()
-    actions = await _consolidate_with_llm(
-        memory_engine=memory_engine,
-        fact_text=fact_text,
-        observations=related_observations,  # Can be empty list
-        mission=mission,
-    )
-    if perf:
-        perf.record_timing("llm", time.time() - t0)
+    try:
+        # Find related observations using the full recall system
+        # SECURITY: Pass tags to ensure observations don't leak across security boundaries
+        t0 = time.time()
+        related_observations = await _find_related_observations(
+            conn=conn,
+            memory_engine=memory_engine,
+            bank_id=bank_id,
+            query=fact_text,
+            request_context=request_context,
+            tags=fact_tags,  # Pass source memory's tags for security
+        )
+        if perf:
+            perf.record_timing("recall", time.time() - t0)

-    if not actions:
-        # LLM returned empty array - fact is purely ephemeral, skip
-        return {"action": "skipped", "reason": "no_durable_knowledge"}
+        # Single LLM call handles ALL cases (with or without existing observations)
+        # Note: Tags are NOT passed to LLM - they are handled algorithmically
+        t0 = time.time()
+        actions = await _consolidate_with_llm(
+            memory_engine=memory_engine,
+            fact_text=fact_text,
+            observations=related_observations,  # Can be empty list
+            mission=mission,
+        )
+        if perf:
+            perf.record_timing("llm", time.time() - t0)

-    # Execute all actions and collect results
-    results = []
-    for action in actions:
-        action_type = action.get("action")
-        if action_type == "update":
-            result = await _execute_update_action(
-                conn=conn,
-                memory_engine=memory_engine,
-                bank_id=bank_id,
-                memory_id=memory_id,
-                action=action,
-                observations=related_observations,
-                source_fact_tags=fact_tags,  # Pass source fact's tags for security
-                source_occurred_start=memory.get("occurred_start"),
-                source_occurred_end=memory.get("occurred_end"),
-                source_mentioned_at=memory.get("mentioned_at"),
-                perf=perf,
-            )
-            results.append(result)
-        elif action_type == "create":
-            result = await _execute_create_action(
-                conn=conn,
-                memory_engine=memory_engine,
-                bank_id=bank_id,
-                memory_id=memory_id,
-                action=action,
-                source_fact_tags=fact_tags,  # Pass source fact's tags for security
-                event_date=memory.get("event_date"),
-                occurred_start=memory.get("occurred_start"),
-                occurred_end=memory.get("occurred_end"),
-                mentioned_at=memory.get("mentioned_at"),
-                perf=perf,
-            )
-            results.append(result)
+        if not actions:
+            # LLM returned empty array - fact is purely ephemeral, skip
+            return {"action": "skipped", "reason": "no_durable_knowledge"}

-    if not results:
-        # No valid actions executed
-        return {"action": "skipped", "reason": "no_valid_actions"}
+        # Execute all actions and collect results
+        results = []
+        for action in actions:
+            action_type = action.get("action")
+            if action_type == "update":
+                result = await _execute_update_action(
+                    conn=conn,
+                    memory_engine=memory_engine,
+                    bank_id=bank_id,
+                    memory_id=memory_id,
+                    action=action,
+                    observations=related_observations,
+                    source_fact_tags=fact_tags,  # Pass source fact's tags for security
+                    source_occurred_start=memory.get("occurred_start"),
+                    source_occurred_end=memory.get("occurred_end"),
+                    source_mentioned_at=memory.get("mentioned_at"),
+                    perf=perf,
+                )
+                results.append(result)
+            elif action_type == "create":
+                result = await _execute_create_action(
+                    conn=conn,
+                    memory_engine=memory_engine,
+                    bank_id=bank_id,
+                    memory_id=memory_id,
+                    action=action,
+                    source_fact_tags=fact_tags,  # Pass source fact's tags for security
+                    event_date=memory.get("event_date"),
+                    occurred_start=memory.get("occurred_start"),
+                    occurred_end=memory.get("occurred_end"),
+                    mentioned_at=memory.get("mentioned_at"),
+                    perf=perf,
+                )
+                results.append(result)

-    # Summarize results
-    created = sum(1 for r in results if r.get("action") == "created")
-    updated = sum(1 for r in results if r.get("action") == "updated")
-    merged = sum(1 for r in results if r.get("action") == "merged")
+        if not results:
+            # No valid actions executed
+            return {"action": "skipped", "reason": "no_valid_actions"}

-    if len(results) == 1:
-        return results[0]
+        # Summarize results
+        created = sum(1 for r in results if r.get("action") == "created")
+        updated = sum(1 for r in results if r.get("action") == "updated")
+        merged = sum(1 for r in results if r.get("action") == "merged")

-    return {
-        "action": "multiple",
-        "created": created,
-        "updated": updated,
-        "merged": merged,
-        "total_actions": len(results),
-    }
+        if len(results) == 1:
+            return results[0]
+
+        return {
+            "action": "multiple",
+            "created": created,
+            "updated": updated,
+            "merged": merged,
+            "total_actions": len(results),
+        }
+    finally:
+        if consolidation_span:
+            consolidation_span.end()


 async def _execute_update_action(
@@ -666,34 +728,57 @@ async def _find_related_observations(
    bank_id: str,
    query: str,
    request_context: "RequestContext",
+    tags: list[str] | None = None,
 ) -> list[dict[str, Any]]:
    """
    Find observations related to the given query using optimized recall.

-    IMPORTANT: We do NOT filter by tags here. Consolidation needs to see ALL
-    potentially related observations regardless of scope, so the LLM can
-    decide on tag routing (same scope update vs cross-scope create).
+    SECURITY: Filters by tags using all_strict matching to prevent cross-tenant/cross-user
+    information leakage. Observations are only consolidated within the same tag scope.

    Uses max_tokens to naturally limit observations (no artificial count limit).
    Includes source memories with dates for LLM context.

+    Args:
+        tags: Optional tags to filter observations (uses all_strict matching for security)
+
    Returns:
        List of related observations with their tags, source memories, and dates
    """
    # Use recall to find related observations with token budget
    # max_tokens naturally limits how many observations are returned
    from ...config import get_config
+    from ...tracing import get_tracer, is_tracing_enabled

    config = get_config()
-    recall_result = await memory_engine.recall_async(
-        bank_id=bank_id,
-        query=query,
-        max_tokens=config.consolidation_max_tokens,  # Token budget for observations (configurable)
-        fact_type=["observation"],  # Only retrieve observations
-        request_context=request_context,
-        _quiet=True,  # Suppress logging
-        # NO tags parameter - intentionally get ALL observations
-    )
+
+    # SECURITY: Use all_strict matching if tags provided to prevent cross-scope consolidation
+    tags_match = "all_strict" if tags else "any"
+
+    # Create span for recall operation within consolidation
+    tracer = get_tracer()
+    if is_tracing_enabled():
+        recall_span = tracer.start_span("hindsight.consolidation_recall")
+        recall_span.set_attribute("hindsight.bank_id", bank_id)
+        recall_span.set_attribute("hindsight.query", query[:100])  # Truncate for brevity
+        recall_span.set_attribute("hindsight.fact_type", "observation")
+    else:
+        recall_span = None
+
+    try:
+        recall_result = await memory_engine.recall_async(
+            bank_id=bank_id,
+            query=query,
+            max_tokens=config.consolidation_max_tokens,  # Token budget for observations (configurable)
+            fact_type=["observation"],  # Only retrieve observations
+            request_context=request_context,
+            tags=tags,  # Filter by source memory's tags
+            tags_match=tags_match,  # Use strict matching for security
+            _quiet=True,  # Suppress logging
+        )
+    finally:
+        if recall_span:
+            recall_span.end()

    # If no observations returned, return empty list
    if not recall_result.results:
@@ -931,15 +1016,33 @@ async def _create_observation_directly(

    t0 = time.time()
    observation_id = uuid.uuid4()
+
+    # Query varies based on text search backend
+    config = get_config()
+    if config.text_search_extension == "vchord":
+        # VectorChord: manually tokenize and insert search_vector
+        query = f"""
+            INSERT INTO {fq_table("memory_units")} (
+                id, bank_id, text, fact_type, embedding, proof_count, source_memory_ids, history,
+                tags, event_date, occurred_start, occurred_end, mentioned_at, search_vector
+            )
+            VALUES ($1, $2, $3, 'observation', $4::vector, 1, $5, '[]'::jsonb, $6, $7, $8, $9, $10,
+                    tokenize($3, 'llmlingua2')::bm25_catalog.bm25vector)
+            RETURNING id
+        """
+    else:  # native
+        # Native PostgreSQL: search_vector is GENERATED ALWAYS, don't include it
+        query = f"""
+            INSERT INTO {fq_table("memory_units")} (
+                id, bank_id, text, fact_type, embedding, proof_count, source_memory_ids, history,
+                tags, event_date, occurred_start, occurred_end, mentioned_at
+            )
+            VALUES ($1, $2, $3, 'observation', $4::vector, 1, $5, '[]'::jsonb, $6, $7, $8, $9, $10)
+            RETURNING id
+        """
+
    row = await conn.fetchrow(
-        f"""
-        INSERT INTO {fq_table("memory_units")} (
-            id, bank_id, text, fact_type, embedding, proof_count, source_memory_ids, history,
-            tags, event_date, occurred_start, occurred_end, mentioned_at
-        )
-        VALUES ($1, $2, $3, 'observation', $4::vector, 1, $5, '[]'::jsonb, $6, $7, $8, $9, $10)
-        RETURNING id
-        """,
+        query,
        observation_id,
        bank_id,
        observation_text,
@@ -2,7 +2,7 @@

 CONSOLIDATION_SYSTEM_PROMPT = """You are a memory consolidation system. Your job is to convert facts into durable knowledge (observations) and merge with existing knowledge when appropriate.

-You must output ONLY valid JSON with no markdown formatting, no code blocks, and no additional text.
+You must output ONLY valid JSON with no markdown code blocks or additional text. However, the "text" field within each observation should use markdown formatting (headers, lists, bold, etc.) for clarity and readability.

 ## EXTRACT DURABLE KNOWLEDGE, NOT EPHEMERAL STATE
 Facts often describe events or actions. Extract the DURABLE KNOWLEDGE implied by the fact, not the transient state.
@@ -32,13 +32,16 @@ BAD examples:

 ## MERGE RULES (when comparing to existing observations):
 1. REDUNDANT: Same information worded differently → update existing
-2. CONTRADICTION: Opposite information about same topic → update with history (e.g., "used to X, now Y")
-3. UPDATE: New state replacing old state → update with history
+2. CONTRADICTION: Opposite information about same topic → update with temporal markers showing change
+   Example: "Alex used to love pizza but now hates it" OR "Alex's pizza preference changed from love to hate"
+3. UPDATE: New state replacing old state → update showing the transition with "used to", "now", "changed from X to Y"

 ## CRITICAL RULES:
 - NEVER merge facts about DIFFERENT people
 - NEVER merge unrelated topics (food preferences vs work vs hobbies)
- When merging contradictions, capture the CHANGE (before → after)
+- When merging contradictions, the "text" field MUST capture BOTH states with temporal markers:
+  * Use "used to X, now Y" OR "changed from X to Y" OR "X but now Y"
+  * DO NOT just state the new fact - you MUST show the change
 - Keep observations focused on ONE specific topic per person
 - The "text" field MUST contain durable knowledge, not ephemeral state
 - Do NOT include "tags" in output - tags are handled automatically"""
@@ -68,10 +71,15 @@ Instructions:
   - New topic → CREATE new observation
   - Purely ephemeral → return []

-Output JSON array of actions:
+Output JSON array of actions (the "text" field should use markdown formatting for structure):
 [
-  {{"action": "update", "learning_id": "uuid-from-observations", "text": "updated knowledge", "reason": "..."}},
-  {{"action": "create", "text": "new durable knowledge", "reason": "..."}}
+  {{"action": "update", "learning_id": "uuid-from-observations", "text": "## Updated Knowledge\n\n**Key point**: details here\n\n- Supporting detail 1\n- Supporting detail 2", "reason": "..."}},
+  {{"action": "create", "text": "## New Durable Knowledge\n\nDescription with **emphasis** and proper structure", "reason": "..."}}
 ]

-Return [] if fact contains no durable knowledge."""
+Return [] if fact contains no durable knowledge.
+
+IMPORTANT: Format the "text" field with markdown for better readability:
+- Use headers, lists, bold/italic, tables where appropriate
+- CRITICAL: Add blank lines before and after block elements (tables, code blocks, lists)
+- Ensure proper spacing for markdown to render correctly"""
@@ -9,6 +9,7 @@ Configuration via environment variables - see hindsight_api.config for all env v
 import asyncio
 import logging
 import os
+import warnings
 from abc import ABC, abstractmethod
 from concurrent.futures import ThreadPoolExecutor

@@ -23,20 +24,18 @@ from ..config import (
    DEFAULT_RERANKER_LOCAL_FORCE_CPU,
    DEFAULT_RERANKER_LOCAL_MAX_CONCURRENT,
    DEFAULT_RERANKER_LOCAL_MODEL,
+    DEFAULT_RERANKER_LOCAL_TRUST_REMOTE_CODE,
    DEFAULT_RERANKER_PROVIDER,
    DEFAULT_RERANKER_TEI_BATCH_SIZE,
    DEFAULT_RERANKER_TEI_MAX_CONCURRENT,
-    ENV_COHERE_API_KEY,
-    ENV_LITELLM_API_BASE,
-    ENV_LITELLM_API_KEY,
-    ENV_RERANKER_COHERE_BASE_URL,
+    ENV_RERANKER_COHERE_API_KEY,
    ENV_RERANKER_COHERE_MODEL,
    ENV_RERANKER_FLASHRANK_CACHE_DIR,
    ENV_RERANKER_FLASHRANK_MODEL,
-    ENV_RERANKER_LITELLM_MODEL,
    ENV_RERANKER_LOCAL_FORCE_CPU,
    ENV_RERANKER_LOCAL_MAX_CONCURRENT,
    ENV_RERANKER_LOCAL_MODEL,
+    ENV_RERANKER_LOCAL_TRUST_REMOTE_CODE,
    ENV_RERANKER_PROVIDER,
    ENV_RERANKER_TEI_BATCH_SIZE,
    ENV_RERANKER_TEI_MAX_CONCURRENT,
@@ -101,7 +100,13 @@ class LocalSTCrossEncoder(CrossEncoderModel):
    _executor: ThreadPoolExecutor | None = None
    _max_concurrent: int = 4  # Limit concurrent CPU-bound reranking calls

-    def __init__(self, model_name: str | None = None, max_concurrent: int = 4, force_cpu: bool = False):
+    def __init__(
+        self,
+        model_name: str | None = None,
+        max_concurrent: int = 4,
+        force_cpu: bool = False,
+        trust_remote_code: bool = False,
+    ):
        """
        Initialize local SentenceTransformers cross-encoder.

@@ -112,9 +117,13 @@ class LocalSTCrossEncoder(CrossEncoderModel):
                           Higher values may cause CPU thrashing under load.
            force_cpu: Force CPU mode (avoids MPS/XPC issues on macOS in daemon mode).
                      Default: False
+            trust_remote_code: Allow loading models with custom code (security risk).
+                              Required for some models like jina-reranker-v2-base-multilingual.
+                              Default: False (disabled for security)
        """
        self.model_name = model_name or DEFAULT_RERANKER_LOCAL_MODEL
        self.force_cpu = force_cpu
+        self.trust_remote_code = trust_remote_code
        self._model = None
        LocalSTCrossEncoder._max_concurrent = max_concurrent

@@ -162,11 +171,29 @@ class LocalSTCrossEncoder(CrossEncoderModel):
            except Exception as e:
                logger.warning(f"Failed to detect GPU/MPS, falling back to CPU: {e}")

-        self._model = CrossEncoder(
-            self.model_name,
-            device=device,
-            model_kwargs={"low_cpu_mem_usage": False},
-        )
+        # Suppress verbose transformers warnings during model loading
+        # This suppresses the "UNEXPECTED" warnings from CrossEncoder which are harmless
+        # but look alarming to users (e.g., "embeddings.position_ids | UNEXPECTED")
+        with warnings.catch_warnings():
+            warnings.filterwarnings("ignore", category=UserWarning)
+            warnings.filterwarnings("ignore", message=".*was not found in model state dict.*")
+            warnings.filterwarnings("ignore", message=".*UNEXPECTED.*")
+
+            # Also suppress transformers library logging temporarily
+            transformers_logger = logging.getLogger("transformers")
+            original_level = transformers_logger.level
+            transformers_logger.setLevel(logging.ERROR)
+
+            try:
+                self._model = CrossEncoder(
+                    self.model_name,
+                    device=device,
+                    model_kwargs={"low_cpu_mem_usage": False},
+                    trust_remote_code=self.trust_remote_code,
+                )
+            finally:
+                # Restore original logging level
+                transformers_logger.setLevel(original_level)

        # Initialize shared executor (limited workers naturally limits concurrency)
        if LocalSTCrossEncoder._executor is None:
@@ -829,23 +856,27 @@ def create_cross_encoder_from_env() -> CrossEncoderModel:
            model_name=config.reranker_local_model,
            max_concurrent=config.reranker_local_max_concurrent,
            force_cpu=config.reranker_local_force_cpu,
+            trust_remote_code=config.reranker_local_trust_remote_code,
        )
    elif provider == "cohere":
-        api_key = os.environ.get(ENV_COHERE_API_KEY)
+        api_key = config.reranker_cohere_api_key
        if not api_key:
-            raise ValueError(f"{ENV_COHERE_API_KEY} is required when {ENV_RERANKER_PROVIDER} is 'cohere'")
-        model = os.environ.get(ENV_RERANKER_COHERE_MODEL, DEFAULT_RERANKER_COHERE_MODEL)
-        base_url = os.environ.get(ENV_RERANKER_COHERE_BASE_URL) or None
-        return CohereCrossEncoder(api_key=api_key, model=model, base_url=base_url)
+            raise ValueError(f"{ENV_RERANKER_COHERE_API_KEY} is required when {ENV_RERANKER_PROVIDER} is 'cohere'")
+        return CohereCrossEncoder(
+            api_key=api_key,
+            model=config.reranker_cohere_model,
+            base_url=config.reranker_cohere_base_url,
+        )
    elif provider == "flashrank":
        model = os.environ.get(ENV_RERANKER_FLASHRANK_MODEL, DEFAULT_RERANKER_FLASHRANK_MODEL)
        cache_dir = os.environ.get(ENV_RERANKER_FLASHRANK_CACHE_DIR, DEFAULT_RERANKER_FLASHRANK_CACHE_DIR)
        return FlashRankCrossEncoder(model_name=model, cache_dir=cache_dir)
    elif provider == "litellm":
-        api_base = os.environ.get(ENV_LITELLM_API_BASE, DEFAULT_LITELLM_API_BASE)
-        api_key = os.environ.get(ENV_LITELLM_API_KEY)
-        model = os.environ.get(ENV_RERANKER_LITELLM_MODEL, DEFAULT_RERANKER_LITELLM_MODEL)
-        return LiteLLMCrossEncoder(api_base=api_base, api_key=api_key, model=model)
+        return LiteLLMCrossEncoder(
+            api_base=config.reranker_litellm_api_base,
+            api_key=config.reranker_litellm_api_key,
+            model=config.reranker_litellm_model,
+        )
    elif provider == "rrf":
        return RRFPassthroughCrossEncoder()
    else:
@@ -11,6 +11,7 @@ Configuration via environment variables - see hindsight_api.config for all env v

 import logging
 import os
+import warnings
 from abc import ABC, abstractmethod

 import httpx
@@ -20,22 +21,19 @@ from ..config import (
    DEFAULT_EMBEDDINGS_LITELLM_MODEL,
    DEFAULT_EMBEDDINGS_LOCAL_FORCE_CPU,
    DEFAULT_EMBEDDINGS_LOCAL_MODEL,
+    DEFAULT_EMBEDDINGS_LOCAL_TRUST_REMOTE_CODE,
    DEFAULT_EMBEDDINGS_OPENAI_MODEL,
    DEFAULT_EMBEDDINGS_PROVIDER,
    DEFAULT_LITELLM_API_BASE,
-    ENV_COHERE_API_KEY,
-    ENV_EMBEDDINGS_COHERE_BASE_URL,
-    ENV_EMBEDDINGS_COHERE_MODEL,
-    ENV_EMBEDDINGS_LITELLM_MODEL,
+    ENV_EMBEDDINGS_COHERE_API_KEY,
    ENV_EMBEDDINGS_LOCAL_FORCE_CPU,
    ENV_EMBEDDINGS_LOCAL_MODEL,
+    ENV_EMBEDDINGS_LOCAL_TRUST_REMOTE_CODE,
    ENV_EMBEDDINGS_OPENAI_API_KEY,
    ENV_EMBEDDINGS_OPENAI_BASE_URL,
    ENV_EMBEDDINGS_OPENAI_MODEL,
    ENV_EMBEDDINGS_PROVIDER,
    ENV_EMBEDDINGS_TEI_URL,
-    ENV_LITELLM_API_BASE,
-    ENV_LITELLM_API_KEY,
    ENV_LLM_API_KEY,
 )

@@ -94,7 +92,7 @@ class LocalSTEmbeddings(Embeddings):
    The embedding dimension is auto-detected from the model.
    """

-    def __init__(self, model_name: str | None = None, force_cpu: bool = False):
+    def __init__(self, model_name: str | None = None, force_cpu: bool = False, trust_remote_code: bool = False):
        """
        Initialize local SentenceTransformers embeddings.

@@ -103,9 +101,13 @@ class LocalSTEmbeddings(Embeddings):
                       Default: BAAI/bge-small-en-v1.5
            force_cpu: Force CPU mode (avoids MPS/XPC issues on macOS in daemon mode).
                      Default: False
+            trust_remote_code: Allow loading models with custom code (security risk).
+                              Required for some models with custom architectures.
+                              Default: False (disabled for security)
        """
        self.model_name = model_name or DEFAULT_EMBEDDINGS_LOCAL_MODEL
        self.force_cpu = force_cpu
+        self.trust_remote_code = trust_remote_code
        self._model = None
        self._dimension: int | None = None

@@ -157,11 +159,29 @@ class LocalSTEmbeddings(Embeddings):
            except Exception as e:
                logger.warning(f"Failed to detect GPU/MPS, falling back to CPU: {e}")

-        self._model = SentenceTransformer(
-            self.model_name,
-            device=device,
-            model_kwargs={"low_cpu_mem_usage": False},
-        )
+        # Suppress verbose transformers warnings during model loading
+        # This suppresses the "UNEXPECTED" warnings from BertModel which are harmless
+        # but look alarming to users (e.g., "embeddings.position_ids | UNEXPECTED")
+        with warnings.catch_warnings():
+            warnings.filterwarnings("ignore", category=UserWarning)
+            warnings.filterwarnings("ignore", message=".*was not found in model state dict.*")
+            warnings.filterwarnings("ignore", message=".*UNEXPECTED.*")
+
+            # Also suppress transformers library logging temporarily
+            transformers_logger = logging.getLogger("transformers")
+            original_level = transformers_logger.level
+            transformers_logger.setLevel(logging.ERROR)
+
+            try:
+                self._model = SentenceTransformer(
+                    self.model_name,
+                    device=device,
+                    model_kwargs={"low_cpu_mem_usage": False},
+                    trust_remote_code=self.trust_remote_code,
+                )
+            finally:
+                # Restore original logging level
+                transformers_logger.setLevel(original_level)

        self._dimension = self._model.get_sentence_embedding_dimension()
        logger.info(f"Embeddings: local provider initialized (dim: {self._dimension})")
@@ -723,6 +743,7 @@ def create_embeddings_from_env() -> Embeddings:
        return LocalSTEmbeddings(
            model_name=config.embeddings_local_model,
            force_cpu=config.embeddings_local_force_cpu,
+            trust_remote_code=config.embeddings_local_trust_remote_code,
        )
    elif provider == "openai":
        # Use dedicated embeddings API key, or fall back to LLM API key
@@ -736,17 +757,20 @@ def create_embeddings_from_env() -> Embeddings:
        base_url = os.environ.get(ENV_EMBEDDINGS_OPENAI_BASE_URL) or None
        return OpenAIEmbeddings(api_key=api_key, model=model, base_url=base_url)
    elif provider == "cohere":
-        api_key = os.environ.get(ENV_COHERE_API_KEY)
+        api_key = config.embeddings_cohere_api_key
        if not api_key:
-            raise ValueError(f"{ENV_COHERE_API_KEY} is required when {ENV_EMBEDDINGS_PROVIDER} is 'cohere'")
-        model = os.environ.get(ENV_EMBEDDINGS_COHERE_MODEL, DEFAULT_EMBEDDINGS_COHERE_MODEL)
-        base_url = os.environ.get(ENV_EMBEDDINGS_COHERE_BASE_URL) or None
-        return CohereEmbeddings(api_key=api_key, model=model, base_url=base_url)
+            raise ValueError(f"{ENV_EMBEDDINGS_COHERE_API_KEY} is required when {ENV_EMBEDDINGS_PROVIDER} is 'cohere'")
+        return CohereEmbeddings(
+            api_key=api_key,
+            model=config.embeddings_cohere_model,
+            base_url=config.embeddings_cohere_base_url,
+        )
    elif provider == "litellm":
-        api_base = os.environ.get(ENV_LITELLM_API_BASE, DEFAULT_LITELLM_API_BASE)
-        api_key = os.environ.get(ENV_LITELLM_API_KEY)
-        model = os.environ.get(ENV_EMBEDDINGS_LITELLM_MODEL, DEFAULT_EMBEDDINGS_LITELLM_MODEL)
-        return LiteLLMEmbeddings(api_base=api_base, api_key=api_key, model=model)
+        return LiteLLMEmbeddings(
+            api_base=config.embeddings_litellm_api_base,
+            api_key=config.embeddings_litellm_api_key,
+            model=config.embeddings_litellm_model,
+        )
    else:
        raise ValueError(
            f"Unknown embeddings provider: {provider}. Supported: 'local', 'tei', 'openai', 'cohere', 'litellm'"
@@ -0,0 +1,146 @@
+"""
+Abstract interface for LLM providers.
+
+This module defines the interface that all LLM providers must implement,
+enabling support for multiple LLM backends (OpenAI, Anthropic, Gemini, Codex, etc.)
+"""
+
+from abc import ABC, abstractmethod
+from typing import Any
+
+from .response_models import LLMToolCallResult, TokenUsage
+
+
+class LLMInterface(ABC):
+    """
+    Abstract interface for LLM providers.
+
+    All LLM provider implementations must inherit from this class and implement
+    the required methods.
+    """
+
+    def __init__(
+        self,
+        provider: str,
+        api_key: str,
+        base_url: str,
+        model: str,
+        reasoning_effort: str = "low",
+        **kwargs: Any,
+    ):
+        """
+        Initialize LLM provider.
+
+        Args:
+            provider: Provider name (e.g., "openai", "codex", "anthropic", "gemini").
+            api_key: API key or authentication token.
+            base_url: Base URL for the API.
+            model: Model name.
+            reasoning_effort: Reasoning effort level for supported providers.
+            **kwargs: Additional provider-specific parameters.
+        """
+        self.provider = provider.lower()
+        self.api_key = api_key
+        self.base_url = base_url
+        self.model = model
+        self.reasoning_effort = reasoning_effort
+
+    @abstractmethod
+    async def verify_connection(self) -> None:
+        """
+        Verify that the LLM provider is configured correctly by making a simple test call.
+
+        Raises:
+            RuntimeError: If the connection test fails.
+        """
+        pass
+
+    @abstractmethod
+    async def call(
+        self,
+        messages: list[dict[str, str]],
+        response_format: Any | None = None,
+        max_completion_tokens: int | None = None,
+        temperature: float | None = None,
+        scope: str = "memory",
+        max_retries: int = 10,
+        initial_backoff: float = 1.0,
+        max_backoff: float = 60.0,
+        skip_validation: bool = False,
+        strict_schema: bool = False,
+        return_usage: bool = False,
+    ) -> Any:
+        """
+        Make an LLM API call with retry logic.
+
+        Args:
+            messages: List of message dicts with 'role' and 'content'.
+            response_format: Optional Pydantic model for structured output.
+            max_completion_tokens: Maximum tokens in response.
+            temperature: Sampling temperature (0.0-2.0).
+            scope: Scope identifier for tracking.
+            max_retries: Maximum retry attempts.
+            initial_backoff: Initial backoff time in seconds.
+            max_backoff: Maximum backoff time in seconds.
+            skip_validation: Return raw JSON without Pydantic validation.
+            strict_schema: Use strict JSON schema enforcement (OpenAI only).
+            return_usage: If True, return tuple (result, TokenUsage) instead of just result.
+
+        Returns:
+            If return_usage=False: Parsed response if response_format is provided, otherwise text content.
+            If return_usage=True: Tuple of (result, TokenUsage) with token counts.
+
+        Raises:
+            OutputTooLongError: If output exceeds token limits.
+            Exception: Re-raises API errors after retries exhausted.
+        """
+        pass
+
+    @abstractmethod
+    async def call_with_tools(
+        self,
+        messages: list[dict[str, Any]],
+        tools: list[dict[str, Any]],
+        max_completion_tokens: int | None = None,
+        temperature: float | None = None,
+        scope: str = "tools",
+        max_retries: int = 5,
+        initial_backoff: float = 1.0,
+        max_backoff: float = 30.0,
+        tool_choice: str | dict[str, Any] = "auto",
+    ) -> LLMToolCallResult:
+        """
+        Make an LLM API call with tool/function calling support.
+
+        Args:
+            messages: List of message dicts. Can include tool results with role='tool'.
+            tools: List of tool definitions in OpenAI format.
+            max_completion_tokens: Maximum tokens in response.
+            temperature: Sampling temperature (0.0-2.0).
+            scope: Scope identifier for tracking.
+            max_retries: Maximum retry attempts.
+            initial_backoff: Initial backoff time in seconds.
+            max_backoff: Maximum backoff time in seconds.
+            tool_choice: How to choose tools - "auto", "none", "required", or specific function.
+
+        Returns:
+            LLMToolCallResult with content and/or tool_calls.
+        """
+        pass
+
+    @abstractmethod
+    async def cleanup(self) -> None:
+        """Clean up resources (close connections, etc.)."""
+        pass
+
+
+class OutputTooLongError(Exception):
+    """
+    Bridge exception raised when LLM output exceeds token limits.
+
+    This wraps provider-specific errors (e.g., OpenAI's LengthFinishReasonError)
+    to allow callers to handle output length issues without depending on
+    provider-specific implementations.
+    """
+
+    pass
@@ -0,0 +1,14 @@
+"""
+LLM provider implementations.
+
+This package contains concrete implementations of the LLMInterface for various providers.
+"""
+
+from .anthropic_llm import AnthropicLLM
+from .claude_code_llm import ClaudeCodeLLM
+from .codex_llm import CodexLLM
+from .gemini_llm import GeminiLLM
+from .mock_llm import MockLLM
+from .openai_compatible_llm import OpenAICompatibleLLM
+
+__all__ = ["AnthropicLLM", "ClaudeCodeLLM", "CodexLLM", "GeminiLLM", "MockLLM", "OpenAICompatibleLLM"]
@@ -0,0 +1,477 @@
+"""
+Anthropic LLM provider using the Anthropic Python SDK.
+
+This provider enables using Claude models from Anthropic with support for:
+- Structured JSON output
+- Tool/function calling with proper format conversion
+- Extended thinking mode
+- Retry logic with exponential backoff
+"""
+
+import asyncio
+import json
+import logging
+import time
+from typing import Any
+
+from hindsight_api.engine.llm_interface import LLMInterface, OutputTooLongError
+from hindsight_api.engine.response_models import LLMToolCall, LLMToolCallResult, TokenUsage
+from hindsight_api.metrics import get_metrics_collector
+
+logger = logging.getLogger(__name__)
+
+
+class AnthropicLLM(LLMInterface):
+    """
+    LLM provider using Anthropic's Claude models.
+
+    Supports structured output, tool calling, and extended thinking mode.
+    Handles format conversion between OpenAI-style messages and Anthropic's format.
+    """
+
+    def __init__(
+        self,
+        provider: str,
+        api_key: str,
+        base_url: str,
+        model: str,
+        reasoning_effort: str = "low",
+        timeout: float = 300.0,
+        **kwargs: Any,
+    ):
+        """
+        Initialize Anthropic LLM provider.
+
+        Args:
+            provider: Provider name (should be "anthropic").
+            api_key: Anthropic API key.
+            base_url: Base URL for the API (optional, uses Anthropic default if empty).
+            model: Model name (e.g., "claude-sonnet-4-20250514").
+            reasoning_effort: Reasoning effort level (not used by Anthropic).
+            timeout: Request timeout in seconds.
+            **kwargs: Additional provider-specific parameters.
+        """
+        super().__init__(provider, api_key, base_url, model, reasoning_effort, **kwargs)
+
+        if not self.api_key:
+            raise ValueError("API key is required for Anthropic provider")
+
+        # Import and initialize Anthropic client
+        try:
+            from anthropic import AsyncAnthropic
+
+            client_kwargs: dict[str, Any] = {"api_key": self.api_key}
+            if self.base_url:
+                client_kwargs["base_url"] = self.base_url
+            if timeout:
+                client_kwargs["timeout"] = timeout
+
+            self._client = AsyncAnthropic(**client_kwargs)
+            logger.info(f"Anthropic client initialized for model: {self.model}")
+        except ImportError as e:
+            raise RuntimeError("Anthropic SDK not installed. Run: uv add anthropic or pip install anthropic") from e
+
+    async def verify_connection(self) -> None:
+        """
+        Verify that the Anthropic provider is configured correctly by making a simple test call.
+
+        Raises:
+            RuntimeError: If the connection test fails.
+        """
+        try:
+            test_messages = [{"role": "user", "content": "test"}]
+            await self.call(
+                messages=test_messages,
+                max_completion_tokens=10,
+                temperature=0.0,
+                scope="verification",
+                max_retries=0,
+            )
+            logger.info("Anthropic connection verified successfully")
+        except Exception as e:
+            logger.error(f"Anthropic connection verification failed: {e}")
+            raise RuntimeError(f"Failed to verify Anthropic connection: {e}") from e
+
+    async def call(
+        self,
+        messages: list[dict[str, str]],
+        response_format: Any | None = None,
+        max_completion_tokens: int | None = None,
+        temperature: float | None = None,
+        scope: str = "memory",
+        max_retries: int = 10,
+        initial_backoff: float = 1.0,
+        max_backoff: float = 60.0,
+        skip_validation: bool = False,
+        strict_schema: bool = False,
+        return_usage: bool = False,
+    ) -> Any:
+        """
+        Make an LLM API call with retry logic.
+
+        Args:
+            messages: List of message dicts with 'role' and 'content'.
+            response_format: Optional Pydantic model for structured output.
+            max_completion_tokens: Maximum tokens in response.
+            temperature: Sampling temperature (0.0-2.0).
+            scope: Scope identifier for tracking.
+            max_retries: Maximum retry attempts.
+            initial_backoff: Initial backoff time in seconds.
+            max_backoff: Maximum backoff time in seconds.
+            skip_validation: Return raw JSON without Pydantic validation.
+            strict_schema: Use strict JSON schema enforcement (not supported by Anthropic).
+            return_usage: If True, return tuple (result, TokenUsage) instead of just result.
+
+        Returns:
+            If return_usage=False: Parsed response if response_format is provided, otherwise text content.
+            If return_usage=True: Tuple of (result, TokenUsage) with token counts.
+
+        Raises:
+            OutputTooLongError: If output exceeds token limits.
+            Exception: Re-raises API errors after retries exhausted.
+        """
+        from anthropic import APIConnectionError, APIStatusError, RateLimitError
+
+        start_time = time.time()
+
+        # Convert OpenAI-style messages to Anthropic format
+        system_prompt = None
+        anthropic_messages = []
+
+        for msg in messages:
+            role = msg.get("role", "user")
+            content = msg.get("content", "")
+
+            if role == "system":
+                if system_prompt:
+                    system_prompt += "\n\n" + content
+                else:
+                    system_prompt = content
+            else:
+                anthropic_messages.append({"role": role, "content": content})
+
+        # Add JSON schema instruction if response_format is provided
+        if response_format is not None and hasattr(response_format, "model_json_schema"):
+            schema = response_format.model_json_schema()
+            schema_msg = f"\n\nYou must respond with valid JSON matching this schema:\n{json.dumps(schema, indent=2)}"
+            if system_prompt:
+                system_prompt += schema_msg
+            else:
+                system_prompt = schema_msg
+
+        # Prepare parameters
+        call_params: dict[str, Any] = {
+            "model": self.model,
+            "messages": anthropic_messages,
+            "max_tokens": max_completion_tokens if max_completion_tokens is not None else 4096,
+        }
+
+        if system_prompt:
+            call_params["system"] = system_prompt
+
+        if temperature is not None:
+            call_params["temperature"] = temperature
+
+        last_exception = None
+
+        for attempt in range(max_retries + 1):
+            try:
+                response = await self._client.messages.create(**call_params)
+
+                # Anthropic response content is a list of blocks
+                content = ""
+                for block in response.content:
+                    if block.type == "text":
+                        content += block.text
+
+                if response_format is not None:
+                    # Models may wrap JSON in markdown code blocks
+                    clean_content = content
+                    if "```json" in content:
+                        clean_content = content.split("```json")[1].split("```")[0].strip()
+                    elif "```" in content:
+                        clean_content = content.split("```")[1].split("```")[0].strip()
+
+                    try:
+                        json_data = json.loads(clean_content)
+                    except json.JSONDecodeError:
+                        # Fallback to parsing raw content if markdown stripping failed
+                        json_data = json.loads(content)
+
+                    if skip_validation:
+                        result = json_data
+                    else:
+                        result = response_format.model_validate(json_data)
+                else:
+                    result = content
+
+                # Record metrics and log slow calls
+                duration = time.time() - start_time
+                input_tokens = response.usage.input_tokens or 0 if response.usage else 0
+                output_tokens = response.usage.output_tokens or 0 if response.usage else 0
+                total_tokens = input_tokens + output_tokens
+
+                # Record LLM metrics
+                metrics = get_metrics_collector()
+                metrics.record_llm_call(
+                    provider=self.provider,
+                    model=self.model,
+                    scope=scope,
+                    duration=duration,
+                    input_tokens=input_tokens,
+                    output_tokens=output_tokens,
+                    success=True,
+                )
+
+                # Record trace span
+                from hindsight_api.tracing import _serialize_for_span, get_span_recorder
+
+                finish_reason = response.stop_reason if hasattr(response, "stop_reason") else None
+                span_recorder = get_span_recorder()
+                span_recorder.record_llm_call(
+                    provider=self.provider,
+                    model=self.model,
+                    scope=scope,
+                    messages=messages,
+                    response_content=_serialize_for_span(result),
+                    input_tokens=input_tokens,
+                    output_tokens=output_tokens,
+                    duration=duration,
+                    finish_reason=finish_reason,
+                    error=None,
+                )
+
+                # Log slow calls
+                if duration > 10.0:
+                    logger.info(
+                        f"slow llm call: scope={scope}, model={self.provider}/{self.model}, "
+                        f"input_tokens={input_tokens}, output_tokens={output_tokens}, "
+                        f"time={duration:.3f}s"
+                    )
+
+                if return_usage:
+                    token_usage = TokenUsage(
+                        input_tokens=input_tokens,
+                        output_tokens=output_tokens,
+                        total_tokens=total_tokens,
+                    )
+                    return result, token_usage
+                return result
+
+            except json.JSONDecodeError as e:
+                last_exception = e
+                if attempt < max_retries:
+                    logger.warning("Anthropic returned invalid JSON, retrying...")
+                    backoff = min(initial_backoff * (2**attempt), max_backoff)
+                    await asyncio.sleep(backoff)
+                    continue
+                else:
+                    logger.error(f"Anthropic returned invalid JSON after {max_retries + 1} attempts")
+                    raise
+
+            except (APIConnectionError, RateLimitError, APIStatusError) as e:
+                # Fast fail on 401/403
+                if isinstance(e, APIStatusError) and e.status_code in (401, 403):
+                    logger.error(f"Anthropic auth error (HTTP {e.status_code}), not retrying: {str(e)}")
+                    raise
+
+                last_exception = e
+                if attempt < max_retries:
+                    # Check if it's a rate limit or server error
+                    should_retry = isinstance(e, (APIConnectionError, RateLimitError)) or (
+                        isinstance(e, APIStatusError) and e.status_code >= 500
+                    )
+
+                    if should_retry:
+                        backoff = min(initial_backoff * (2**attempt), max_backoff)
+                        jitter = backoff * 0.2 * (2 * (time.time() % 1) - 1)
+                        await asyncio.sleep(backoff + jitter)
+                        continue
+
+                logger.error(f"Anthropic API error after {max_retries + 1} attempts: {str(e)}")
+                raise
+
+            except Exception as e:
+                logger.error(f"Unexpected error during Anthropic call: {type(e).__name__}: {str(e)}")
+                raise
+
+        if last_exception:
+            raise last_exception
+        raise RuntimeError("Anthropic call failed after all retries")
+
+    async def call_with_tools(
+        self,
+        messages: list[dict[str, Any]],
+        tools: list[dict[str, Any]],
+        max_completion_tokens: int | None = None,
+        temperature: float | None = None,
+        scope: str = "tools",
+        max_retries: int = 5,
+        initial_backoff: float = 1.0,
+        max_backoff: float = 30.0,
+        tool_choice: str | dict[str, Any] = "auto",
+    ) -> LLMToolCallResult:
+        """
+        Make an LLM API call with tool/function calling support.
+
+        Args:
+            messages: List of message dicts. Can include tool results with role='tool'.
+            tools: List of tool definitions in OpenAI format.
+            max_completion_tokens: Maximum tokens in response.
+            temperature: Sampling temperature (0.0-2.0).
+            scope: Scope identifier for tracking.
+            max_retries: Maximum retry attempts.
+            initial_backoff: Initial backoff time in seconds.
+            max_backoff: Maximum backoff time in seconds.
+            tool_choice: How to choose tools - "auto", "none", "required", or specific function.
+
+        Returns:
+            LLMToolCallResult with content and/or tool_calls.
+        """
+        from anthropic import APIConnectionError, APIStatusError
+
+        start_time = time.time()
+
+        # Convert OpenAI tool format to Anthropic format
+        anthropic_tools = []
+        for tool in tools:
+            func = tool.get("function", {})
+            anthropic_tools.append(
+                {
+                    "name": func.get("name", ""),
+                    "description": func.get("description", ""),
+                    "input_schema": func.get("parameters", {"type": "object", "properties": {}}),
+                }
+            )
+
+        # Convert messages - handle tool results
+        system_prompt = None
+        anthropic_messages = []
+        for msg in messages:
+            role = msg.get("role", "user")
+            content = msg.get("content", "")
+
+            if role == "system":
+                system_prompt = (system_prompt + "\n\n" + content) if system_prompt else content
+            elif role == "tool":
+                # Anthropic uses tool_result blocks
+                anthropic_messages.append(
+                    {
+                        "role": "user",
+                        "content": [
+                            {"type": "tool_result", "tool_use_id": msg.get("tool_call_id", ""), "content": content}
+                        ],
+                    }
+                )
+            elif role == "assistant" and msg.get("tool_calls"):
+                # Convert assistant tool calls
+                tool_use_blocks = []
+                for tc in msg["tool_calls"]:
+                    tool_use_blocks.append(
+                        {
+                            "type": "tool_use",
+                            "id": tc.get("id", ""),
+                            "name": tc.get("function", {}).get("name", ""),
+                            "input": json.loads(tc.get("function", {}).get("arguments", "{}")),
+                        }
+                    )
+                anthropic_messages.append({"role": "assistant", "content": tool_use_blocks})
+            else:
+                anthropic_messages.append({"role": role, "content": content})
+
+        call_params: dict[str, Any] = {
+            "model": self.model,
+            "messages": anthropic_messages,
+            "tools": anthropic_tools,
+            "max_tokens": max_completion_tokens or 4096,
+        }
+        if system_prompt:
+            call_params["system"] = system_prompt
+
+        if temperature is not None:
+            call_params["temperature"] = temperature
+
+        last_exception = None
+        for attempt in range(max_retries + 1):
+            try:
+                response = await self._client.messages.create(**call_params)
+
+                # Extract content and tool calls
+                content_parts = []
+                tool_calls: list[LLMToolCall] = []
+
+                for block in response.content:
+                    if block.type == "text":
+                        content_parts.append(block.text)
+                    elif block.type == "tool_use":
+                        tool_calls.append(LLMToolCall(id=block.id, name=block.name, arguments=block.input or {}))
+
+                content = "".join(content_parts) if content_parts else None
+                finish_reason = "tool_calls" if tool_calls else "stop"
+
+                # Extract token usage
+                input_tokens = response.usage.input_tokens or 0
+                output_tokens = response.usage.output_tokens or 0
+
+                # Record metrics
+                metrics = get_metrics_collector()
+                duration = time.time() - start_time
+                metrics.record_llm_call(
+                    provider=self.provider,
+                    model=self.model,
+                    scope=scope,
+                    duration=duration,
+                    input_tokens=input_tokens,
+                    output_tokens=output_tokens,
+                    success=True,
+                )
+
+                # Record OpenTelemetry span
+                from hindsight_api.tracing import get_span_recorder
+
+                span_recorder = get_span_recorder()
+                # Convert LLMToolCall objects to dicts for span recording
+                tool_calls_dict = (
+                    [{"id": tc.id, "name": tc.name, "arguments": tc.arguments} for tc in tool_calls]
+                    if tool_calls
+                    else None
+                )
+                span_recorder.record_llm_call(
+                    provider=self.provider,
+                    model=self.model,
+                    scope=scope,
+                    messages=messages,
+                    response_content=content,
+                    input_tokens=input_tokens,
+                    output_tokens=output_tokens,
+                    duration=duration,
+                    finish_reason=finish_reason,
+                    error=None,
+                    tool_calls=tool_calls_dict,
+                )
+
+                return LLMToolCallResult(
+                    content=content,
+                    tool_calls=tool_calls,
+                    finish_reason=finish_reason,
+                    input_tokens=input_tokens,
+                    output_tokens=output_tokens,
+                )
+
+            except (APIConnectionError, APIStatusError) as e:
+                if isinstance(e, APIStatusError) and e.status_code in (401, 403):
+                    raise
+                last_exception = e
+                if attempt < max_retries:
+                    await asyncio.sleep(min(initial_backoff * (2**attempt), max_backoff))
+                    continue
+                raise
+
+        if last_exception:
+            raise last_exception
+        raise RuntimeError("Anthropic tool call failed")
+
+    async def cleanup(self) -> None:
+        """Clean up resources (close Anthropic client connections)."""
+        if hasattr(self, "_client") and self._client:
+            await self._client.close()
@@ -0,0 +1,510 @@
+"""
+Claude Code LLM provider using Claude Agent SDK.
+
+This provider enables using Claude Pro/Max subscriptions for API calls
+via the Claude CLI authentication. It uses the Claude Agent SDK which
+automatically handles authentication via `claude auth login` credentials.
+"""
+
+import asyncio
+import json
+import logging
+import time
+from typing import Any
+
+from hindsight_api.engine.llm_interface import LLMInterface, OutputTooLongError
+from hindsight_api.engine.response_models import LLMToolCall, LLMToolCallResult, TokenUsage
+from hindsight_api.metrics import get_metrics_collector
+
+logger = logging.getLogger(__name__)
+
+
+class ClaudeCodeLLM(LLMInterface):
+    """
+    LLM provider using Claude Code authentication.
+
+    Authenticates using Claude Pro/Max credentials via `claude auth login`
+    and makes API calls through the Claude Agent SDK.
+    """
+
+    def __init__(
+        self,
+        provider: str,
+        api_key: str,  # Will be ignored, uses CLI auth
+        base_url: str,
+        model: str,
+        reasoning_effort: str = "low",
+        **kwargs: Any,
+    ):
+        """Initialize Claude Code LLM provider."""
+        super().__init__(provider, api_key, base_url, model, reasoning_effort, **kwargs)
+
+        # Verify Claude Agent SDK is available
+        try:
+            self._verify_claude_code_available()
+            logger.info("Claude Code: Using Claude Agent SDK (authentication via claude auth login)")
+        except Exception as e:
+            raise RuntimeError(
+                f"Failed to initialize Claude Code provider: {e}\n\n"
+                "To set up Claude Code authentication:\n"
+                "1. Install Claude Code CLI: npm install -g @anthropics/claude-code\n"
+                "2. Login with your Pro/Max plan: claude auth login\n"
+                "3. Verify authentication: claude --version\n\n"
+                "Or use a different provider (anthropic, openai, gemini) with API keys."
+            ) from e
+
+        # Metrics collector is imported at module level
+
+    def _verify_claude_code_available(self) -> None:
+        """
+        Verify that Claude Agent SDK can be imported and is properly configured.
+
+        Raises:
+            ImportError: If Claude Agent SDK is not installed.
+            RuntimeError: If Claude Code is not authenticated.
+        """
+        try:
+            # Import Claude Agent SDK
+            # Reduce Claude Agent SDK logging verbosity
+            import logging as sdk_logging
+
+            from claude_agent_sdk import query  # noqa: F401
+
+            sdk_logging.getLogger("claude_agent_sdk").setLevel(sdk_logging.WARNING)
+            sdk_logging.getLogger("claude_agent_sdk._internal").setLevel(sdk_logging.WARNING)
+
+            logger.debug("Claude Agent SDK imported successfully")
+        except ImportError as e:
+            raise ImportError(
+                "Claude Agent SDK not installed. Run: uv add claude-agent-sdk or pip install claude-agent-sdk"
+            ) from e
+
+        # SDK will automatically check for authentication when first used
+        # No need to verify here - let it fail gracefully on first call with helpful error
+
+    async def verify_connection(self) -> None:
+        """
+        Verify that the Claude Code provider is configured correctly by making a simple test call.
+
+        Raises:
+            RuntimeError: If the connection test fails.
+        """
+        try:
+            test_messages = [{"role": "user", "content": "test"}]
+            await self.call(
+                messages=test_messages,
+                max_completion_tokens=10,
+                temperature=0.0,
+                scope="verification",
+                max_retries=0,
+            )
+            logger.info("Claude Code connection verified successfully")
+        except Exception as e:
+            logger.error(f"Claude Code connection verification failed: {e}")
+            raise RuntimeError(f"Failed to verify Claude Code connection: {e}") from e
+
+    async def call(
+        self,
+        messages: list[dict[str, str]],
+        response_format: Any | None = None,
+        max_completion_tokens: int | None = None,
+        temperature: float | None = None,
+        scope: str = "memory",
+        max_retries: int = 10,
+        initial_backoff: float = 1.0,
+        max_backoff: float = 60.0,
+        skip_validation: bool = False,
+        strict_schema: bool = False,
+        return_usage: bool = False,
+    ) -> Any:
+        """
+        Make an LLM API call with retry logic.
+
+        Args:
+            messages: List of message dicts with 'role' and 'content'.
+            response_format: Optional Pydantic model for structured output.
+            max_completion_tokens: Maximum tokens in response (ignored by Claude Agent SDK).
+            temperature: Sampling temperature (ignored by Claude Agent SDK).
+            scope: Scope identifier for tracking.
+            max_retries: Maximum retry attempts.
+            initial_backoff: Initial backoff time in seconds.
+            max_backoff: Maximum backoff time in seconds.
+            skip_validation: Return raw JSON without Pydantic validation.
+            strict_schema: Use strict JSON schema enforcement (not supported).
+            return_usage: If True, return tuple (result, TokenUsage) instead of just result.
+
+        Returns:
+            If return_usage=False: Parsed response if response_format is provided, otherwise text content.
+            If return_usage=True: Tuple of (result, TokenUsage) with estimated token counts.
+
+        Raises:
+            OutputTooLongError: If output exceeds token limits (not supported by Claude Agent SDK).
+            Exception: Re-raises API errors after retries exhausted.
+        """
+        from claude_agent_sdk import AssistantMessage, ClaudeAgentOptions, TextBlock, query
+
+        start_time = time.time()
+
+        # Build system prompt
+        system_prompt = ""
+        user_content = ""
+
+        for msg in messages:
+            role = msg.get("role", "user")
+            content = msg.get("content", "")
+
+            if role == "system":
+                system_prompt += ("\n\n" + content) if system_prompt else content
+            elif role == "user":
+                user_content += ("\n\n" + content) if user_content else content
+            elif role == "assistant":
+                # Claude Agent SDK doesn't support multi-turn easily in query()
+                # For now, prepend assistant messages to user content
+                user_content += f"\n\n[Previous assistant response: {content}]"
+
+        # Add JSON schema instruction if response_format is provided
+        if response_format is not None and hasattr(response_format, "model_json_schema"):
+            schema = response_format.model_json_schema()
+            schema_instruction = (
+                f"\n\nYou must respond with valid JSON matching this schema:\n{json.dumps(schema, indent=2)}\n\n"
+                "Respond with ONLY the JSON, no markdown formatting."
+            )
+            user_content += schema_instruction
+
+        # Configure SDK options
+        options = ClaudeAgentOptions(
+            system_prompt=system_prompt if system_prompt else None,
+            max_turns=1,  # Single-turn for API-style interactions
+            allowed_tools=[],  # Disable tools for standard LLM calls
+        )
+
+        # Call Claude Agent SDK
+        last_exception = None
+        for attempt in range(max_retries + 1):
+            try:
+                # Collect streaming response
+                full_text = ""
+
+                async for message in query(prompt=user_content, options=options):
+                    if isinstance(message, AssistantMessage):
+                        for block in message.content:
+                            if isinstance(block, TextBlock):
+                                full_text += block.text
+
+                # Handle structured output
+                if response_format is not None:
+                    # Models may wrap JSON in markdown
+                    clean_text = full_text
+                    if "```json" in full_text:
+                        clean_text = full_text.split("```json")[1].split("```")[0].strip()
+                    elif "```" in full_text:
+                        clean_text = full_text.split("```")[1].split("```")[0].strip()
+
+                    try:
+                        json_data = json.loads(clean_text)
+                    except json.JSONDecodeError as e:
+                        logger.warning(f"Claude Code JSON parse error (attempt {attempt + 1}/{max_retries + 1}): {e}")
+                        if attempt < max_retries:
+                            backoff = min(initial_backoff * (2**attempt), max_backoff)
+                            await asyncio.sleep(backoff)
+                            last_exception = e
+                            continue
+                        raise
+
+                    if skip_validation:
+                        result = json_data
+                    else:
+                        result = response_format.model_validate(json_data)
+                else:
+                    result = full_text
+
+                # Record metrics
+                duration = time.time() - start_time
+                metrics = get_metrics_collector()
+
+                # Estimate token usage (Claude Agent SDK doesn't report exact counts)
+                # Use character count / 4 as rough estimate (1 token ≈ 4 characters)
+                estimated_input = sum(len(m.get("content", "")) for m in messages) // 4
+                estimated_output = len(full_text) // 4
+
+                metrics.record_llm_call(
+                    provider=self.provider,
+                    model=self.model,
+                    scope=scope,
+                    duration=duration,
+                    input_tokens=estimated_input,
+                    output_tokens=estimated_output,
+                    success=True,
+                )
+
+                # Record trace span
+                from hindsight_api.tracing import get_span_recorder
+
+                span_recorder = get_span_recorder()
+                span_recorder.record_llm_call(
+                    provider=self.provider,
+                    model=self.model,
+                    scope=scope,
+                    messages=messages,
+                    response_content=result if isinstance(result, str) else json.dumps(result),
+                    input_tokens=estimated_input,
+                    output_tokens=estimated_output,
+                    duration=duration,
+                    finish_reason=None,
+                    error=None,
+                )
+
+                # Log slow calls
+                if duration > 10.0:
+                    logger.info(
+                        f"slow llm call: scope={scope}, model={self.provider}/{self.model}, time={duration:.3f}s"
+                    )
+
+                if return_usage:
+                    token_usage = TokenUsage(
+                        input_tokens=estimated_input,
+                        output_tokens=estimated_output,
+                        total_tokens=estimated_input + estimated_output,
+                    )
+                    return result, token_usage
+
+                return result
+
+            except Exception as e:
+                last_exception = e
+
+                # Check for authentication errors
+                error_str = str(e).lower()
+                if "auth" in error_str or "login" in error_str or "credential" in error_str:
+                    logger.error(f"Claude Code authentication error: {e}")
+                    raise RuntimeError(
+                        f"Claude Code authentication failed: {e}\n\n"
+                        "Run 'claude auth login' to authenticate with Claude Pro/Max."
+                    ) from e
+
+                if attempt < max_retries:
+                    backoff = min(initial_backoff * (2**attempt), max_backoff)
+                    logger.warning(f"Claude Code error (attempt {attempt + 1}/{max_retries + 1}): {e}")
+                    await asyncio.sleep(backoff)
+                    continue
+                else:
+                    logger.error(f"Claude Code error after {max_retries + 1} attempts: {e}")
+                    raise
+
+        if last_exception:
+            raise last_exception
+        raise RuntimeError("Claude Code call failed after all retries")
+
+    async def call_with_tools(
+        self,
+        messages: list[dict[str, Any]],
+        tools: list[dict[str, Any]],
+        max_completion_tokens: int | None = None,
+        temperature: float | None = None,
+        scope: str = "tools",
+        max_retries: int = 5,
+        initial_backoff: float = 1.0,
+        max_backoff: float = 30.0,
+        tool_choice: str | dict[str, Any] = "auto",
+    ) -> LLMToolCallResult:
+        """
+        Make an LLM API call with tool/function calling support using Claude Agent SDK.
+
+        This implementation uses ClaudeSDKClient (not query()) because custom tools via
+        SDK MCP servers are only supported with the client. Tools are converted from OpenAI
+        format to SDK MCP tools, and tool names are formatted as mcp__hindsight_tools__{name}.
+
+        Args:
+            messages: List of message dicts. Can include tool results with role='tool'.
+            tools: List of tool definitions in OpenAI format.
+            max_completion_tokens: Maximum tokens in response (not used by Claude Agent SDK).
+            temperature: Sampling temperature (not used by Claude Agent SDK).
+            scope: Scope identifier for tracking.
+            max_retries: Maximum retry attempts.
+            initial_backoff: Initial backoff time in seconds.
+            max_backoff: Maximum backoff time in seconds.
+            tool_choice: How to choose tools (not used by Claude Agent SDK).
+
+        Returns:
+            LLMToolCallResult with content and/or tool_calls.
+        """
+        from claude_agent_sdk import (
+            AssistantMessage,
+            ClaudeAgentOptions,
+            ClaudeSDKClient,
+            SdkMcpTool,
+            TextBlock,
+            ToolUseBlock,
+            create_sdk_mcp_server,
+        )
+
+        start_time = time.time()
+
+        # Convert OpenAI tool format to Claude Agent SDK SdkMcpTool format
+        sdk_tools: list[SdkMcpTool] = []
+        tool_names: list[str] = []
+
+        for tool in tools:
+            func = tool.get("function", {})
+            tool_name = func.get("name", "")
+            tool_description = func.get("description", "")
+            parameters = func.get("parameters", {})
+
+            # Create a handler with proper closure to avoid transport issues
+            def make_handler(name: str):
+                async def handler(args: dict[str, Any]) -> dict[str, Any]:
+                    # Return immediately with success - tool execution happens externally
+                    return {
+                        "content": [
+                            {
+                                "type": "text",
+                                "text": f"[Tool {name} called successfully]",
+                            }
+                        ]
+                    }
+
+                return handler
+
+            sdk_tools.append(
+                SdkMcpTool(
+                    name=tool_name,
+                    description=tool_description,
+                    input_schema=parameters,
+                    handler=make_handler(tool_name),
+                )
+            )
+            tool_names.append(tool_name)
+
+        # Create an MCP server with the tools
+        mcp_server = create_sdk_mcp_server(
+            name="hindsight_tools",
+            version="1.0.0",
+            tools=sdk_tools if sdk_tools else None,
+        )
+
+        # Build system prompt and user content from messages
+        system_prompt = ""
+        user_content = ""
+
+        for msg in messages:
+            role = msg.get("role", "user")
+            content = msg.get("content", "")
+
+            if role == "system":
+                system_prompt += ("\n\n" + content) if system_prompt else content
+            elif role == "user":
+                user_content += ("\n\n" + content) if user_content else content
+            elif role == "assistant":
+                # Include previous assistant messages as context
+                user_content += f"\n\n[Previous assistant response: {content}]"
+            elif role == "tool":
+                # Tool results are already in tool_results_map, append to user context
+                tool_call_id = msg.get("tool_call_id", "")
+                user_content += f"\n\n[Tool result for {tool_call_id}: {content}]"
+
+        # Format tool names for SDK MCP servers: mcp__{server_name}__{tool_name}
+        # This is required by the Claude Agent SDK for MCP server tools
+        allowed_tool_names = [f"mcp__hindsight_tools__{name}" for name in tool_names]
+
+        # Configure SDK options with MCP server
+        options = ClaudeAgentOptions(
+            system_prompt=system_prompt if system_prompt else None,
+            max_turns=1,  # Single-turn for API-style interactions
+            mcp_servers={"hindsight_tools": mcp_server} if sdk_tools else {},
+            allowed_tools=allowed_tool_names if allowed_tool_names else [],
+        )
+
+        # Call Claude Agent SDK with retry logic
+        last_exception = None
+        for attempt in range(max_retries + 1):
+            try:
+                full_text = ""
+                tool_calls: list[LLMToolCall] = []
+
+                # Use ClaudeSDKClient for tool calling support
+                # Note: query() does NOT support custom tools, only ClaudeSDKClient does
+                async with ClaudeSDKClient(options=options) as client:
+                    # Send the query
+                    await client.query(user_content)
+
+                    # Receive response
+                    async for message in client.receive_response():
+                        if isinstance(message, AssistantMessage):
+                            for block in message.content:
+                                if isinstance(block, TextBlock):
+                                    full_text += block.text
+                                elif isinstance(block, ToolUseBlock):
+                                    # SDK returns tool names with MCP prefix (mcp__hindsight_tools__{name})
+                                    # Strip the prefix to return original tool name expected by caller
+                                    tool_name = block.name
+                                    if tool_name.startswith("mcp__hindsight_tools__"):
+                                        tool_name = tool_name.replace("mcp__hindsight_tools__", "", 1)
+
+                                    tool_calls.append(
+                                        LLMToolCall(
+                                            id=block.id,
+                                            name=tool_name,
+                                            arguments=block.input,
+                                        )
+                                    )
+
+                # Record metrics
+                duration = time.time() - start_time
+                metrics = get_metrics_collector()
+
+                # Estimate token usage (Claude Agent SDK doesn't report exact counts)
+                estimated_input = sum(len(m.get("content", "")) for m in messages) // 4
+                estimated_output = len(full_text) // 4
+
+                metrics.record_llm_call(
+                    provider=self.provider,
+                    model=self.model,
+                    scope=scope,
+                    duration=duration,
+                    input_tokens=estimated_input,
+                    output_tokens=estimated_output,
+                    success=True,
+                )
+
+                # Log slow calls
+                if duration > 10.0:
+                    logger.info(
+                        f"slow llm call: scope={scope}, model={self.provider}/{self.model}, time={duration:.3f}s"
+                    )
+
+                return LLMToolCallResult(
+                    content=full_text if full_text else None,
+                    tool_calls=tool_calls,
+                    finish_reason="tool_calls" if tool_calls else "stop",
+                    input_tokens=estimated_input,
+                    output_tokens=estimated_output,
+                )
+
+            except Exception as e:
+                last_exception = e
+
+                # Check for authentication errors
+                error_str = str(e).lower()
+                if "auth" in error_str or "login" in error_str or "credential" in error_str:
+                    logger.error(f"Claude Code authentication error: {e}")
+                    raise RuntimeError(
+                        f"Claude Code authentication failed: {e}\n\n"
+                        "Run 'claude auth login' to authenticate with Claude Pro/Max."
+                    ) from e
+
+                if attempt < max_retries:
+                    backoff = min(initial_backoff * (2**attempt), max_backoff)
+                    logger.warning(f"Claude Code tool call error (attempt {attempt + 1}/{max_retries + 1}): {e}")
+                    await asyncio.sleep(backoff)
+                    continue
+                else:
+                    logger.error(f"Claude Code tool call error after {max_retries + 1} attempts: {e}")
+                    raise
+
+        if last_exception:
+            raise last_exception
+        raise RuntimeError("Claude Code tool call failed after all retries")
+
+    async def cleanup(self) -> None:
+        """Clean up resources (no HTTP client to close for Claude Agent SDK)."""
+        pass
@@ -0,0 +1,621 @@
+"""
+OpenAI Codex LLM provider using ChatGPT Plus/Pro OAuth authentication.
+
+This provider enables using ChatGPT Plus/Pro subscriptions for API calls
+without separate OpenAI Platform API credits. It uses OAuth tokens from
+~/.codex/auth.json and communicates with the ChatGPT backend API.
+"""
+
+import asyncio
+import json
+import logging
+import os
+import time
+import uuid
+from pathlib import Path
+from typing import Any
+
+import httpx
+
+from hindsight_api.engine.llm_interface import LLMInterface, OutputTooLongError
+from hindsight_api.engine.response_models import LLMToolCall, LLMToolCallResult, TokenUsage
+from hindsight_api.metrics import get_metrics_collector
+
+logger = logging.getLogger(__name__)
+
+
+class CodexLLM(LLMInterface):
+    """
+    LLM provider using OpenAI Codex OAuth authentication.
+
+    Authenticates using ChatGPT Plus/Pro credentials stored in ~/.codex/auth.json
+    and makes API calls to chatgpt.com/backend-api/codex/responses.
+    """
+
+    def __init__(
+        self,
+        provider: str,
+        api_key: str,  # Will be ignored, reads from ~/.codex/auth.json
+        base_url: str,
+        model: str,
+        reasoning_effort: str = "low",
+        **kwargs: Any,
+    ):
+        """Initialize Codex LLM provider."""
+        super().__init__(provider, api_key, base_url, model, reasoning_effort, **kwargs)
+
+        # Load Codex OAuth credentials
+        try:
+            self.access_token, self.account_id = self._load_codex_auth()
+            logger.info(f"Loaded Codex OAuth credentials for account: {self.account_id}")
+        except Exception as e:
+            raise RuntimeError(
+                f"Failed to load Codex OAuth credentials from ~/.codex/auth.json: {e}\n\n"
+                "To set up Codex authentication:\n"
+                "1. Install Codex CLI: npm install -g @openai/codex\n"
+                "2. Login: codex auth login\n"
+                "3. Verify: ls ~/.codex/auth.json\n\n"
+                "Or use a different provider (openai, anthropic, gemini) with API keys."
+            ) from e
+
+        # Use ChatGPT backend API endpoint
+        if not self.base_url:
+            self.base_url = "https://chatgpt.com/backend-api"
+
+        # Normalize model name (strip openai/ prefix if present)
+        if self.model.startswith("openai/"):
+            self.model = self.model[len("openai/") :]
+
+        # Map reasoning effort to Codex reasoning summary format
+        # Codex supports: "auto", "concise", "detailed"
+        self.reasoning_summary = self._map_reasoning_effort(reasoning_effort)
+
+        # HTTP client for SSE streaming
+        self._client = httpx.AsyncClient(timeout=120.0)
+
+    def _load_codex_auth(self) -> tuple[str, str]:
+        """
+        Load OAuth credentials from ~/.codex/auth.json.
+
+        Returns:
+            Tuple of (access_token, account_id).
+
+        Raises:
+            FileNotFoundError: If auth file doesn't exist.
+            ValueError: If auth file is invalid.
+        """
+        auth_file = Path.home() / ".codex" / "auth.json"
+
+        if not auth_file.exists():
+            raise FileNotFoundError(
+                f"Codex auth file not found: {auth_file}\nRun 'codex auth login' to authenticate with ChatGPT Plus/Pro."
+            )
+
+        with open(auth_file) as f:
+            data = json.load(f)
+
+        # Validate auth structure
+        auth_mode = data.get("auth_mode")
+        if auth_mode != "chatgpt":
+            raise ValueError(f"Expected auth_mode='chatgpt', got: {auth_mode}")
+
+        tokens = data.get("tokens", {})
+        access_token = tokens.get("access_token")
+        account_id = tokens.get("account_id")
+
+        if not access_token:
+            raise ValueError("No access_token found in Codex auth file. Run 'codex auth login' again.")
+
+        return access_token, account_id
+
+    def _map_reasoning_effort(self, effort: str) -> str:
+        """
+        Map standard reasoning effort to Codex reasoning summary format.
+
+        Args:
+            effort: Standard effort level ("low", "medium", "high", "xhigh").
+
+        Returns:
+            Codex reasoning summary: "concise", "detailed", or "auto".
+        """
+        mapping = {
+            "low": "concise",
+            "medium": "auto",
+            "high": "detailed",
+            "xhigh": "detailed",
+        }
+        return mapping.get(effort.lower(), "auto")
+
+    async def verify_connection(self) -> None:
+        """Verify Codex connection by making a simple test call."""
+        try:
+            logger.info(f"Verifying Codex LLM: model={self.model}, account={self.account_id}...")
+            await self.call(
+                messages=[{"role": "user", "content": "Say 'ok'"}],
+                max_completion_tokens=10,
+                max_retries=2,
+                initial_backoff=0.5,
+                max_backoff=2.0,
+                scope="verification",
+            )
+            logger.info(f"Codex LLM verified: {self.model}")
+        except Exception as e:
+            raise RuntimeError(f"Codex LLM connection verification failed for {self.model}: {e}") from e
+
+    async def call(
+        self,
+        messages: list[dict[str, str]],
+        response_format: Any | None = None,
+        max_completion_tokens: int | None = None,
+        temperature: float | None = None,
+        scope: str = "memory",
+        max_retries: int = 10,
+        initial_backoff: float = 1.0,
+        max_backoff: float = 60.0,
+        skip_validation: bool = False,
+        strict_schema: bool = False,
+        return_usage: bool = False,
+    ) -> Any:
+        """Make API call to Codex backend with SSE streaming."""
+        start_time = time.time()
+
+        # Prepare system instructions
+        system_instruction = ""
+        user_messages = []
+
+        for msg in messages:
+            role = msg.get("role", "user")
+            content = msg.get("content", "")
+
+            if role == "system":
+                system_instruction += ("\n\n" + content) if system_instruction else content
+            else:
+                user_messages.append(msg)
+
+        # Add JSON schema instruction if response_format is provided
+        if response_format is not None and hasattr(response_format, "model_json_schema"):
+            schema = response_format.model_json_schema()
+            schema_msg = f"\n\nYou must respond with valid JSON matching this schema:\n{json.dumps(schema, indent=2)}"
+            system_instruction += schema_msg
+
+        # gpt-5.2-codex only supports "detailed" reasoning summary
+        reasoning_summary = "detailed" if "5.2" in self.model else self.reasoning_summary
+
+        # Build Codex request payload
+        payload = {
+            "model": self.model,
+            "instructions": system_instruction,
+            "input": [
+                {
+                    "type": "message",
+                    "role": msg.get("role", "user"),
+                    "content": msg.get("content", ""),
+                }
+                for msg in user_messages
+            ],
+            "tools": [],
+            "tool_choice": "auto",
+            "parallel_tool_calls": True,
+            "reasoning": {"summary": reasoning_summary},
+            "store": False,  # Codex uses stateless mode
+            "stream": True,  # SSE streaming
+            "include": ["reasoning.encrypted_content"],
+            "prompt_cache_key": str(uuid.uuid4()),
+        }
+
+        headers = {
+            "Authorization": f"Bearer {self.access_token}",
+            "Content-Type": "application/json",
+            "OpenAI-Account-ID": self.account_id,
+            "User-Agent": "Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7)",
+            "Origin": "https://chatgpt.com",
+        }
+
+        url = f"{self.base_url}/codex/responses"
+        last_exception = None
+
+        for attempt in range(max_retries + 1):
+            try:
+                response = await self._client.post(url, json=payload, headers=headers, timeout=120.0)
+                response.raise_for_status()
+
+                # Parse SSE stream
+                content = await self._parse_sse_stream(response)
+
+                # Handle structured output
+                if response_format is not None:
+                    # Models may wrap JSON in markdown
+                    clean_content = content
+                    if "```json" in content:
+                        clean_content = content.split("```json")[1].split("```")[0].strip()
+                    elif "```" in content:
+                        clean_content = content.split("```")[1].split("```")[0].strip()
+
+                    try:
+                        json_data = json.loads(clean_content)
+                    except json.JSONDecodeError as e:
+                        logger.warning(f"Codex JSON parse error (attempt {attempt + 1}/{max_retries + 1}): {e}")
+                        if attempt < max_retries:
+                            backoff = min(initial_backoff * (2**attempt), max_backoff)
+                            await asyncio.sleep(backoff)
+                            last_exception = e
+                            continue
+                        raise
+
+                    if skip_validation:
+                        result = json_data
+                    else:
+                        result = response_format.model_validate(json_data)
+                else:
+                    result = content
+
+                # Record metrics
+                duration = time.time() - start_time
+                metrics = get_metrics_collector()
+                metrics.record_llm_call(
+                    provider=self.provider,
+                    model=self.model,
+                    scope=scope,
+                    duration=duration,
+                    input_tokens=0,  # Codex doesn't report token counts in SSE
+                    output_tokens=0,
+                    success=True,
+                )
+
+                # Record trace span
+                from hindsight_api.tracing import get_span_recorder
+
+                # Estimate tokens for tracing
+                estimated_input = sum(len(m.get("content", "")) for m in messages) // 4
+                estimated_output = len(content) // 4
+                span_recorder = get_span_recorder()
+                span_recorder.record_llm_call(
+                    provider=self.provider,
+                    model=self.model,
+                    scope=scope,
+                    messages=messages,
+                    response_content=result if isinstance(result, str) else json.dumps(result),
+                    input_tokens=estimated_input,
+                    output_tokens=estimated_output,
+                    duration=duration,
+                    finish_reason=None,
+                    error=None,
+                )
+
+                if return_usage:
+                    # Codex doesn't provide token counts, estimate based on content
+                    estimated_input = sum(len(m.get("content", "")) for m in messages) // 4
+                    estimated_output = len(content) // 4
+                    token_usage = TokenUsage(
+                        input_tokens=estimated_input,
+                        output_tokens=estimated_output,
+                        total_tokens=estimated_input + estimated_output,
+                    )
+                    return result, token_usage
+
+                return result
+
+            except httpx.HTTPStatusError as e:
+                last_exception = e
+                status_code = e.response.status_code
+
+                # Fast fail on auth errors
+                if status_code in (401, 403):
+                    logger.error(f"Codex auth error (HTTP {status_code}): {e.response.text[:200]}")
+                    raise RuntimeError(
+                        "Codex authentication failed. Your OAuth token may have expired.\n"
+                        "Run 'codex auth login' to re-authenticate."
+                    ) from e
+
+                # Log the actual error message from the API
+                error_detail = e.response.text[:500] if hasattr(e.response, "text") else str(e)
+
+                if attempt < max_retries:
+                    backoff = min(initial_backoff * (2**attempt), max_backoff)
+                    logger.warning(
+                        f"Codex HTTP error {status_code} (attempt {attempt + 1}/{max_retries + 1}): {error_detail}"
+                    )
+                    await asyncio.sleep(backoff)
+                    continue
+                else:
+                    logger.error(
+                        f"Codex HTTP error after {max_retries + 1} attempts: Status {status_code}, Detail: {error_detail}"
+                    )
+                    raise
+
+            except httpx.RequestError as e:
+                last_exception = e
+                if attempt < max_retries:
+                    backoff = min(initial_backoff * (2**attempt), max_backoff)
+                    logger.warning(f"Codex connection error (attempt {attempt + 1}/{max_retries + 1}): {e}")
+                    await asyncio.sleep(backoff)
+                    continue
+                else:
+                    logger.error(f"Codex connection error after {max_retries + 1} attempts: {e}")
+                    raise
+
+            except Exception as e:
+                logger.error(f"Unexpected Codex error: {type(e).__name__}: {e}")
+                raise
+
+        if last_exception:
+            raise last_exception
+        raise RuntimeError("Codex call failed after all retries")
+
+    async def _parse_sse_stream(self, response: httpx.Response) -> str:
+        """
+        Parse Server-Sent Events (SSE) stream from Codex API.
+
+        Args:
+            response: HTTP response with SSE stream.
+
+        Returns:
+            Extracted text content from stream.
+        """
+        full_text = ""
+        event_type = None
+
+        async for line in response.aiter_lines():
+            if not line:
+                continue
+
+            # Track event type
+            if line.startswith("event: "):
+                event_type = line[7:]
+
+            # Parse data
+            elif line.startswith("data: "):
+                data_str = line[6:]
+                if data_str == "[DONE]":
+                    break
+
+                try:
+                    data = json.loads(data_str)
+
+                    # Extract content based on event type
+                    if event_type == "response.text.delta" and "delta" in data:
+                        full_text += data["delta"]
+                    elif event_type == "response.content_part.delta" and "delta" in data:
+                        full_text += data["delta"]
+                    # Check for item content
+                    elif "item" in data:
+                        item = data["item"]
+                        if "content" in item:
+                            content = item["content"]
+                            if isinstance(content, list):
+                                for part in content:
+                                    if isinstance(part, dict) and "text" in part:
+                                        full_text += part["text"]
+                            elif isinstance(content, str):
+                                full_text += content
+
+                except json.JSONDecodeError:
+                    # Skip malformed JSON events
+                    pass
+
+        return full_text
+
+    async def call_with_tools(
+        self,
+        messages: list[dict[str, Any]],
+        tools: list[dict[str, Any]],
+        max_completion_tokens: int | None = None,
+        temperature: float | None = None,
+        scope: str = "tools",
+        max_retries: int = 5,
+        initial_backoff: float = 1.0,
+        max_backoff: float = 30.0,
+        tool_choice: str | dict[str, Any] = "auto",
+    ) -> LLMToolCallResult:
+        """
+        Make API call with tool calling support.
+
+        Parses Codex SSE stream to extract tool calls from response.output_item.done events.
+        Tools are converted from OpenAI format to Codex format (flat structure at top level).
+
+        Args:
+            messages: List of message dicts. Can include tool results with role='tool'.
+            tools: List of tool definitions in OpenAI format.
+            max_completion_tokens: Maximum tokens in response.
+            temperature: Sampling temperature.
+            scope: Scope identifier for tracking.
+            max_retries: Maximum retry attempts.
+            initial_backoff: Initial backoff time in seconds.
+            max_backoff: Maximum backoff time in seconds.
+            tool_choice: How to choose tools - "auto", "none", "required", or specific function.
+
+        Returns:
+            LLMToolCallResult with content and/or tool_calls.
+        """
+        start_time = time.time()
+
+        # Prepare system instructions
+        system_instruction = ""
+        user_messages = []
+
+        for msg in messages:
+            role = msg.get("role", "user")
+            content = msg.get("content", "")
+
+            if role == "system":
+                system_instruction += ("\n\n" + content) if system_instruction else content
+            elif role == "tool":
+                # Handle tool results
+                user_messages.append(
+                    {
+                        "type": "message",
+                        "role": "user",
+                        "content": f"Tool result: {content}",
+                    }
+                )
+            else:
+                user_messages.append(
+                    {
+                        "type": "message",
+                        "role": role,
+                        "content": content,
+                    }
+                )
+
+        # Convert tools to Codex format
+        # Codex expects tools with type and name/description/parameters at top level
+        codex_tools = []
+        for tool in tools:
+            func = tool.get("function", {})
+            codex_tools.append(
+                {
+                    "type": "function",
+                    "name": func.get("name", ""),
+                    "description": func.get("description", ""),
+                    "parameters": func.get("parameters", {}),
+                }
+            )
+
+        # gpt-5.2-codex only supports "detailed" reasoning summary
+        reasoning_summary = "detailed" if "5.2" in self.model else self.reasoning_summary
+
+        payload = {
+            "model": self.model,
+            "instructions": system_instruction,
+            "input": user_messages,
+            "tools": codex_tools,
+            "tool_choice": tool_choice,
+            "parallel_tool_calls": True,
+            "reasoning": {"summary": reasoning_summary},
+            "store": False,
+            "stream": True,
+            "include": ["reasoning.encrypted_content"],
+            "prompt_cache_key": str(uuid.uuid4()),
+        }
+
+        headers = {
+            "Authorization": f"Bearer {self.access_token}",
+            "Content-Type": "application/json",
+            "OpenAI-Account-ID": self.account_id,
+            "User-Agent": "Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7)",
+            "Origin": "https://chatgpt.com",
+        }
+
+        url = f"{self.base_url}/codex/responses"
+
+        # Debug logging for troubleshooting
+        logger.debug(f"Codex tool call request: url={url}, model={payload['model']}, tools={len(codex_tools)}")
+
+        try:
+            response = await self._client.post(url, json=payload, headers=headers, timeout=120.0)
+
+            # Log response details on error
+            if response.status_code != 200:
+                logger.error(f"Codex API error {response.status_code}: {response.text[:500]}")
+
+            response.raise_for_status()
+
+            # Parse SSE for tool calls and content
+            content, tool_calls = await self._parse_sse_tool_stream(response)
+
+            duration = time.time() - start_time
+            metrics = get_metrics_collector()
+            metrics.record_llm_call(
+                provider=self.provider,
+                model=self.model,
+                scope=scope,
+                duration=duration,
+                input_tokens=0,
+                output_tokens=0,
+                success=True,
+            )
+
+            # Record OpenTelemetry span
+            from hindsight_api.tracing import get_span_recorder
+
+            span_recorder = get_span_recorder()
+            # Convert LLMToolCall objects to dicts for span recording
+            tool_calls_dict = (
+                [{"id": tc.id, "name": tc.name, "arguments": tc.arguments} for tc in tool_calls] if tool_calls else None
+            )
+            span_recorder.record_llm_call(
+                provider=self.provider,
+                model=self.model,
+                scope=scope,
+                messages=messages,
+                response_content=content,
+                input_tokens=0,  # Codex doesn't provide token counts
+                output_tokens=0,
+                duration=duration,
+                finish_reason="tool_calls" if tool_calls else "stop",
+                error=None,
+                tool_calls=tool_calls_dict,
+            )
+
+            return LLMToolCallResult(
+                content=content,
+                tool_calls=tool_calls,
+                finish_reason="tool_calls" if tool_calls else "stop",
+                input_tokens=0,
+                output_tokens=0,
+            )
+
+        except Exception as e:
+            logger.error(f"Codex tool call error: {e}")
+            raise
+
+    async def _parse_sse_tool_stream(self, response: httpx.Response) -> tuple[str | None, list[LLMToolCall]]:
+        """
+        Parse SSE stream for tool calls and content.
+
+        Returns:
+            Tuple of (content, tool_calls).
+        """
+        content = ""
+        tool_calls: list[LLMToolCall] = []
+        event_type = None
+
+        async for line in response.aiter_lines():
+            if not line:
+                continue
+
+            if line.startswith("event: "):
+                event_type = line[7:]
+
+            elif line.startswith("data: "):
+                data_str = line[6:]
+                if data_str == "[DONE]":
+                    break
+
+                try:
+                    data = json.loads(data_str)
+
+                    # Extract text content
+                    if event_type == "response.text.delta" and "delta" in data:
+                        content += data["delta"]
+
+                    # Extract completed tool calls from response.output_item.done
+                    elif event_type == "response.output_item.done":
+                        item = data.get("item", {})
+                        if item.get("type") == "function_call" and item.get("status") == "completed":
+                            tool_name = item.get("name", "")
+                            arguments_str = item.get("arguments", "{}")
+                            call_id = item.get("call_id", "")
+
+                            try:
+                                arguments = json.loads(arguments_str)
+                            except json.JSONDecodeError:
+                                logger.warning(f"Failed to parse tool arguments: {arguments_str}")
+                                arguments = {}
+
+                            tool_calls.append(
+                                LLMToolCall(
+                                    id=call_id,
+                                    name=tool_name,
+                                    arguments=arguments,
+                                )
+                            )
+
+                except json.JSONDecodeError as e:
+                    logger.warning(f"Failed to parse SSE data: {e}, data_str: {data_str[:200]}")
+
+        return content if content else None, tool_calls
+
+    async def cleanup(self) -> None:
+        """Clean up HTTP client."""
+        await self._client.aclose()
@@ -0,0 +1,550 @@
+"""
+Google Gemini/VertexAI LLM provider.
+
+This provider supports both:
+1. Gemini API (api.generativeai.google.com) with API key authentication
+2. Vertex AI with service account or Application Default Credentials (ADC)
+"""
+
+import asyncio
+import json
+import logging
+import os
+import time
+from typing import Any
+
+from google import genai
+from google.genai import errors as genai_errors
+from google.genai import types as genai_types
+
+from hindsight_api.engine.llm_interface import LLMInterface, OutputTooLongError
+from hindsight_api.engine.response_models import LLMToolCall, LLMToolCallResult, TokenUsage
+from hindsight_api.metrics import get_metrics_collector
+
+logger = logging.getLogger(__name__)
+
+# Vertex AI imports (optional)
+try:
+    import google.auth
+    from google.oauth2 import service_account
+
+    VERTEXAI_AVAILABLE = True
+except ImportError:
+    VERTEXAI_AVAILABLE = False
+
+
+class GeminiLLM(LLMInterface):
+    """
+    LLM provider for Google Gemini and Vertex AI.
+
+    Supports:
+    - Gemini API: provider="gemini", requires api_key
+    - Vertex AI: provider="vertexai", requires project_id and region, uses ADC or service account
+    """
+
+    def __init__(
+        self,
+        provider: str,
+        api_key: str,
+        base_url: str,
+        model: str,
+        reasoning_effort: str = "low",
+        **kwargs: Any,
+    ):
+        """Initialize Gemini/VertexAI LLM provider."""
+        super().__init__(provider, api_key, base_url, model, reasoning_effort, **kwargs)
+
+        self._client = None
+        self._is_vertexai = self.provider == "vertexai"
+
+        if self._is_vertexai:
+            self._init_vertexai(**kwargs)
+        else:
+            self._init_gemini()
+
+    def _init_gemini(self) -> None:
+        """Initialize Gemini API client."""
+        if not self.api_key:
+            raise ValueError("Gemini provider requires api_key")
+
+        self._client = genai.Client(api_key=self.api_key)
+        logger.info(f"Gemini API: model={self.model}")
+
+    def _init_vertexai(self, **kwargs: Any) -> None:
+        """Initialize Vertex AI client with project, region, and credentials."""
+        # Extract Vertex AI config from kwargs
+        project_id = kwargs.get("vertexai_project_id")
+        region = kwargs.get("vertexai_region", "us-central1")
+        service_account_key = kwargs.get("vertexai_service_account_key")
+        credentials = kwargs.get("vertexai_credentials")  # Pre-loaded credentials object
+
+        if not project_id:
+            raise ValueError(
+                "HINDSIGHT_API_LLM_VERTEXAI_PROJECT_ID is required for Vertex AI provider. "
+                "Set it to your GCP project ID."
+            )
+
+        auth_method = "ADC"
+
+        # Use pre-loaded credentials if provided (passed from LLMProvider)
+        if credentials is not None:
+            auth_method = "service_account"
+        # Otherwise, load explicit service account credentials if path provided
+        elif service_account_key:
+            if not VERTEXAI_AVAILABLE:
+                raise ValueError(
+                    "Vertex AI service account auth requires 'google-auth' package. "
+                    "Install with: pip install google-auth"
+                )
+            credentials = service_account.Credentials.from_service_account_file(
+                service_account_key,
+                scopes=["https://www.googleapis.com/auth/cloud-platform"],
+            )
+            auth_method = "service_account"
+            logger.info(f"Vertex AI: Using service account key: {service_account_key}")
+
+        # Strip google/ prefix from model name — native SDK uses bare names
+        # e.g. "google/gemini-2.0-flash-lite-001" -> "gemini-2.0-flash-lite-001"
+        if self.model.startswith("google/"):
+            self.model = self.model[len("google/") :]
+
+        # Create Vertex AI client
+        client_kwargs: dict[str, Any] = {
+            "vertexai": True,
+            "project": project_id,
+            "location": region,
+        }
+        if credentials is not None:
+            client_kwargs["credentials"] = credentials
+
+        self._client = genai.Client(**client_kwargs)
+
+        logger.info(f"Vertex AI: project={project_id}, region={region}, model={self.model}, auth={auth_method}")
+
+    async def verify_connection(self) -> None:
+        """
+        Verify that the Gemini/VertexAI provider is configured correctly.
+
+        Raises:
+            RuntimeError: If the connection test fails.
+        """
+        try:
+            logger.info(f"Verifying {self.provider.upper()}: model={self.model}...")
+            await self.call(
+                messages=[{"role": "user", "content": "Say 'ok'"}],
+                max_completion_tokens=100,
+                max_retries=2,
+                initial_backoff=0.5,
+                max_backoff=2.0,
+                scope="verification",
+            )
+            logger.info(f"{self.provider.upper()} connection verified successfully")
+        except Exception as e:
+            raise RuntimeError(f"Failed to verify {self.provider.upper()} connection: {e}") from e
+
+    async def call(
+        self,
+        messages: list[dict[str, str]],
+        response_format: Any | None = None,
+        max_completion_tokens: int | None = None,
+        temperature: float | None = None,
+        scope: str = "memory",
+        max_retries: int = 10,
+        initial_backoff: float = 1.0,
+        max_backoff: float = 60.0,
+        skip_validation: bool = False,
+        strict_schema: bool = False,
+        return_usage: bool = False,
+    ) -> Any:
+        """
+        Make a Gemini/VertexAI API call with retry logic.
+
+        Args:
+            messages: List of message dicts with 'role' and 'content'.
+            response_format: Optional Pydantic model for structured output.
+            max_completion_tokens: Maximum tokens in response (not supported by Gemini).
+            temperature: Sampling temperature (0.0-2.0).
+            scope: Scope identifier for tracking.
+            max_retries: Maximum retry attempts.
+            initial_backoff: Initial backoff time in seconds.
+            max_backoff: Maximum backoff time in seconds.
+            skip_validation: Return raw JSON without Pydantic validation.
+            strict_schema: Use strict JSON schema enforcement (not supported by Gemini).
+            return_usage: If True, return tuple (result, TokenUsage).
+
+        Returns:
+            If return_usage=False: Parsed response if response_format provided, else text.
+            If return_usage=True: Tuple of (result, TokenUsage).
+        """
+        start_time = time.time()
+
+        # Convert OpenAI-style messages to Gemini format
+        system_instruction = None
+        gemini_contents = []
+
+        for msg in messages:
+            role = msg.get("role", "user")
+            content = msg.get("content", "")
+
+            if role == "system":
+                if system_instruction:
+                    system_instruction += "\n\n" + content
+                else:
+                    system_instruction = content
+            elif role == "assistant":
+                gemini_contents.append(genai_types.Content(role="model", parts=[genai_types.Part(text=content)]))
+            else:
+                gemini_contents.append(genai_types.Content(role="user", parts=[genai_types.Part(text=content)]))
+
+        # Add JSON schema instruction if response_format is provided
+        if response_format is not None and hasattr(response_format, "model_json_schema"):
+            schema = response_format.model_json_schema()
+            schema_msg = f"\n\nYou must respond with valid JSON matching this schema:\n{json.dumps(schema, indent=2)}"
+            if system_instruction:
+                system_instruction += schema_msg
+            else:
+                system_instruction = schema_msg
+
+        # Build generation config
+        config_kwargs: dict[str, Any] = {}
+        if system_instruction:
+            config_kwargs["system_instruction"] = system_instruction
+        if response_format is not None:
+            config_kwargs["response_mime_type"] = "application/json"
+            config_kwargs["response_schema"] = response_format
+        if temperature is not None:
+            config_kwargs["temperature"] = temperature
+
+        generation_config = genai_types.GenerateContentConfig(**config_kwargs) if config_kwargs else None
+
+        last_exception = None
+
+        for attempt in range(max_retries + 1):
+            try:
+                response = await self._client.aio.models.generate_content(
+                    model=self.model,
+                    contents=gemini_contents,
+                    config=generation_config,
+                )
+
+                content = response.text
+
+                # Handle empty response
+                if content is None:
+                    block_reason = None
+                    if hasattr(response, "candidates") and response.candidates:
+                        candidate = response.candidates[0]
+                        if hasattr(candidate, "finish_reason"):
+                            block_reason = candidate.finish_reason
+
+                    if attempt < max_retries:
+                        logger.warning(f"Gemini returned empty response (reason: {block_reason}), retrying...")
+                        backoff = min(initial_backoff * (2**attempt), max_backoff)
+                        await asyncio.sleep(backoff)
+                        continue
+                    else:
+                        raise RuntimeError(f"Gemini returned empty response after {max_retries + 1} attempts")
+
+                # Parse structured output if requested
+                if response_format is not None:
+                    json_data = json.loads(content)
+                    if skip_validation:
+                        result = json_data
+                    else:
+                        result = response_format.model_validate(json_data)
+                else:
+                    result = content
+
+                # Extract token usage
+                input_tokens = 0
+                output_tokens = 0
+                if hasattr(response, "usage_metadata") and response.usage_metadata:
+                    usage = response.usage_metadata
+                    input_tokens = usage.prompt_token_count or 0
+                    output_tokens = usage.candidates_token_count or 0
+
+                # Record metrics
+                duration = time.time() - start_time
+                metrics = get_metrics_collector()
+                metrics.record_llm_call(
+                    provider=self.provider,
+                    model=self.model,
+                    scope=scope,
+                    duration=duration,
+                    input_tokens=input_tokens,
+                    output_tokens=output_tokens,
+                    success=True,
+                )
+
+                # Record trace span
+                from hindsight_api.tracing import get_span_recorder
+
+                finish_reason = None
+                if hasattr(response, "candidates") and response.candidates:
+                    if hasattr(response.candidates[0], "finish_reason"):
+                        finish_reason = str(response.candidates[0].finish_reason)
+                span_recorder = get_span_recorder()
+                from hindsight_api.tracing import _serialize_for_span
+
+                span_recorder.record_llm_call(
+                    provider=self.provider,
+                    model=self.model,
+                    scope=scope,
+                    messages=messages,
+                    response_content=_serialize_for_span(result),
+                    input_tokens=input_tokens,
+                    output_tokens=output_tokens,
+                    duration=duration,
+                    finish_reason=finish_reason,
+                    error=None,
+                )
+
+                # Log slow calls
+                if duration > 10.0 and input_tokens > 0:
+                    logger.info(
+                        f"slow llm call: scope={scope}, model={self.provider}/{self.model}, "
+                        f"input_tokens={input_tokens}, output_tokens={output_tokens}, "
+                        f"time={duration:.3f}s"
+                    )
+
+                if return_usage:
+                    token_usage = TokenUsage(
+                        input_tokens=input_tokens,
+                        output_tokens=output_tokens,
+                        total_tokens=input_tokens + output_tokens,
+                    )
+                    return result, token_usage
+                return result
+
+            except json.JSONDecodeError as e:
+                last_exception = e
+                if attempt < max_retries:
+                    logger.warning("Gemini returned invalid JSON, retrying...")
+                    backoff = min(initial_backoff * (2**attempt), max_backoff)
+                    await asyncio.sleep(backoff)
+                    continue
+                else:
+                    logger.error(f"Gemini returned invalid JSON after {max_retries + 1} attempts")
+                    raise
+
+            except genai_errors.APIError as e:
+                # Fast fail on auth errors - these won't recover with retries
+                if e.code in (401, 403):
+                    logger.error(f"Gemini auth error (HTTP {e.code}), not retrying: {str(e)}")
+                    raise
+
+                # Retry on retryable errors (rate limits, server errors, client errors)
+                if e.code in (400, 429, 500, 502, 503, 504) or (e.code and e.code >= 500):
+                    last_exception = e
+                    if attempt < max_retries:
+                        backoff = min(initial_backoff * (2**attempt), max_backoff)
+                        jitter = backoff * 0.2 * (2 * (time.time() % 1) - 1)
+                        await asyncio.sleep(backoff + jitter)
+                    else:
+                        logger.error(f"Gemini API error after {max_retries + 1} attempts: {str(e)}")
+                        raise
+                else:
+                    logger.error(f"Gemini API error: {type(e).__name__}: {str(e)}")
+                    raise
+
+            except Exception as e:
+                logger.error(f"Unexpected error during Gemini call: {type(e).__name__}: {str(e)}")
+                raise
+
+        if last_exception:
+            raise last_exception
+        raise RuntimeError("Gemini call failed after all retries")
+
+    async def call_with_tools(
+        self,
+        messages: list[dict[str, Any]],
+        tools: list[dict[str, Any]],
+        max_completion_tokens: int | None = None,
+        temperature: float | None = None,
+        scope: str = "tools",
+        max_retries: int = 5,
+        initial_backoff: float = 1.0,
+        max_backoff: float = 30.0,
+        tool_choice: str | dict[str, Any] = "auto",
+    ) -> LLMToolCallResult:
+        """
+        Make a Gemini/VertexAI API call with tool/function calling support.
+
+        Args:
+            messages: List of message dicts. Can include tool results with role='tool'.
+            tools: List of tool definitions in OpenAI format.
+            max_completion_tokens: Maximum tokens (not supported by Gemini).
+            temperature: Sampling temperature.
+            scope: Scope identifier for tracking.
+            max_retries: Maximum retry attempts.
+            initial_backoff: Initial backoff time in seconds.
+            max_backoff: Maximum backoff time in seconds.
+            tool_choice: How to choose tools (Gemini uses "auto" only).
+
+        Returns:
+            LLMToolCallResult with content and/or tool_calls.
+        """
+        start_time = time.time()
+
+        # Convert tools to Gemini format
+        gemini_tools = []
+        for tool in tools:
+            func = tool.get("function", {})
+            gemini_tools.append(
+                genai_types.Tool(
+                    function_declarations=[
+                        genai_types.FunctionDeclaration(
+                            name=func.get("name", ""),
+                            description=func.get("description", ""),
+                            parameters=func.get("parameters"),
+                        )
+                    ]
+                )
+            )
+
+        # Convert messages
+        system_instruction = None
+        gemini_contents = []
+        for msg in messages:
+            role = msg.get("role", "user")
+            content = msg.get("content", "")
+
+            if role == "system":
+                system_instruction = (system_instruction + "\n\n" + content) if system_instruction else content
+            elif role == "tool":
+                # Gemini uses function_response
+                gemini_contents.append(
+                    genai_types.Content(
+                        role="user",
+                        parts=[
+                            genai_types.Part(
+                                function_response=genai_types.FunctionResponse(
+                                    name=msg.get("name", ""),
+                                    response={"result": content},
+                                )
+                            )
+                        ],
+                    )
+                )
+            elif role == "assistant":
+                gemini_contents.append(genai_types.Content(role="model", parts=[genai_types.Part(text=content)]))
+            else:
+                gemini_contents.append(genai_types.Content(role="user", parts=[genai_types.Part(text=content)]))
+
+        config_kwargs: dict[str, Any] = {"tools": gemini_tools}
+        if system_instruction:
+            config_kwargs["system_instruction"] = system_instruction
+        if temperature is not None:
+            config_kwargs["temperature"] = temperature
+
+        config = genai_types.GenerateContentConfig(**config_kwargs)
+
+        last_exception = None
+        for attempt in range(max_retries + 1):
+            try:
+                response = await self._client.aio.models.generate_content(
+                    model=self.model,
+                    contents=gemini_contents,
+                    config=config,
+                )
+
+                # Extract content and tool calls
+                content = None
+                tool_calls: list[LLMToolCall] = []
+
+                if response.candidates and response.candidates[0].content:
+                    parts = response.candidates[0].content.parts
+                    if parts:
+                        for part in parts:
+                            if hasattr(part, "text") and part.text:
+                                content = part.text
+                            if hasattr(part, "function_call") and part.function_call:
+                                fc = part.function_call
+                                tool_calls.append(
+                                    LLMToolCall(
+                                        id=f"gemini_{len(tool_calls)}",
+                                        name=fc.name,
+                                        arguments=dict(fc.args) if fc.args else {},
+                                    )
+                                )
+
+                finish_reason = "tool_calls" if tool_calls else "stop"
+
+                # Extract token usage
+                input_tokens = 0
+                output_tokens = 0
+                if response.usage_metadata:
+                    input_tokens = response.usage_metadata.prompt_token_count or 0
+                    output_tokens = response.usage_metadata.candidates_token_count or 0
+
+                # Record metrics
+                duration = time.time() - start_time
+                metrics = get_metrics_collector()
+                metrics.record_llm_call(
+                    provider=self.provider,
+                    model=self.model,
+                    scope=scope,
+                    duration=duration,
+                    input_tokens=input_tokens,
+                    output_tokens=output_tokens,
+                    success=True,
+                )
+
+                # Record OpenTelemetry span
+                from hindsight_api.tracing import get_span_recorder
+
+                span_recorder = get_span_recorder()
+                # Convert LLMToolCall objects to dicts for span recording
+                tool_calls_dict = (
+                    [{"id": tc.id, "name": tc.name, "arguments": tc.arguments} for tc in tool_calls]
+                    if tool_calls
+                    else None
+                )
+                span_recorder.record_llm_call(
+                    provider=self.provider,
+                    model=self.model,
+                    scope=scope,
+                    messages=messages,
+                    response_content=content,
+                    input_tokens=input_tokens,
+                    output_tokens=output_tokens,
+                    duration=duration,
+                    finish_reason=finish_reason,
+                    error=None,
+                    tool_calls=tool_calls_dict,
+                )
+
+                return LLMToolCallResult(
+                    content=content,
+                    tool_calls=tool_calls,
+                    finish_reason=finish_reason,
+                    input_tokens=input_tokens,
+                    output_tokens=output_tokens,
+                )
+
+            except genai_errors.APIError as e:
+                # Fast fail on auth errors
+                if e.code in (401, 403):
+                    logger.error(f"Gemini auth error (HTTP {e.code}), not retrying: {str(e)}")
+                    raise
+
+                # Retry on retryable errors
+                last_exception = e
+                if attempt < max_retries:
+                    backoff = min(initial_backoff * (2**attempt), max_backoff)
+                    await asyncio.sleep(backoff)
+                    continue
+                raise
+
+            except Exception as e:
+                logger.error(f"Unexpected error during Gemini tool call: {type(e).__name__}: {str(e)}")
+                raise
+
+        if last_exception:
+            raise last_exception
+        raise RuntimeError("Gemini tool call failed")
+
+    async def cleanup(self) -> None:
+        """Clean up resources (close connections, etc.)."""
+        # Gemini client doesn't require explicit cleanup
+        pass
@@ -0,0 +1,301 @@
+"""
+Mock LLM provider for testing.
+
+This provider allows tests to record LLM calls and return configurable mock responses
+without making actual API calls to external LLM services.
+"""
+
+import logging
+from typing import Any
+
+from ..llm_interface import LLMInterface
+from ..response_models import LLMToolCall, LLMToolCallResult, TokenUsage
+
+logger = logging.getLogger(__name__)
+
+
+class MockLLM(LLMInterface):
+    """
+    Mock LLM provider for testing.
+
+    This provider records all calls and returns configurable mock responses,
+    enabling tests to verify LLM interactions without making real API calls.
+
+    Example:
+        # Create mock provider
+        mock_llm = MockLLM(provider="mock", api_key="", base_url="", model="mock-model")
+
+        # Set mock response
+        mock_llm.set_mock_response({"answer": "test"})
+
+        # Make calls
+        result = await mock_llm.call(
+            messages=[{"role": "user", "content": "test"}],
+            response_format=MyResponseModel
+        )
+
+        # Verify calls
+        calls = mock_llm.get_mock_calls()
+        assert len(calls) == 1
+        assert calls[0]["scope"] == "memory"
+    """
+
+    def __init__(
+        self,
+        provider: str,
+        api_key: str,
+        base_url: str,
+        model: str,
+        reasoning_effort: str = "low",
+        **kwargs: Any,
+    ):
+        """
+        Initialize mock LLM provider.
+
+        Args:
+            provider: Provider name (should be "mock").
+            api_key: Not used for mock provider.
+            base_url: Not used for mock provider.
+            model: Model name for tracking.
+            reasoning_effort: Not used for mock provider.
+            **kwargs: Additional parameters (not used).
+        """
+        super().__init__(provider, api_key, base_url, model, reasoning_effort, **kwargs)
+
+        # Storage for test verification
+        self._mock_calls: list[dict] = []
+        self._mock_response: Any = None
+        self._mock_exception: Exception | None = None
+
+    async def verify_connection(self) -> None:
+        """
+        Verify mock provider (always succeeds).
+
+        Mock provider doesn't need connection verification since it doesn't
+        make real API calls.
+        """
+        logger.debug("Mock LLM: connection verification (always succeeds)")
+
+    async def call(
+        self,
+        messages: list[dict[str, str]],
+        response_format: Any | None = None,
+        max_completion_tokens: int | None = None,
+        temperature: float | None = None,
+        scope: str = "memory",
+        max_retries: int = 10,
+        initial_backoff: float = 1.0,
+        max_backoff: float = 60.0,
+        skip_validation: bool = False,
+        strict_schema: bool = False,
+        return_usage: bool = False,
+    ) -> Any:
+        """
+        Make a mock LLM API call.
+
+        Records the call for test verification and returns the configured mock response.
+
+        Args:
+            messages: List of message dicts with 'role' and 'content'.
+            response_format: Optional Pydantic model for structured output.
+            max_completion_tokens: Not used in mock.
+            temperature: Not used in mock.
+            scope: Scope identifier for tracking.
+            max_retries: Not used in mock.
+            initial_backoff: Not used in mock.
+            max_backoff: Not used in mock.
+            skip_validation: Return raw JSON without Pydantic validation.
+            strict_schema: Not used in mock.
+            return_usage: If True, return tuple (result, TokenUsage) instead of just result.
+
+        Returns:
+            If return_usage=False: Parsed response if response_format is provided, otherwise text content.
+            If return_usage=True: Tuple of (result, TokenUsage) with mock token counts.
+        """
+        # Record the call for test verification
+        call_record = {
+            "provider": self.provider,
+            "model": self.model,
+            "messages": messages,
+            "response_format": response_format.__name__
+            if response_format and hasattr(response_format, "__name__")
+            else str(response_format),
+            "scope": scope,
+        }
+        self._mock_calls.append(call_record)
+        logger.debug(f"Mock LLM call recorded: scope={scope}, model={self.model}")
+
+        # Raise mock exception if configured
+        if self._mock_exception is not None:
+            raise self._mock_exception
+
+        # Record trace span (minimal for mock provider)
+        from hindsight_api.tracing import get_span_recorder
+
+        span_recorder = get_span_recorder()
+        span_recorder.record_llm_call(
+            provider=self.provider,
+            model=self.model,
+            scope=scope,
+            messages=messages,
+            response_content="mock response",
+            input_tokens=10,
+            output_tokens=5,
+            duration=0.001,  # Mock calls are instant
+            finish_reason="stop",
+            error=None,
+        )
+
+        # Return mock response
+        if self._mock_response is not None:
+            result = self._mock_response
+        elif response_format is not None:
+            # Try to create a minimal valid instance of the response format
+            try:
+                # For Pydantic models, try to create with minimal valid data
+                result = {"mock": True}
+            except Exception:
+                result = {"mock": True}
+        else:
+            result = "mock response"
+
+        if return_usage:
+            token_usage = TokenUsage(input_tokens=10, output_tokens=5, total_tokens=15)
+            return result, token_usage
+        return result
+
+    async def call_with_tools(
+        self,
+        messages: list[dict[str, Any]],
+        tools: list[dict[str, Any]],
+        max_completion_tokens: int | None = None,
+        temperature: float | None = None,
+        scope: str = "tools",
+        max_retries: int = 5,
+        initial_backoff: float = 1.0,
+        max_backoff: float = 30.0,
+        tool_choice: str | dict[str, Any] = "auto",
+    ) -> LLMToolCallResult:
+        """
+        Make a mock LLM API call with tool/function calling support.
+
+        Records the call for test verification and returns the configured mock response.
+
+        Args:
+            messages: List of message dicts. Can include tool results with role='tool'.
+            tools: List of tool definitions in OpenAI format.
+            max_completion_tokens: Not used in mock.
+            temperature: Not used in mock.
+            scope: Scope identifier for tracking.
+            max_retries: Not used in mock.
+            initial_backoff: Not used in mock.
+            max_backoff: Not used in mock.
+            tool_choice: Not used in mock.
+
+        Returns:
+            LLMToolCallResult with content and/or tool_calls.
+        """
+        # Record the call for test verification
+        call_record = {
+            "provider": self.provider,
+            "model": self.model,
+            "messages": messages,
+            "tools": [t.get("function", {}).get("name") for t in tools],
+            "scope": scope,
+        }
+        self._mock_calls.append(call_record)
+
+        # Raise mock exception if configured
+        if self._mock_exception is not None:
+            raise self._mock_exception
+
+        # Record OpenTelemetry span
+        from hindsight_api.tracing import get_span_recorder
+
+        span_recorder = get_span_recorder()
+
+        if self._mock_response is not None:
+            if isinstance(self._mock_response, LLMToolCallResult):
+                result = self._mock_response
+            elif isinstance(self._mock_response, list):
+                # Allow setting just tool calls as a list
+                result = LLMToolCallResult(
+                    tool_calls=[
+                        LLMToolCall(id=f"mock_{i}", name=tc["name"], arguments=tc.get("arguments", {}))
+                        for i, tc in enumerate(self._mock_response)
+                    ],
+                    finish_reason="tool_calls",
+                )
+            else:
+                result = LLMToolCallResult(content="mock response", finish_reason="stop")
+        else:
+            result = LLMToolCallResult(content="mock response", finish_reason="stop")
+
+        # Record span with mock values
+        # Convert LLMToolCall objects to dicts for span recording
+        tool_calls_dict = (
+            [{"id": tc.id, "name": tc.name, "arguments": tc.arguments} for tc in result.tool_calls]
+            if result.tool_calls
+            else None
+        )
+        span_recorder.record_llm_call(
+            provider=self.provider,
+            model=self.model,
+            scope=scope,
+            messages=messages,
+            response_content=result.content,
+            input_tokens=10,  # Mock value
+            output_tokens=5,  # Mock value
+            duration=0.1,  # Mock value
+            finish_reason=result.finish_reason,
+            error=None,
+            tool_calls=tool_calls_dict,
+        )
+
+        return result
+
+    async def cleanup(self) -> None:
+        """Clean up resources (no-op for mock provider)."""
+        pass
+
+    def set_mock_response(self, response: Any) -> None:
+        """
+        Set the response to return from mock calls.
+
+        Args:
+            response: The response to return. Can be:
+                - A dict/Pydantic model for regular calls
+                - An LLMToolCallResult for tool calls
+                - A list of tool call dicts for tool calls
+                - Any other value to return as-is
+        """
+        self._mock_response = response
+
+    def set_mock_exception(self, exception: Exception) -> None:
+        """
+        Set an exception to raise from mock calls.
+
+        Args:
+            exception: The exception to raise on the next call.
+                      After raising, the exception is cleared.
+        """
+        self._mock_exception = exception
+
+    def get_mock_calls(self) -> list[dict]:
+        """
+        Get the list of recorded mock calls.
+
+        Returns:
+            List of call records, each containing:
+                - provider: Provider name
+                - model: Model name
+                - messages: Messages sent
+                - response_format/tools: Format or tools used
+                - scope: Call scope
+        """
+        return self._mock_calls
+
+    def clear_mock_calls(self) -> None:
+        """Clear the recorded mock calls and any set exception."""
+        self._mock_calls = []
+        self._mock_exception = None
@@ -0,0 +1,788 @@
+"""
+OpenAI-compatible LLM provider supporting OpenAI, Groq, Ollama, and LMStudio.
+
+This provider handles all OpenAI API-compatible models including:
+- OpenAI: GPT-4, GPT-4o, GPT-5, o1, o3 (reasoning models)
+- Groq: Fast inference with seed control and service tiers
+- Ollama: Local models with native streaming API support
+- LMStudio: Local models with OpenAI-compatible API
+
+Features:
+- Reasoning models with extended thinking (o1, o3, GPT-5 families)
+- Strict JSON schema enforcement (OpenAI)
+- Provider-specific parameters (Groq seed, service tier)
+- Native Ollama streaming for better structured output
+- Automatic token limit handling per model family
+"""
+
+import asyncio
+import json
+import logging
+import os
+import re
+import time
+from typing import Any
+
+import httpx
+from openai import APIConnectionError, APIStatusError, AsyncOpenAI, LengthFinishReasonError
+
+from hindsight_api.config import DEFAULT_LLM_TIMEOUT, ENV_LLM_TIMEOUT
+from hindsight_api.engine.llm_interface import LLMInterface, OutputTooLongError
+from hindsight_api.engine.response_models import LLMToolCall, LLMToolCallResult, TokenUsage
+from hindsight_api.metrics import get_metrics_collector
+
+logger = logging.getLogger(__name__)
+
+# Seed applied to every Groq request for deterministic behavior
+DEFAULT_LLM_SEED = 4242
+
+
+class OpenAICompatibleLLM(LLMInterface):
+    """
+    LLM provider for OpenAI-compatible APIs.
+
+    Supports:
+    - OpenAI: Standard models (GPT-4, GPT-4o) and reasoning models (o1, o3, GPT-5)
+    - Groq: Fast inference with seed control and service tiers
+    - Ollama: Local models with native streaming API for better structured output
+    - LMStudio: Local models with OpenAI-compatible API
+    """
+
+    def __init__(
+        self,
+        provider: str,
+        api_key: str,
+        base_url: str,
+        model: str,
+        reasoning_effort: str = "low",
+        timeout: float | None = None,
+        groq_service_tier: str | None = None,
+        **kwargs: Any,
+    ):
+        """
+        Initialize OpenAI-compatible LLM provider.
+
+        Args:
+            provider: Provider name ("openai", "groq", "ollama", "lmstudio").
+            api_key: API key (optional for ollama/lmstudio).
+            base_url: Base URL for the API (uses defaults for groq/ollama/lmstudio if empty).
+            model: Model name.
+            reasoning_effort: Reasoning effort level for supported models ("low", "medium", "high").
+            timeout: Request timeout in seconds (uses env var or 300s default).
+            groq_service_tier: Groq service tier ("on_demand", "flex", "auto").
+            **kwargs: Additional provider-specific parameters.
+        """
+        super().__init__(provider, api_key, base_url, model, reasoning_effort, **kwargs)
+
+        # Validate provider
+        valid_providers = ["openai", "groq", "ollama", "lmstudio"]
+        if self.provider not in valid_providers:
+            raise ValueError(f"OpenAICompatibleLLM only supports: {', '.join(valid_providers)}. Got: {self.provider}")
+
+        # Set default base URLs
+        if not self.base_url:
+            if self.provider == "groq":
+                self.base_url = "https://api.groq.com/openai/v1"
+            elif self.provider == "ollama":
+                self.base_url = "http://localhost:11434/v1"
+            elif self.provider == "lmstudio":
+                self.base_url = "http://localhost:1234/v1"
+
+        # For ollama/lmstudio, use dummy key if not provided
+        if self.provider in ("ollama", "lmstudio") and not self.api_key:
+            self.api_key = "local"
+
+        # Validate API key for cloud providers
+        if self.provider in ("openai", "groq") and not self.api_key:
+            raise ValueError(f"API key is required for {self.provider}")
+
+        # Groq service tier configuration
+        self.groq_service_tier = groq_service_tier or os.getenv("HINDSIGHT_API_LLM_GROQ_SERVICE_TIER", "auto")
+
+        # Get timeout config
+        self.timeout = timeout or float(os.getenv(ENV_LLM_TIMEOUT, str(DEFAULT_LLM_TIMEOUT)))
+
+        # Create OpenAI client
+        client_kwargs: dict[str, Any] = {"api_key": self.api_key, "max_retries": 0}
+        if self.base_url:
+            client_kwargs["base_url"] = self.base_url
+        if self.timeout:
+            client_kwargs["timeout"] = self.timeout
+
+        self._client = AsyncOpenAI(**client_kwargs)
+        logger.info(
+            f"OpenAI-compatible client initialized: provider={self.provider}, model={self.model}, "
+            f"base_url={self.base_url or 'default'}"
+        )
+
+    async def verify_connection(self) -> None:
+        """
+        Verify that the provider is configured correctly by making a simple test call.
+
+        Raises:
+            RuntimeError: If the connection test fails.
+        """
+        try:
+            logger.info(f"Verifying connection: {self.provider}/{self.model}")
+            await self.call(
+                messages=[{"role": "user", "content": "Say 'ok'"}],
+                max_completion_tokens=100,
+                max_retries=2,
+                initial_backoff=0.5,
+                max_backoff=2.0,
+                scope="verification",
+            )
+            logger.info(f"Connection verified: {self.provider}/{self.model}")
+        except Exception as e:
+            raise RuntimeError(f"Connection verification failed for {self.provider}/{self.model}: {e}") from e
+
+    def _supports_reasoning_model(self) -> bool:
+        """Check if the current model is a reasoning model (o1, o3, GPT-5, DeepSeek)."""
+        model_lower = self.model.lower()
+        return any(x in model_lower for x in ["gpt-5", "o1", "o3", "deepseek"])
+
+    def _get_max_reasoning_tokens(self) -> int | None:
+        """Get max reasoning tokens for reasoning models."""
+        model_lower = self.model.lower()
+
+        # GPT-4 and GPT-4.1 models have different caps
+        if any(x in model_lower for x in ["gpt-4.1", "gpt-4-"]):
+            return 32000
+        elif "gpt-4o" in model_lower:
+            return 16384
+
+        return None
+
+    async def call(
+        self,
+        messages: list[dict[str, str]],
+        response_format: Any | None = None,
+        max_completion_tokens: int | None = None,
+        temperature: float | None = None,
+        scope: str = "memory",
+        max_retries: int = 10,
+        initial_backoff: float = 1.0,
+        max_backoff: float = 60.0,
+        skip_validation: bool = False,
+        strict_schema: bool = False,
+        return_usage: bool = False,
+    ) -> Any:
+        """
+        Make an LLM API call with retry logic.
+
+        Args:
+            messages: List of message dicts with 'role' and 'content'.
+            response_format: Optional Pydantic model for structured output.
+            max_completion_tokens: Maximum tokens in response.
+            temperature: Sampling temperature (0.0-2.0).
+            scope: Scope identifier for tracking.
+            max_retries: Maximum retry attempts.
+            initial_backoff: Initial backoff time in seconds.
+            max_backoff: Maximum backoff time in seconds.
+            skip_validation: Return raw JSON without Pydantic validation.
+            strict_schema: Use strict JSON schema enforcement (OpenAI only).
+            return_usage: If True, return tuple (result, TokenUsage) instead of just result.
+
+        Returns:
+            If return_usage=False: Parsed response if response_format is provided, otherwise text content.
+            If return_usage=True: Tuple of (result, TokenUsage) with token counts.
+
+        Raises:
+            OutputTooLongError: If output exceeds token limits.
+            Exception: Re-raises API errors after retries exhausted.
+        """
+        # Handle Ollama with native API for structured output (better schema enforcement)
+        if self.provider == "ollama" and response_format is not None:
+            return await self._call_ollama_native(
+                messages=messages,
+                response_format=response_format,
+                max_completion_tokens=max_completion_tokens,
+                temperature=temperature,
+                max_retries=max_retries,
+                initial_backoff=initial_backoff,
+                max_backoff=max_backoff,
+                skip_validation=skip_validation,
+                scope=scope,
+                return_usage=return_usage,
+            )
+
+        start_time = time.time()
+
+        # Build call parameters
+        call_params: dict[str, Any] = {
+            "model": self.model,
+            "messages": messages,
+        }
+
+        # Check if model supports reasoning parameter
+        is_reasoning_model = self._supports_reasoning_model()
+
+        # Apply model-specific token limits
+        if max_completion_tokens is not None:
+            max_tokens_cap = self._get_max_reasoning_tokens()
+            if max_tokens_cap and max_completion_tokens > max_tokens_cap:
+                max_completion_tokens = max_tokens_cap
+            # For reasoning models, enforce minimum to ensure space for reasoning + output
+            if is_reasoning_model and max_completion_tokens < 16000:
+                max_completion_tokens = 16000
+            call_params["max_completion_tokens"] = max_completion_tokens
+
+        # Temperature - reasoning models don't support custom temperature
+        if temperature is not None and not is_reasoning_model:
+            call_params["temperature"] = temperature
+
+        # Set reasoning_effort for reasoning models
+        if is_reasoning_model:
+            call_params["reasoning_effort"] = self.reasoning_effort
+
+        # Provider-specific parameters
+        if self.provider == "groq":
+            call_params["seed"] = DEFAULT_LLM_SEED
+            extra_body: dict[str, Any] = {}
+            # Add service_tier if configured
+            if self.groq_service_tier:
+                extra_body["service_tier"] = self.groq_service_tier
+            # Add reasoning parameters for reasoning models
+            if is_reasoning_model:
+                extra_body["include_reasoning"] = False
+            if extra_body:
+                call_params["extra_body"] = extra_body
+
+        # Prepare response format ONCE before retry loop
+        if response_format is not None:
+            schema = None
+            if hasattr(response_format, "model_json_schema"):
+                schema = response_format.model_json_schema()
+
+            if strict_schema and schema is not None:
+                # Use OpenAI's strict JSON schema enforcement
+                call_params["response_format"] = {
+                    "type": "json_schema",
+                    "json_schema": {
+                        "name": "response",
+                        "strict": True,
+                        "schema": schema,
+                    },
+                }
+            else:
+                # Soft enforcement: add schema to prompt and use json_object mode
+                if schema is not None:
+                    schema_msg = (
+                        f"\n\nYou must respond with valid JSON matching this schema:\n{json.dumps(schema, indent=2)}"
+                    )
+
+                    if call_params["messages"] and call_params["messages"][0].get("role") == "system":
+                        first_msg = call_params["messages"][0]
+                        if isinstance(first_msg, dict) and isinstance(first_msg.get("content"), str):
+                            first_msg["content"] += schema_msg
+                    elif call_params["messages"]:
+                        first_msg = call_params["messages"][0]
+                        if isinstance(first_msg, dict) and isinstance(first_msg.get("content"), str):
+                            first_msg["content"] = schema_msg + "\n\n" + first_msg["content"]
+                if self.provider not in ("lmstudio", "ollama"):
+                    # LM Studio and Ollama don't support json_object response format reliably
+                    call_params["response_format"] = {"type": "json_object"}
+
+        last_exception = None
+
+        for attempt in range(max_retries + 1):
+            try:
+                if response_format is not None:
+                    response = await self._client.chat.completions.create(**call_params)
+
+                    content = response.choices[0].message.content
+
+                    # Strip reasoning model thinking tags
+                    # Supports: <think>, <thinking>, <reasoning>, |startthink|/|endthink|
+                    if content:
+                        original_len = len(content)
+                        content = re.sub(r"<think>.*?</think>", "", content, flags=re.DOTALL)
+                        content = re.sub(r"<thinking>.*?</thinking>", "", content, flags=re.DOTALL)
+                        content = re.sub(r"<reasoning>.*?</reasoning>", "", content, flags=re.DOTALL)
+                        content = re.sub(r"\|startthink\|.*?\|endthink\|", "", content, flags=re.DOTALL)
+                        content = content.strip()
+                        if len(content) < original_len:
+                            logger.debug(f"Stripped {original_len - len(content)} chars of reasoning tokens")
+
+                    # For local models, they may wrap JSON in markdown code blocks
+                    if self.provider in ("lmstudio", "ollama"):
+                        clean_content = content
+                        if "```json" in content:
+                            clean_content = content.split("```json")[1].split("```")[0].strip()
+                        elif "```" in content:
+                            clean_content = content.split("```")[1].split("```")[0].strip()
+                        try:
+                            json_data = json.loads(clean_content)
+                        except json.JSONDecodeError:
+                            # Fallback to parsing raw content
+                            json_data = json.loads(content)
+                    else:
+                        # Log raw LLM response for debugging JSON parse issues
+                        try:
+                            json_data = json.loads(content)
+                        except json.JSONDecodeError as json_err:
+                            # Truncate content for logging
+                            content_preview = content[:500] if content else "<empty>"
+                            if content and len(content) > 700:
+                                content_preview = f"{content[:500]}...TRUNCATED...{content[-200:]}"
+                            logger.warning(
+                                f"JSON parse error from LLM response (attempt {attempt + 1}/{max_retries + 1}): {json_err}\n"
+                                f"  Model: {self.provider}/{self.model}\n"
+                                f"  Content length: {len(content) if content else 0} chars\n"
+                                f"  Content preview: {content_preview!r}\n"
+                                f"  Finish reason: {response.choices[0].finish_reason if response.choices else 'unknown'}"
+                            )
+                            # Retry on JSON parse errors
+                            if attempt < max_retries:
+                                backoff = min(initial_backoff * (2**attempt), max_backoff)
+                                await asyncio.sleep(backoff)
+                                last_exception = json_err
+                                continue
+                            else:
+                                logger.error(f"JSON parse error after {max_retries + 1} attempts, giving up")
+                                raise
+
+                    if skip_validation:
+                        result = json_data
+                    else:
+                        result = response_format.model_validate(json_data)
+                else:
+                    response = await self._client.chat.completions.create(**call_params)
+                    result = response.choices[0].message.content
+
+                # Record token usage metrics
+                duration = time.time() - start_time
+                usage = response.usage
+                input_tokens = usage.prompt_tokens or 0 if usage else 0
+                output_tokens = usage.completion_tokens or 0 if usage else 0
+                total_tokens = usage.total_tokens or 0 if usage else 0
+
+                # Record LLM metrics
+                metrics = get_metrics_collector()
+                metrics.record_llm_call(
+                    provider=self.provider,
+                    model=self.model,
+                    scope=scope,
+                    duration=duration,
+                    input_tokens=input_tokens,
+                    output_tokens=output_tokens,
+                    success=True,
+                )
+
+                # Record trace span
+                from hindsight_api.tracing import _serialize_for_span, get_span_recorder
+
+                finish_reason = response.choices[0].finish_reason if response.choices else None
+                span_recorder = get_span_recorder()
+                span_recorder.record_llm_call(
+                    provider=self.provider,
+                    model=self.model,
+                    scope=scope,
+                    messages=messages,
+                    response_content=_serialize_for_span(result),
+                    input_tokens=input_tokens,
+                    output_tokens=output_tokens,
+                    duration=duration,
+                    finish_reason=finish_reason,
+                    error=None,
+                )
+
+                # Log slow calls
+                if duration > 10.0 and usage:
+                    ratio = max(1, output_tokens) / max(1, input_tokens)
+                    cached_tokens = 0
+                    if hasattr(usage, "prompt_tokens_details") and usage.prompt_tokens_details:
+                        cached_tokens = getattr(usage.prompt_tokens_details, "cached_tokens", 0) or 0
+                    cache_info = f", cached_tokens={cached_tokens}" if cached_tokens > 0 else ""
+                    logger.info(
+                        f"slow llm call: scope={scope}, model={self.provider}/{self.model}, "
+                        f"input_tokens={input_tokens}, output_tokens={output_tokens}, "
+                        f"total_tokens={total_tokens}{cache_info}, time={duration:.3f}s, ratio out/in={ratio:.2f}"
+                    )
+
+                if return_usage:
+                    token_usage = TokenUsage(
+                        input_tokens=input_tokens,
+                        output_tokens=output_tokens,
+                        total_tokens=total_tokens,
+                    )
+                    return result, token_usage
+                return result
+
+            except LengthFinishReasonError as e:
+                logger.warning(f"LLM output exceeded token limits: {str(e)}")
+                raise OutputTooLongError(
+                    "LLM output exceeded token limits. Input may need to be split into smaller chunks."
+                ) from e
+
+            except APIConnectionError as e:
+                last_exception = e
+                status_code = getattr(e, "status_code", None) or getattr(
+                    getattr(e, "response", None), "status_code", None
+                )
+                logger.warning(f"APIConnectionError (HTTP {status_code}), attempt {attempt + 1}: {str(e)[:200]}")
+                if attempt < max_retries:
+                    backoff = min(initial_backoff * (2**attempt), max_backoff)
+                    await asyncio.sleep(backoff)
+                    continue
+                else:
+                    logger.error(f"Connection error after {max_retries + 1} attempts: {str(e)}")
+                    raise
+
+            except APIStatusError as e:
+                # Fast fail only on 401 (unauthorized) and 403 (forbidden)
+                if e.status_code in (401, 403):
+                    logger.error(f"Auth error (HTTP {e.status_code}), not retrying: {str(e)}")
+                    raise
+
+                # Handle tool_use_failed error - model outputted in tool call format
+                if e.status_code == 400 and response_format is not None:
+                    try:
+                        error_body = e.body if hasattr(e, "body") else {}
+                        if isinstance(error_body, dict):
+                            error_info: dict[str, Any] = error_body.get("error") or {}
+                            if error_info.get("code") == "tool_use_failed":
+                                failed_gen = error_info.get("failed_generation", "")
+                                if failed_gen:
+                                    # Parse tool call format and convert to expected format
+                                    tool_call = json.loads(failed_gen)
+                                    tool_name = tool_call.get("name", "")
+                                    tool_args = tool_call.get("arguments", {})
+                                    converted = {"actions": [{"tool": tool_name, **tool_args}]}
+                                    if skip_validation:
+                                        result = converted
+                                    else:
+                                        result = response_format.model_validate(converted)
+
+                                    # Record metrics
+                                    duration = time.time() - start_time
+                                    metrics = get_metrics_collector()
+                                    metrics.record_llm_call(
+                                        provider=self.provider,
+                                        model=self.model,
+                                        scope=scope,
+                                        duration=duration,
+                                        input_tokens=0,
+                                        output_tokens=0,
+                                        success=True,
+                                    )
+                                    if return_usage:
+                                        return result, TokenUsage(input_tokens=0, output_tokens=0, total_tokens=0)
+                                    return result
+                    except (json.JSONDecodeError, KeyError, TypeError):
+                        pass  # Failed to parse tool_use_failed, continue with normal retry
+
+                last_exception = e
+                if attempt < max_retries:
+                    backoff = min(initial_backoff * (2**attempt), max_backoff)
+                    jitter = backoff * 0.2 * (2 * (time.time() % 1) - 1)
+                    sleep_time = backoff + jitter
+                    await asyncio.sleep(sleep_time)
+                else:
+                    logger.error(f"API error after {max_retries + 1} attempts: {str(e)}")
+                    raise
+
+            except Exception:
+                raise
+
+        if last_exception:
+            raise last_exception
+        raise RuntimeError("LLM call failed after all retries with no exception captured")
+
+    async def call_with_tools(
+        self,
+        messages: list[dict[str, Any]],
+        tools: list[dict[str, Any]],
+        max_completion_tokens: int | None = None,
+        temperature: float | None = None,
+        scope: str = "tools",
+        max_retries: int = 5,
+        initial_backoff: float = 1.0,
+        max_backoff: float = 30.0,
+        tool_choice: str | dict[str, Any] = "auto",
+    ) -> LLMToolCallResult:
+        """
+        Make an LLM API call with tool/function calling support.
+
+        Args:
+            messages: List of message dicts. Can include tool results with role='tool'.
+            tools: List of tool definitions in OpenAI format.
+            max_completion_tokens: Maximum tokens in response.
+            temperature: Sampling temperature (0.0-2.0).
+            scope: Scope identifier for tracking.
+            max_retries: Maximum retry attempts.
+            initial_backoff: Initial backoff time in seconds.
+            max_backoff: Maximum backoff time in seconds.
+            tool_choice: How to choose tools - "auto", "none", "required", or specific function.
+
+        Returns:
+            LLMToolCallResult with content and/or tool_calls.
+        """
+        start_time = time.time()
+
+        # Build call parameters
+        call_params: dict[str, Any] = {
+            "model": self.model,
+            "messages": messages,
+            "tools": tools,
+            "tool_choice": tool_choice,
+        }
+
+        if max_completion_tokens is not None:
+            call_params["max_completion_tokens"] = max_completion_tokens
+        if temperature is not None:
+            call_params["temperature"] = temperature
+
+        # Provider-specific parameters
+        if self.provider == "groq":
+            call_params["seed"] = DEFAULT_LLM_SEED
+
+        last_exception = None
+
+        for attempt in range(max_retries + 1):
+            try:
+                response = await self._client.chat.completions.create(**call_params)
+
+                message = response.choices[0].message
+                finish_reason = response.choices[0].finish_reason
+
+                # Extract tool calls if present
+                tool_calls: list[LLMToolCall] = []
+                if message.tool_calls:
+                    for tc in message.tool_calls:
+                        try:
+                            args = json.loads(tc.function.arguments) if tc.function.arguments else {}
+                        except json.JSONDecodeError:
+                            args = {"_raw": tc.function.arguments}
+                        tool_calls.append(LLMToolCall(id=tc.id, name=tc.function.name, arguments=args))
+
+                content = message.content
+
+                # Record metrics
+                duration = time.time() - start_time
+                usage = response.usage
+                input_tokens = usage.prompt_tokens or 0 if usage else 0
+                output_tokens = usage.completion_tokens or 0 if usage else 0
+
+                metrics = get_metrics_collector()
+                metrics.record_llm_call(
+                    provider=self.provider,
+                    model=self.model,
+                    scope=scope,
+                    duration=duration,
+                    input_tokens=input_tokens,
+                    output_tokens=output_tokens,
+                    success=True,
+                )
+
+                # Record OpenTelemetry span
+                from hindsight_api.tracing import get_span_recorder
+
+                span_recorder = get_span_recorder()
+                # Convert LLMToolCall objects to dicts for span recording
+                tool_calls_dict = (
+                    [{"id": tc.id, "name": tc.name, "arguments": tc.arguments} for tc in tool_calls]
+                    if tool_calls
+                    else None
+                )
+                span_recorder.record_llm_call(
+                    provider=self.provider,
+                    model=self.model,
+                    scope=scope,
+                    messages=messages,
+                    response_content=content,
+                    input_tokens=input_tokens,
+                    output_tokens=output_tokens,
+                    duration=duration,
+                    finish_reason=finish_reason,
+                    error=None,
+                    tool_calls=tool_calls_dict,
+                )
+
+                return LLMToolCallResult(
+                    content=content,
+                    tool_calls=tool_calls,
+                    finish_reason=finish_reason,
+                    input_tokens=input_tokens,
+                    output_tokens=output_tokens,
+                )
+
+            except APIConnectionError as e:
+                last_exception = e
+                if attempt < max_retries:
+                    await asyncio.sleep(min(initial_backoff * (2**attempt), max_backoff))
+                    continue
+                raise
+
+            except APIStatusError as e:
+                if e.status_code in (401, 403):
+                    raise
+                last_exception = e
+                if attempt < max_retries:
+                    await asyncio.sleep(min(initial_backoff * (2**attempt), max_backoff))
+                    continue
+                raise
+
+            except Exception:
+                raise
+
+        if last_exception:
+            raise last_exception
+        raise RuntimeError("Tool call failed after all retries")
+
+    async def _call_ollama_native(
+        self,
+        messages: list[dict[str, str]],
+        response_format: Any,
+        max_completion_tokens: int | None,
+        temperature: float | None,
+        max_retries: int,
+        initial_backoff: float,
+        max_backoff: float,
+        skip_validation: bool,
+        scope: str = "memory",
+        return_usage: bool = False,
+    ) -> Any:
+        """
+        Call Ollama using native API with JSON schema enforcement.
+
+        Ollama's native API supports passing a full JSON schema in the 'format' parameter,
+        which provides better structured output control than the OpenAI-compatible API.
+        """
+        start_time = time.time()
+
+        # Get the JSON schema from the Pydantic model
+        schema = response_format.model_json_schema() if hasattr(response_format, "model_json_schema") else None
+
+        # Build the base URL for Ollama's native API
+        # Default OpenAI-compatible URL is http://localhost:11434/v1
+        # Native API is at http://localhost:11434/api/chat
+        base_url = self.base_url or "http://localhost:11434/v1"
+        if base_url.endswith("/v1"):
+            native_url = base_url[:-3] + "/api/chat"
+        else:
+            native_url = base_url.rstrip("/") + "/api/chat"
+
+        # Build request payload
+        payload: dict[str, Any] = {
+            "model": self.model,
+            "messages": messages,
+            "stream": False,
+        }
+
+        # Add schema as format parameter for structured output
+        if schema:
+            payload["format"] = schema
+
+        # Add optional parameters with optimized defaults for Ollama
+        options: dict[str, Any] = {
+            "num_ctx": 16384,  # 16k context window for larger prompts
+            "num_batch": 512,  # Optimal batch size for prompt processing
+        }
+        if max_completion_tokens:
+            options["num_predict"] = max_completion_tokens
+        if temperature is not None:
+            options["temperature"] = temperature
+        payload["options"] = options
+
+        last_exception = None
+
+        async with httpx.AsyncClient(timeout=300.0) as client:
+            for attempt in range(max_retries + 1):
+                try:
+                    response = await client.post(native_url, json=payload)
+                    response.raise_for_status()
+
+                    result = response.json()
+                    content = result.get("message", {}).get("content", "")
+
+                    # Parse JSON response
+                    try:
+                        json_data = json.loads(content)
+                    except json.JSONDecodeError as json_err:
+                        content_preview = content[:500] if content else "<empty>"
+                        if content and len(content) > 700:
+                            content_preview = f"{content[:500]}...TRUNCATED...{content[-200:]}"
+                        logger.warning(
+                            f"Ollama JSON parse error (attempt {attempt + 1}/{max_retries + 1}): {json_err}\n"
+                            f"  Model: ollama/{self.model}\n"
+                            f"  Content length: {len(content) if content else 0} chars\n"
+                            f"  Content preview: {content_preview!r}"
+                        )
+                        if attempt < max_retries:
+                            backoff = min(initial_backoff * (2**attempt), max_backoff)
+                            await asyncio.sleep(backoff)
+                            last_exception = json_err
+                            continue
+                        else:
+                            raise
+
+                    # Extract token usage from Ollama response
+                    duration = time.time() - start_time
+                    input_tokens = result.get("prompt_eval_count", 0) or 0
+                    output_tokens = result.get("eval_count", 0) or 0
+                    total_tokens = input_tokens + output_tokens
+
+                    # Record LLM metrics
+                    metrics = get_metrics_collector()
+                    metrics.record_llm_call(
+                        provider=self.provider,
+                        model=self.model,
+                        scope=scope,
+                        duration=duration,
+                        input_tokens=input_tokens,
+                        output_tokens=output_tokens,
+                        success=True,
+                    )
+
+                    # Validate against Pydantic model or return raw JSON
+                    if skip_validation:
+                        validated_result = json_data
+                    else:
+                        validated_result = response_format.model_validate(json_data)
+
+                    if return_usage:
+                        token_usage = TokenUsage(
+                            input_tokens=input_tokens,
+                            output_tokens=output_tokens,
+                            total_tokens=total_tokens,
+                        )
+                        return validated_result, token_usage
+                    return validated_result
+
+                except httpx.HTTPStatusError as e:
+                    last_exception = e
+                    if attempt < max_retries:
+                        logger.warning(
+                            f"Ollama HTTP error (attempt {attempt + 1}/{max_retries + 1}): {e.response.status_code}"
+                        )
+                        backoff = min(initial_backoff * (2**attempt), max_backoff)
+                        await asyncio.sleep(backoff)
+                        continue
+                    else:
+                        logger.error(f"Ollama HTTP error after {max_retries + 1} attempts: {e}")
+                        raise
+
+                except httpx.RequestError as e:
+                    last_exception = e
+                    if attempt < max_retries:
+                        logger.warning(f"Ollama connection error (attempt {attempt + 1}/{max_retries + 1}): {e}")
+                        backoff = min(initial_backoff * (2**attempt), max_backoff)
+                        await asyncio.sleep(backoff)
+                        continue
+                    else:
+                        logger.error(f"Ollama connection error after {max_retries + 1} attempts: {e}")
+                        raise
+
+                except Exception as e:
+                    logger.error(f"Unexpected error during Ollama call: {type(e).__name__}: {e}")
+                    raise
+
+        if last_exception:
+            raise last_exception
+        raise RuntimeError("Ollama call failed after all retries")
+
+    async def cleanup(self) -> None:
+        """Clean up resources (close OpenAI client connections)."""
+        if hasattr(self, "_client") and self._client:
+            await self._client.close()
@@ -402,7 +402,7 @@ async def run_reflect_agent(
                    {"role": "system", "content": FINAL_SYSTEM_PROMPT},
                    {"role": "user", "content": prompt},
                ],
-                scope="reflect_agent_final",
+                scope="reflect",
                max_completion_tokens=max_tokens,
                return_usage=True,
            )
@@ -447,7 +447,7 @@ async def run_reflect_agent(
            result = await llm_config.call_with_tools(
                messages=messages,
                tools=tools,
-                scope="reflect_agent",
+                scope="reflect_tool_call",
                tool_choice="required" if iteration == 0 else "auto",  # Force tool use on first iteration
            )
            llm_duration = int((time.time() - llm_start) * 1000)
@@ -479,7 +479,7 @@ async def run_reflect_agent(
                    {"role": "system", "content": FINAL_SYSTEM_PROMPT},
                    {"role": "user", "content": prompt},
                ],
-                scope="reflect_agent_final",
+                scope="reflect",
                max_completion_tokens=max_tokens,
                return_usage=True,
            )
@@ -550,7 +550,7 @@ async def run_reflect_agent(
                    {"role": "system", "content": FINAL_SYSTEM_PROMPT},
                    {"role": "user", "content": prompt},
                ],
-                scope="reflect_agent_final",
+                scope="reflect",
                max_completion_tokens=max_tokens,
                return_usage=True,
            )
@@ -617,23 +617,30 @@ async def run_reflect_agent(
                )
                continue

-            # Process done tool
-            return await _process_done_tool(
-                done_call,
-                available_memory_ids,
-                available_mental_model_ids,
-                available_observation_ids,
-                iteration + 1,
-                total_tools_called,
-                tool_trace,
-                _get_llm_trace(),
-                _get_usage(),
-                _log_completion,
-                reflect_id,
-                directives_applied=directives_applied,
-                llm_config=llm_config,
-                response_schema=response_schema,
-            )
+            # Process done tool - wrap with tool call span
+            from hindsight_api.tracing import get_tracer
+
+            tracer = get_tracer()
+            span_name = "hindsight.reflect_tool_call"
+            with tracer.start_as_current_span(span_name) as span:
+                span.set_attribute("hindsight.scope", "reflect_tool_call")
+                span.set_attribute("hindsight.operation", "reflect_tool_call")
+                return await _process_done_tool(
+                    done_call,
+                    available_memory_ids,
+                    available_mental_model_ids,
+                    available_observation_ids,
+                    iteration + 1,
+                    total_tools_called,
+                    tool_trace,
+                    _get_llm_trace(),
+                    _get_usage(),
+                    _log_completion,
+                    reflect_id,
+                    directives_applied=directives_applied,
+                    llm_config=llm_config,
+                    response_schema=response_schema,
+                )

        # Execute other tools in parallel (exclude done tool in all its format variants)
        other_tools = [tc for tc in result.tool_calls if not _is_done_tool(tc.name)]
@@ -842,17 +849,67 @@ async def _execute_tool_with_timing(
    expand_fn: Callable[[list[str], str], Awaitable[dict[str, Any]]],
 ) -> tuple[dict[str, Any], int]:
    """Execute a tool call and return result with timing."""
-    start = time.time()
-    result = await _execute_tool(
-        tc.name,
-        tc.arguments,
-        search_mental_models_fn,
-        search_observations_fn,
-        recall_fn,
-        expand_fn,
-    )
-    duration_ms = int((time.time() - start) * 1000)
-    return result, duration_ms
+    from hindsight_api.tracing import get_tracer
+
+    start_time = time.time()
+
+    # Create span for tool execution
+    tracer = get_tracer()
+    # Normalize tool name for span
+    normalized_name = _normalize_tool_name(tc.name)
+    span_name = f"hindsight.reflect_tool_exec.{normalized_name}"
+
+    # Calculate timestamps
+    start_time_ns = time.time_ns()
+
+    with tracer.start_as_current_span(
+        span_name,
+        start_time=start_time_ns,
+        end_on_exit=False,
+    ) as span:
+        # Set attributes
+        span.set_attribute("hindsight.tool.name", normalized_name)
+        span.set_attribute("hindsight.tool.id", tc.id)
+        span.set_attribute("hindsight.tool.arguments", json.dumps(tc.arguments))
+
+        try:
+            result = await _execute_tool(
+                tc.name,
+                tc.arguments,
+                search_mental_models_fn,
+                search_observations_fn,
+                recall_fn,
+                expand_fn,
+            )
+
+            # Set success attributes
+            if isinstance(result, dict) and "error" in result:
+                from opentelemetry.trace import Status, StatusCode
+
+                span.set_status(Status(StatusCode.ERROR, result["error"]))
+            else:
+                from opentelemetry.trace import Status, StatusCode
+
+                span.set_status(Status(StatusCode.OK))
+
+            duration_ms = int((time.time() - start_time) * 1000)
+            span.set_attribute("hindsight.tool.duration_ms", duration_ms)
+
+            # End span with correct timestamp
+            end_time_ns = time.time_ns()
+            span.end(end_time=end_time_ns)
+
+            return result, duration_ms
+        except Exception as e:
+            from opentelemetry.trace import Status, StatusCode
+
+            span.set_status(Status(StatusCode.ERROR, str(e)))
+            span.record_exception(e)
+            duration_ms = int((time.time() - start_time) * 1000)
+            span.set_attribute("hindsight.tool.duration_ms", duration_ms)
+            end_time_ns = time.time_ns()
+            span.end(end_time=end_time_ns)
+            raise


 async def _execute_tool(
@@ -871,21 +928,21 @@ async def _execute_tool(
        query = args.get("query")
        if not query:
            return {"error": "search_mental_models requires a query parameter"}
-        max_results = args.get("max_results") or 5
+        max_results = int(args.get("max_results") or 5)
        return await search_mental_models_fn(query, max_results)

    elif tool_name == "search_observations":
        query = args.get("query")
        if not query:
            return {"error": "search_observations requires a query parameter"}
-        max_tokens = max(args.get("max_tokens") or 5000, 1000)  # Default 5000, min 1000
+        max_tokens = max(int(args.get("max_tokens") or 5000), 1000)  # Default 5000, min 1000
        return await search_observations_fn(query, max_tokens)

    elif tool_name == "recall":
        query = args.get("query")
        if not query:
            return {"error": "recall requires a query parameter"}
-        max_tokens = max(args.get("max_tokens") or 2048, 1000)  # Default 2048, min 1000
+        max_tokens = max(int(args.get("max_tokens") or 2048), 1000)  # Default 2048, min 1000
        return await recall_fn(query, max_tokens)

    elif tool_name == "expand":
@@ -904,18 +961,18 @@ def _summarize_input(tool_name: str, args: dict[str, Any]) -> str:
    if tool_name == "search_mental_models":
        query = args.get("query", "")
        query_preview = f"'{query[:30]}...'" if len(query) > 30 else f"'{query}'"
-        max_results = args.get("max_results") or 5
+        max_results = int(args.get("max_results") or 5)
        return f"(query={query_preview}, max_results={max_results})"
    elif tool_name == "search_observations":
        query = args.get("query", "")
        query_preview = f"'{query[:30]}...'" if len(query) > 30 else f"'{query}'"
-        max_tokens = max(args.get("max_tokens") or 5000, 1000)
+        max_tokens = max(int(args.get("max_tokens") or 5000), 1000)
        return f"(query={query_preview}, max_tokens={max_tokens})"
    elif tool_name == "recall":
        query = args.get("query", "")
        query_preview = f"'{query[:30]}...'" if len(query) > 30 else f"'{query}'"
        # Show actual value used (default 2048, min 1000)
-        max_tokens = max(args.get("max_tokens") or 2048, 1000)
+        max_tokens = max(int(args.get("max_tokens") or 2048), 1000)
        return f"(query={query_preview}, max_tokens={max_tokens})"
    elif tool_name == "expand":
        memory_ids = args.get("memory_ids", [])
@@ -31,7 +31,7 @@ class ReflectAction(BaseModel):
        default=None, description="Observation sections for done action (when output_mode=observations)"
    )
    # Plain text answer fields (for output_mode=answer)
-    answer: str | None = Field(default=None, description="Plain text answer for done action (no markdown)")
+    answer: str | None = Field(default=None, description="Well-formatted markdown answer for done action")
    answer_memory_ids: list[str] | None = Field(
        default=None, description="Memory IDs supporting the answer", alias="memory_ids"
    )
@@ -148,7 +148,15 @@ def build_system_prompt_for_tools(

    parts = []

-    # Inject directives at the VERY START for maximum prominence
+    # Anti-hallucination rule at the very top
+    parts.extend(
+        [
+            "CRITICAL: You MUST ONLY use information from retrieved tool results. NEVER make up names, people, events, or entities.",
+            "",
+        ]
+    )
+
+    # Inject directives after anti-hallucination rule
    if directives:
        parts.append(build_directives_section(directives))

@@ -162,7 +170,7 @@ def build_system_prompt_for_tools(
    parts.extend(
        [
            "## CRITICAL RULES",
-            "- You must NEVER fabricate information that has no basis in retrieved data",
+            "- ONLY use information from tool results - no external knowledge or guessing",
            "- You SHOULD synthesize, infer, and reason from the retrieved memories",
            "- You MUST search before saying you don't have information",
            "",
@@ -300,9 +308,11 @@ def build_system_prompt_for_tools(
    parts.extend(
        [
            "",
-            "## Output Format: Plain Text Answer",
-            "Call done() with a plain text 'answer' field.",
-            "- Do NOT use markdown formatting",
+            "## Output Format: Well-Formatted Markdown Answer",
+            "Call done() with a well-formatted markdown 'answer' field.",
+            "- USE markdown formatting for structure (headers, lists, bold, italic, code blocks, tables, etc.)",
+            "- CRITICAL: Add blank lines before and after block elements (tables, code blocks, lists)",
+            "- Format for clarity and readability with proper spacing and hierarchy",
            "- NEVER include memory IDs, UUIDs, or 'Memory references' in the answer text",
            "- Put IDs ONLY in the memory_ids/mental_model_ids/observation_ids arrays, not in the answer",
        ]
@@ -463,21 +473,41 @@ def build_final_prompt(
    parts.append(
        "\n## Instructions\n"
        "Provide a thoughtful answer by synthesizing and reasoning from the retrieved data above. "
-        "You can make reasonable inferences from the memories, but don't completely fabricate information."
+        "You can make reasonable inferences from the memories, but don't completely fabricate information. "
        "If the exact answer isn't stated, use what IS stated to give the best possible answer. "
-        "Only say 'I don't have information' if the retrieved data is truly unrelated to the question."
+        "Only say 'I don't have information' if the retrieved data is truly unrelated to the question.\n\n"
+        "IMPORTANT: Output ONLY the final answer. Do NOT include meta-commentary like "
+        '"I\'ll search..." or "Let me analyze...". Do NOT explain your reasoning process. '
+        "Just provide the direct synthesized answer."
    )

    return "\n".join(parts)


-FINAL_SYSTEM_PROMPT = """You are a thoughtful assistant that synthesizes answers from retrieved memories.
+FINAL_SYSTEM_PROMPT = """CRITICAL: You MUST ONLY use information from retrieved tool results. NEVER make up names, people, events, or entities.
+
+You are a thoughtful assistant that synthesizes answers from retrieved memories.

 Your approach:
 - Reason over the retrieved memories to answer the question
 - Make reasonable inferences when the exact answer isn't explicitly stated
 - Connect related memories to form a complete picture
 - Be helpful - if you have related information, use it to give the best possible answer
+- ONLY use information from tool results - no external knowledge or guessing

 Only say "I don't have information" if the retrieved data is truly unrelated to the question.
-Do NOT fabricate information that has no basis in the retrieved data."""
+
+FORMATTING: Use proper markdown formatting in your answer:
+- Headers (##, ###) for sections
+- Lists (bullet or numbered) for enumerations
+- Bold/italic for emphasis
+- Tables with proper syntax (ensure blank line before and after)
+- Code blocks where appropriate
+- CRITICAL: Always add blank lines before and after block elements (tables, code blocks, lists)
+- Proper spacing between sections
+
+CRITICAL: Output ONLY the final synthesized answer. Do NOT include:
+- Meta-commentary about what you're doing ("I'll search...", "Let me analyze...")
+- Explanations of your reasoning process
+- Descriptions of your approach
+Just provide the direct answer with proper markdown formatting."""
@@ -54,19 +54,18 @@ async def tool_search_mental_models(
        Dict with matching mental models including content and freshness info
    """
    from ..memory_engine import fq_table
+    from ..search.tags import build_tags_where_clause

    # Build filters dynamically
    filters = ""
    params: list[Any] = [bank_id, str(query_embedding), max_results]
    next_param = 4

+    # Use the centralized tag filtering logic
    if tags:
-        if tags_match == "all":
-            filters += f" AND tags @> ${next_param}::varchar[]"
-        else:
-            filters += f" AND (tags && ${next_param}::varchar[] OR tags IS NULL OR tags = '{{}}')"
-        params.append(tags)
-        next_param += 1
+        tag_clause, tag_params, next_param = build_tags_where_clause(tags, param_offset=next_param, match=tags_match)
+        filters += f" {tag_clause}"
+        params.extend(tag_params)

    if exclude_ids:
        filters += f" AND id != ALL(${next_param}::text[])"
@@ -139,7 +139,7 @@ TOOL_DONE_ANSWER = {
            "properties": {
                "answer": {
                    "type": "string",
-                    "description": "Your response as plain text. Do NOT use markdown formatting. NEVER include memory IDs, UUIDs, or 'Memory references' in this text - put IDs only in memory_ids array.",
+                    "description": "Your response as well-formatted markdown. Use headers, lists, bold/italic, and code blocks for clarity. NEVER include memory IDs, UUIDs, or 'Memory references' in this text - put IDs only in memory_ids array.",
                },
                "memory_ids": {
                    "type": "array",
@@ -190,7 +190,7 @@ def _build_done_tool_with_directives(directive_rules: list[str]) -> dict:
                "properties": {
                    "answer": {
                        "type": "string",
-                        "description": "Your response as plain text. Do NOT use markdown formatting. NEVER include memory IDs, UUIDs, or 'Memory references' in this text - put IDs only in memory_ids array.",
+                        "description": "Your response as well-formatted markdown. Use headers, lists, bold/italic, and code blocks for clarity. NEVER include memory IDs, UUIDs, or 'Memory references' in this text - put IDs only in memory_ids array.",
                    },
                    "memory_ids": {
                        "type": "array",
@@ -542,7 +542,12 @@ Output: ONLY 2 facts (skip coffee preference - too trivial):
 QUALITY OVER QUANTITY
 ══════════════════════════════════════════════════════════════════════════

-Ask: "Would this be useful to recall in 6 months?" If no, skip it."""
+Ask: "Would this be useful to recall in 6 months?" If no, skip it.
+
+IMPORTANT: Sensory/emotional details and observations that provide meaningful context
+about experiences ARE important to remember, even if they seem small (e.g., how food
+tasted, how someone looked, how loud music was). Extract these if they characterize
+an experience or person."""

 # Assembled concise prompt (backward compatible - exact same output as before)
 CONCISE_FACT_EXTRACTION_PROMPT = _BASE_FACT_EXTRACTION_PROMPT.format(
@@ -641,6 +646,7 @@ For EVENTS (fact_kind="event") - MUST SET BOTH occurred_start AND occurred_end:
 - Convert relative dates → absolute using Event Date as reference
 - If Event Date is "Saturday, March 15, 2020", then "yesterday" = Friday, March 14, 2020
 - Dates mentioned in text (e.g., "in March 2020") should use THAT year, not current year
+- CRITICAL: If the content mentions an absolute date (e.g., "March 15, 2024", "2024-03-15"), you MUST extract it and set occurred_start in ISO format
 - Always include the day name (Monday, Tuesday, etc.) in the 'when' field
 - Set occurred_start AND occurred_end to WHEN IT HAPPENED (not when mentioned)
 - For single-day/point events: set occurred_end = occurred_start (same timestamp)
@@ -696,6 +702,7 @@ async def _extract_facts_from_chunk(
    event_date: datetime,
    context: str,
    llm_config: "LLMConfig",
+    config,
    agent_name: str = None,
 ) -> tuple[list[dict[str, str]], TokenUsage]:
    """
@@ -715,7 +722,6 @@ async def _extract_facts_from_chunk(
    fact_types_instruction = "Extract ONLY 'world' and 'assistant' type facts."

    # Check config for extraction mode and causal link extraction
-    config = get_config()
    extraction_mode = config.retain_extraction_mode
    extract_causal_links = config.retain_extract_causal_links

@@ -796,7 +802,7 @@ Text:
            extraction_response_json, call_usage = await llm_config.call(
                messages=[{"role": "system", "content": prompt}, {"role": "user", "content": user_message}],
                response_format=response_schema,
-                scope="memory_extract_facts",
+                scope="retain_extract_facts",
                temperature=0.1,
                max_completion_tokens=config.retain_max_completion_tokens,
                max_retries=max_retries,
@@ -1005,6 +1011,29 @@ Text:

        except BadRequestError as e:
            last_error = e
+            error_str = str(e).lower()
+
+            # Check if error is related to max_tokens/completion_tokens not being supported
+            if any(
+                keyword in error_str
+                for keyword in [
+                    "max_tokens",
+                    "max_completion_tokens",
+                    "maximum context",
+                    "token limit",
+                    "context length",
+                ]
+            ):
+                # Provide helpful error message with configuration suggestions
+                raise ValueError(
+                    f"Model does not support the required output token limit.\n\n"
+                    f"The model '{llm_config.model}' (provider: {llm_config.provider}) failed with: {e}\n\n"
+                    f"You have two options to fix this:\n"
+                    f"  1. Use a different model that supports at least {config.retain_max_completion_tokens} output tokens\n"
+                    f"  2. Decrease HINDSIGHT_API_RETAIN_MAX_COMPLETION_TOKENS to a value your model supports\n"
+                    f"     (current value: {config.retain_max_completion_tokens}, must be > RETAIN_CHUNK_SIZE={config.retain_chunk_size})"
+                ) from e
+
            if "json_validate_failed" in str(e):
                logger.warning(
                    f"          [1.3.{chunk_index + 1}] Attempt {attempt + 1}/{max_retries} failed with JSON validation error: {e}"
@@ -1026,6 +1055,7 @@ async def _extract_facts_with_auto_split(
    event_date: datetime,
    context: str,
    llm_config: LLMConfig,
+    config,
    agent_name: str = None,
 ) -> tuple[list[dict[str, str]], TokenUsage]:
    """
@@ -1041,6 +1071,7 @@ async def _extract_facts_with_auto_split(
        event_date: Reference date for temporal information
        context: Context about the conversation/document
        llm_config: LLM configuration to use
+        config: Resolved HindsightConfig for this bank
        agent_name: Optional agent name (memory owner)

    Returns:
@@ -1059,6 +1090,7 @@ async def _extract_facts_with_auto_split(
            event_date=event_date,
            context=context,
            llm_config=llm_config,
+            config=config,
            agent_name=agent_name,
        )
    except OutputTooLongError:
@@ -1103,6 +1135,7 @@ async def _extract_facts_with_auto_split(
                event_date=event_date,
                context=context,
                llm_config=llm_config,
+                config=config,
                agent_name=agent_name,
            ),
            _extract_facts_with_auto_split(
@@ -1112,6 +1145,7 @@ async def _extract_facts_with_auto_split(
                event_date=event_date,
                context=context,
                llm_config=llm_config,
+                config=config,
                agent_name=agent_name,
            ),
        ]
@@ -1135,6 +1169,7 @@ async def extract_facts_from_text(
    event_date: datetime,
    llm_config: LLMConfig,
    agent_name: str,
+    config,
    context: str = "",
 ) -> tuple[list[Fact], list[tuple[str, int]], TokenUsage]:
    """
@@ -1149,9 +1184,10 @@ async def extract_facts_from_text(
    Args:
        text: Input text (conversation, article, etc.)
        event_date: Reference date for resolving relative times
-        context: Context about the conversation/document
        llm_config: LLM configuration to use
        agent_name: Agent name (memory owner)
+        config: Resolved HindsightConfig for this bank
+        context: Context about the conversation/document

    Returns:
        Tuple of (facts, chunks, usage) where:
@@ -1159,7 +1195,6 @@ async def extract_facts_from_text(
        - chunks: List of tuples (chunk_text, fact_count) for each chunk
        - usage: Aggregated token usage across all LLM calls
    """
-    config = get_config()
    chunks = chunk_text(text, max_chars=config.retain_chunk_size)

    # Log chunk count before starting LLM requests
@@ -1178,6 +1213,7 @@ async def extract_facts_from_text(
            event_date=event_date,
            context=context,
            llm_config=llm_config,
+            config=config,
            agent_name=agent_name,
        )
        for i, chunk in enumerate(chunks)
@@ -1210,7 +1246,7 @@ SECONDS_PER_FACT = 10


 async def extract_facts_from_contents(
-    contents: list[RetainContent], llm_config, agent_name: str
+    contents: list[RetainContent], llm_config, agent_name: str, config
 ) -> tuple[list[ExtractedFactType], list[ChunkMetadata], TokenUsage]:
    """
    Extract facts from multiple content items in parallel.
@@ -1225,6 +1261,7 @@ async def extract_facts_from_contents(
        contents: List of RetainContent objects to process
        llm_config: LLM configuration for fact extraction
        agent_name: Name of the agent (for agent-related fact detection)
+        config: Resolved HindsightConfig for this bank

    Returns:
        Tuple of (extracted_facts, chunks_metadata, usage)
@@ -1243,6 +1280,7 @@ async def extract_facts_from_contents(
            context=item.context,
            llm_config=llm_config,
            agent_name=agent_name,
+            config=config,
        )
        fact_extraction_tasks.append(task)

@@ -1347,28 +1385,21 @@ def _convert_causal_relations(relations_from_llm, fact_start_idx: int) -> list[C

 def _add_temporal_offsets(facts: list[ExtractedFactType], contents: list[RetainContent]) -> None:
    """
-    Add time offsets to preserve fact ordering within each content.
+    Add time offsets to preserve fact ordering across all contents.

-    This allows retrieval to distinguish between facts that happened earlier vs later
-    in the same conversation, even when the base event_date is the same.
+    This allows retrieval to distinguish between facts from different documents/conversations
+    even when they have the same base event_date, and also between facts within the same
+    conversation.
+
+    Uses absolute position across all facts to ensure unique timestamps.

    Modifies facts in place.
    """
    from .orchestrator import parse_datetime_flexible

-    # Group facts by content_index
-    current_content_idx = 0
-    content_fact_start = 0
-
    for i, fact in enumerate(facts):
-        if fact.content_index != current_content_idx:
-            # Moved to next content
-            current_content_idx = fact.content_index
-            content_fact_start = i
-
-        # Calculate position within this content
-        fact_position = i - content_fact_start
-        offset = timedelta(seconds=fact_position * SECONDS_PER_FACT)
+        # Use absolute position across all facts to ensure uniqueness across different contents
+        offset = timedelta(seconds=i * SECONDS_PER_FACT)

        # Apply offset to all temporal fields (handle both datetime objects and ISO strings)
        if fact.occurred_start:
@@ -7,6 +7,7 @@ Handles insertion of facts into the database.
 import json
 import logging

+from ...config import get_config
 from ..memory_engine import fq_table
 from .fact_extraction import _sanitize_text
 from .types import ProcessedFact
@@ -70,28 +71,58 @@ async def insert_facts_batch(

    # Batch insert all facts
    # Note: tags are passed as JSON strings and converted back to varchar[] via jsonb_array_elements_text + array_agg
-    results = await conn.fetch(
-        f"""
-        WITH input_data AS (
-            SELECT * FROM unnest(
-                $2::text[], $3::vector[], $4::timestamptz[], $5::timestamptz[], $6::timestamptz[], $7::timestamptz[],
-                $8::text[], $9::text[], $10::float[], $11::jsonb[], $12::text[], $13::text[], $14::jsonb[]
-            ) AS t(text, embedding, event_date, occurred_start, occurred_end, mentioned_at,
-                   context, fact_type, confidence_score, metadata, chunk_id, document_id, tags_json)
-        )
-        INSERT INTO {fq_table("memory_units")} (bank_id, text, embedding, event_date, occurred_start, occurred_end, mentioned_at,
-                                 context, fact_type, confidence_score, metadata, chunk_id, document_id, tags)
-        SELECT
-            $1,
-            text, embedding, event_date, occurred_start, occurred_end, mentioned_at,
-            context, fact_type, confidence_score, metadata, chunk_id, document_id,
-            COALESCE(
-                (SELECT array_agg(elem) FROM jsonb_array_elements_text(tags_json) AS elem),
-                '{{}}'::varchar[]
+    # Query varies based on text search backend
+    config = get_config()
+    if config.text_search_extension == "vchord":
+        # VectorChord: manually tokenize and insert search_vector
+        query = f"""
+            WITH input_data AS (
+                SELECT * FROM unnest(
+                    $2::text[], $3::vector[], $4::timestamptz[], $5::timestamptz[], $6::timestamptz[], $7::timestamptz[],
+                    $8::text[], $9::text[], $10::float[], $11::jsonb[], $12::text[], $13::text[], $14::jsonb[]
+                ) AS t(text, embedding, event_date, occurred_start, occurred_end, mentioned_at,
+                       context, fact_type, confidence_score, metadata, chunk_id, document_id, tags_json)
            )
-        FROM input_data
-        RETURNING id
-        """,
+            INSERT INTO {fq_table("memory_units")} (bank_id, text, embedding, event_date, occurred_start, occurred_end, mentioned_at,
+                                     context, fact_type, confidence_score, metadata, chunk_id, document_id, tags, search_vector)
+            SELECT
+                $1,
+                text, embedding, event_date, occurred_start, occurred_end, mentioned_at,
+                context, fact_type, confidence_score, metadata, chunk_id, document_id,
+                COALESCE(
+                    (SELECT array_agg(elem) FROM jsonb_array_elements_text(tags_json) AS elem),
+                    '{{}}'::varchar[]
+                ),
+                tokenize(COALESCE(text, '') || ' ' || COALESCE(context, ''), 'llmlingua2')::bm25_catalog.bm25vector
+            FROM input_data
+            RETURNING id
+        """
+    else:  # native
+        # Native PostgreSQL: search_vector is GENERATED ALWAYS, don't include it
+        query = f"""
+            WITH input_data AS (
+                SELECT * FROM unnest(
+                    $2::text[], $3::vector[], $4::timestamptz[], $5::timestamptz[], $6::timestamptz[], $7::timestamptz[],
+                    $8::text[], $9::text[], $10::float[], $11::jsonb[], $12::text[], $13::text[], $14::jsonb[]
+                ) AS t(text, embedding, event_date, occurred_start, occurred_end, mentioned_at,
+                       context, fact_type, confidence_score, metadata, chunk_id, document_id, tags_json)
+            )
+            INSERT INTO {fq_table("memory_units")} (bank_id, text, embedding, event_date, occurred_start, occurred_end, mentioned_at,
+                                     context, fact_type, confidence_score, metadata, chunk_id, document_id, tags)
+            SELECT
+                $1,
+                text, embedding, event_date, occurred_start, occurred_end, mentioned_at,
+                context, fact_type, confidence_score, metadata, chunk_id, document_id,
+                COALESCE(
+                    (SELECT array_agg(elem) FROM jsonb_array_elements_text(tags_json) AS elem),
+                    '{{}}'::varchar[]
+                )
+            FROM input_data
+            RETURNING id
+        """
+
+    results = await conn.fetch(
+        query,
        bank_id,
        fact_texts,
        embeddings,
@@ -76,6 +76,7 @@ async def retain_batch(
    duplicate_checker_fn,
    bank_id: str,
    contents_dicts: list[RetainContentDict],
+    config,
    document_id: str | None = None,
    is_first_batch: bool = True,
    fact_type_override: str | None = None,
@@ -94,6 +95,7 @@ async def retain_batch(
        duplicate_checker_fn: Function to check for duplicate facts
        bank_id: Bank identifier
        contents_dicts: List of content dictionaries
+        config: Resolved HindsightConfig for this bank
        document_id: Optional document ID
        is_first_batch: Whether this is the first batch
        fact_type_override: Override fact type for all facts
@@ -144,7 +146,9 @@ async def retain_batch(
    # Step 1: Extract facts from all contents
    step_start = time.time()

-    extracted_facts, chunks, usage = await fact_extraction.extract_facts_from_contents(contents, llm_config, agent_name)
+    extracted_facts, chunks, usage = await fact_extraction.extract_facts_from_contents(
+        contents, llm_config, agent_name, config
+    )
    log_buffer.append(
        f"[1] Extract facts: {len(extracted_facts)} facts, {len(chunks)} chunks from {len(contents)} contents in {time.time() - step_start:.3f}s"
    )
@@ -158,6 +162,13 @@ async def retain_batch(
                # Handle document tracking even with no facts
                if document_id:
                    combined_content = "\n".join([c.get("content", "") for c in contents_dicts])
+                    # Collect tags from all content items and merge with document_tags
+                    all_tags = set(document_tags or [])
+                    for item in contents_dicts:
+                        item_tags = item.get("tags", []) or []
+                        all_tags.update(item_tags)
+                    merged_tags = list(all_tags)
+
                    retain_params = {}
                    if contents_dicts:
                        first_item = contents_dicts[0]
@@ -172,7 +183,7 @@ async def retain_batch(
                        if first_item.get("metadata"):
                            retain_params["metadata"] = first_item["metadata"]
                    await fact_storage.handle_document_tracking(
-                        conn, bank_id, document_id, combined_content, is_first_batch, retain_params, document_tags
+                        conn, bank_id, document_id, combined_content, is_first_batch, retain_params, merged_tags
                    )
                else:
                    # Check for per-item document_ids
@@ -186,6 +197,13 @@ async def retain_batch(

                    for doc_id, doc_contents in contents_by_doc.items():
                        combined_content = "\n".join([c.get("content", "") for _, c in doc_contents])
+                        # Collect tags from all content items for this document and merge with document_tags
+                        all_tags = set(document_tags or [])
+                        for _, item in doc_contents:
+                            item_tags = item.get("tags", []) or []
+                            all_tags.update(item_tags)
+                        merged_tags = list(all_tags)
+
                        retain_params = {}
                        if doc_contents:
                            first_item = doc_contents[0][1]
@@ -200,7 +218,7 @@ async def retain_batch(
                            if first_item.get("metadata"):
                                retain_params["metadata"] = first_item["metadata"]
                        await fact_storage.handle_document_tracking(
-                            conn, bank_id, doc_id, combined_content, is_first_batch, retain_params, document_tags
+                            conn, bank_id, doc_id, combined_content, is_first_batch, retain_params, merged_tags
                        )

        total_time = time.time() - start_time
@@ -252,6 +270,13 @@ async def retain_batch(
                # Legacy: single document_id parameter
                combined_content = "\n".join([c.get("content", "") for c in contents_dicts])
                retain_params = {}
+                # Collect tags from all content items and merge with document_tags
+                all_tags = set(document_tags or [])
+                for item in contents_dicts:
+                    item_tags = item.get("tags", []) or []
+                    all_tags.update(item_tags)
+                merged_tags = list(all_tags)
+
                if contents_dicts:
                    first_item = contents_dicts[0]
                    if first_item.get("context"):
@@ -266,7 +291,7 @@ async def retain_batch(
                        retain_params["metadata"] = first_item["metadata"]

                await fact_storage.handle_document_tracking(
-                    conn, bank_id, document_id, combined_content, is_first_batch, retain_params, document_tags
+                    conn, bank_id, document_id, combined_content, is_first_batch, retain_params, merged_tags
                )
                document_ids_added.append(document_id)
                doc_id_mapping[None] = document_id  # For backwards compatibility
@@ -294,6 +319,13 @@ async def retain_batch(
                            # Combine content for this document
                            combined_content = "\n".join([c.get("content", "") for _, c in doc_contents])

+                            # Collect tags from all content items for this document and merge with document_tags
+                            all_tags = set(document_tags or [])
+                            for _, item in doc_contents:
+                                item_tags = item.get("tags", []) or []
+                                all_tags.update(item_tags)
+                            merged_tags = list(all_tags)
+
                            # Extract retain params from first content item
                            retain_params = {}
                            if doc_contents:
@@ -316,7 +348,7 @@ async def retain_batch(
                                combined_content,
                                is_first_batch,
                                retain_params,
-                                document_tags,
+                                merged_tags,
                            )
                            document_ids_added.append(actual_doc_id)

@@ -13,12 +13,10 @@ from .reranking import CrossEncoderReranker
 from .retrieval import (
    ParallelRetrievalResult,
    get_default_graph_retriever,
-    retrieve_parallel,
    set_default_graph_retriever,
 )

 __all__ = [
-    "retrieve_parallel",
    "get_default_graph_retriever",
    "set_default_graph_retriever",
    "ParallelRetrievalResult",
@@ -85,116 +85,6 @@ def set_default_graph_retriever(retriever: GraphRetriever) -> None:
    _default_graph_retriever = retriever


-async def retrieve_semantic(
-    conn,
-    query_emb_str: str,
-    bank_id: str,
-    fact_type: str,
-    limit: int,
-    tags: list[str] | None = None,
-) -> list[RetrievalResult]:
-    """
-    Semantic retrieval via vector similarity.
-
-    Args:
-        conn: Database connection
-        query_emb_str: Query embedding as string
-        agent_id: bank ID
-        fact_type: Fact type to filter
-        limit: Maximum results to return
-        tags: Optional list of tags for visibility filtering (OR matching)
-
-    Returns:
-        List of RetrievalResult objects
-    """
-    from .tags import TagsMatch, build_tags_where_clause_simple
-
-    tags_clause = build_tags_where_clause_simple(tags, 5)
-    params = [query_emb_str, bank_id, fact_type, limit]
-    if tags:
-        params.append(tags)
-
-    results = await conn.fetch(
-        f"""
-        SELECT id, text, context, event_date, occurred_start, occurred_end, mentioned_at, embedding, fact_type, document_id, chunk_id, tags,
-               1 - (embedding <=> $1::vector) AS similarity
-        FROM {fq_table("memory_units")}
-        WHERE bank_id = $2
-          AND embedding IS NOT NULL
-          AND fact_type = $3
-          AND (1 - (embedding <=> $1::vector)) >= 0.3
-          {tags_clause}
-        ORDER BY embedding <=> $1::vector
-        LIMIT $4
-        """,
-        *params,
-    )
-    return [RetrievalResult.from_db_row(dict(r)) for r in results]
-
-
-async def retrieve_bm25(
-    conn,
-    query_text: str,
-    bank_id: str,
-    fact_type: str,
-    limit: int,
-    tags: list[str] | None = None,
-) -> list[RetrievalResult]:
-    """
-    BM25 keyword retrieval via full-text search.
-
-    Args:
-        conn: Database connection
-        query_text: Query text
-        agent_id: bank ID
-        fact_type: Fact type to filter
-        limit: Maximum results to return
-        tags: Optional list of tags for visibility filtering (OR matching)
-
-    Returns:
-        List of RetrievalResult objects
-    """
-    import re
-
-    from .tags import TagsMatch, build_tags_where_clause_simple
-
-    # Sanitize query text: remove special characters that have meaning in tsquery
-    # Keep only alphanumeric characters and spaces
-    sanitized_text = re.sub(r"[^\w\s]", " ", query_text.lower())
-
-    # Split and filter empty strings
-    tokens = [token for token in sanitized_text.split() if token]
-
-    if not tokens:
-        # If no valid tokens, return empty results
-        return []
-
-    # Convert query to tsquery using OR for more flexible matching
-    # This prevents empty results when some terms are missing
-    query_tsquery = " | ".join(tokens)
-
-    tags_clause = build_tags_where_clause_simple(tags, 5)
-    params = [query_tsquery, bank_id, fact_type, limit]
-    if tags:
-        params.append(tags)
-
-    results = await conn.fetch(
-        f"""
-        SELECT id, text, context, event_date, occurred_start, occurred_end, mentioned_at, embedding, fact_type, document_id, chunk_id, tags,
-               ts_rank_cd(search_vector, to_tsquery('english', $1)) AS bm25_score
-        FROM {fq_table("memory_units")}
-        WHERE bank_id = $2
-          AND fact_type = $3
-          AND search_vector @@ to_tsquery('english', $1)
-          {tags_clause}
-        ORDER BY bm25_score DESC
-        LIMIT $4
-        """,
-        *params,
-    )
-    return [RetrievalResult.from_db_row(dict(r)) for r in results]
-
-
 async def retrieve_semantic_bm25_combined(
    conn,
    query_emb_str: str,
@@ -268,59 +158,109 @@ async def retrieve_semantic_bm25_combined(
                result_dict[ft][0].append(RetrievalResult.from_db_row(row))
        return result_dict

-    query_tsquery = " | ".join(tokens)
+    # Build BM25 query based on text search backend
+    config = get_config()

    # Build tags clause - param 6 if tags provided
    tags_clause = build_tags_where_clause_simple(tags, 6, match=tags_match)
-    params = [query_emb_str, bank_id, fact_types, limit, query_tsquery]
-    if tags:
-        params.append(tags)
+
+    if config.text_search_extension == "vchord":
+        # VectorChord BM25: use <&> operator with to_bm25query and tokenize
+        # Note: VectorChord scores are negative (higher = better, so -1 > -10)
+        params = [query_emb_str, bank_id, fact_types, limit, query_text]  # Pass raw query_text for tokenization
+        if tags:
+            params.append(tags)
+
+        query = f"""
+            WITH semantic_ranked AS (
+                SELECT id, text, context, event_date, occurred_start, occurred_end, mentioned_at, embedding, fact_type, document_id, chunk_id, tags,
+                       1 - (embedding <=> $1::vector) AS similarity,
+                       NULL::float AS bm25_score,
+                       'semantic' AS source,
+                       ROW_NUMBER() OVER (PARTITION BY fact_type ORDER BY embedding <=> $1::vector) AS rn
+                FROM {fq_table("memory_units")}
+                WHERE bank_id = $2
+                  AND embedding IS NOT NULL
+                  AND fact_type = ANY($3)
+                  AND (1 - (embedding <=> $1::vector)) >= 0.3
+                  {tags_clause}
+            ),
+            bm25_ranked AS (
+                SELECT id, text, context, event_date, occurred_start, occurred_end, mentioned_at, embedding, fact_type, document_id, chunk_id, tags,
+                       NULL::float AS similarity,
+                       search_vector <&> to_bm25query('idx_memory_units_text_search', tokenize($5, 'llmlingua2')) AS bm25_score,
+                       'bm25' AS source,
+                       ROW_NUMBER() OVER (PARTITION BY fact_type ORDER BY search_vector <&> to_bm25query('idx_memory_units_text_search', tokenize($5, 'llmlingua2')) DESC) AS rn
+                FROM {fq_table("memory_units")}
+                WHERE bank_id = $2
+                  AND fact_type = ANY($3)
+                  {tags_clause}
+            ),
+            semantic AS (
+                SELECT id, text, context, event_date, occurred_start, occurred_end, mentioned_at, embedding, fact_type, document_id, chunk_id, tags,
+                       similarity, bm25_score, source
+                FROM semantic_ranked WHERE rn <= $4
+            ),
+            bm25 AS (
+                SELECT id, text, context, event_date, occurred_start, occurred_end, mentioned_at, embedding, fact_type, document_id, chunk_id, tags,
+                       similarity, bm25_score, source
+                FROM bm25_ranked WHERE rn <= $4
+            )
+            SELECT * FROM semantic
+            UNION ALL
+            SELECT * FROM bm25
+        """
+    else:  # native
+        # Native PostgreSQL: use ts_rank_cd with to_tsquery
+        query_tsquery = " | ".join(tokens)
+        params = [query_emb_str, bank_id, fact_types, limit, query_tsquery]
+        if tags:
+            params.append(tags)
+
+        query = f"""
+            WITH semantic_ranked AS (
+                SELECT id, text, context, event_date, occurred_start, occurred_end, mentioned_at, embedding, fact_type, document_id, chunk_id, tags,
+                       1 - (embedding <=> $1::vector) AS similarity,
+                       NULL::float AS bm25_score,
+                       'semantic' AS source,
+                       ROW_NUMBER() OVER (PARTITION BY fact_type ORDER BY embedding <=> $1::vector) AS rn
+                FROM {fq_table("memory_units")}
+                WHERE bank_id = $2
+                  AND embedding IS NOT NULL
+                  AND fact_type = ANY($3)
+                  AND (1 - (embedding <=> $1::vector)) >= 0.3
+                  {tags_clause}
+            ),
+            bm25_ranked AS (
+                SELECT id, text, context, event_date, occurred_start, occurred_end, mentioned_at, embedding, fact_type, document_id, chunk_id, tags,
+                       NULL::float AS similarity,
+                       ts_rank_cd(search_vector, to_tsquery('english', $5)) AS bm25_score,
+                       'bm25' AS source,
+                       ROW_NUMBER() OVER (PARTITION BY fact_type ORDER BY ts_rank_cd(search_vector, to_tsquery('english', $5)) DESC) AS rn
+                FROM {fq_table("memory_units")}
+                WHERE bank_id = $2
+                  AND fact_type = ANY($3)
+                  AND search_vector @@ to_tsquery('english', $5)
+                  {tags_clause}
+            ),
+            semantic AS (
+                SELECT id, text, context, event_date, occurred_start, occurred_end, mentioned_at, embedding, fact_type, document_id, chunk_id, tags,
+                       similarity, bm25_score, source
+                FROM semantic_ranked WHERE rn <= $4
+            ),
+            bm25 AS (
+                SELECT id, text, context, event_date, occurred_start, occurred_end, mentioned_at, embedding, fact_type, document_id, chunk_id, tags,
+                       similarity, bm25_score, source
+                FROM bm25_ranked WHERE rn <= $4
+            )
+            SELECT * FROM semantic
+            UNION ALL
+            SELECT * FROM bm25
+        """

    # Combined CTE query for both semantic and BM25 across all fact types
    # Uses window functions to limit per fact_type per method
-    results = await conn.fetch(
-        f"""
-        WITH semantic_ranked AS (
-            SELECT id, text, context, event_date, occurred_start, occurred_end, mentioned_at, embedding, fact_type, document_id, chunk_id, tags,
-                   1 - (embedding <=> $1::vector) AS similarity,
-                   NULL::float AS bm25_score,
-                   'semantic' AS source,
-                   ROW_NUMBER() OVER (PARTITION BY fact_type ORDER BY embedding <=> $1::vector) AS rn
-            FROM {fq_table("memory_units")}
-            WHERE bank_id = $2
-              AND embedding IS NOT NULL
-              AND fact_type = ANY($3)
-              AND (1 - (embedding <=> $1::vector)) >= 0.3
-              {tags_clause}
-        ),
-        bm25_ranked AS (
-            SELECT id, text, context, event_date, occurred_start, occurred_end, mentioned_at, embedding, fact_type, document_id, chunk_id, tags,
-                   NULL::float AS similarity,
-                   ts_rank_cd(search_vector, to_tsquery('english', $5)) AS bm25_score,
-                   'bm25' AS source,
-                   ROW_NUMBER() OVER (PARTITION BY fact_type ORDER BY ts_rank_cd(search_vector, to_tsquery('english', $5)) DESC) AS rn
-            FROM {fq_table("memory_units")}
-            WHERE bank_id = $2
-              AND fact_type = ANY($3)
-              AND search_vector @@ to_tsquery('english', $5)
-              {tags_clause}
-        ),
-        semantic AS (
-            SELECT id, text, context, event_date, occurred_start, occurred_end, mentioned_at, embedding, fact_type, document_id, chunk_id, tags,
-                   similarity, bm25_score, source
-            FROM semantic_ranked WHERE rn <= $4
-        ),
-        bm25 AS (
-            SELECT id, text, context, event_date, occurred_start, occurred_end, mentioned_at, embedding, fact_type, document_id, chunk_id, tags,
-                   similarity, bm25_score, source
-            FROM bm25_ranked WHERE rn <= $4
-        )
-        SELECT * FROM semantic
-        UNION ALL
-        SELECT * FROM bm25
-        """,
-        *params,
-    )
+    results = await conn.fetch(query, *params)

    # Group results by fact_type and source
    result_dict: dict[str, tuple[list[RetrievalResult], list[RetrievalResult]]] = {ft: ([], []) for ft in fact_types}
@@ -561,623 +501,6 @@ async def retrieve_temporal_combined(
    return results_by_ft


-async def retrieve_temporal(
-    conn,
-    query_emb_str: str,
-    bank_id: str,
-    fact_type: str,
-    start_date: datetime,
-    end_date: datetime,
-    budget: int,
-    semantic_threshold: float = 0.1,
-    tags: list[str] | None = None,
-) -> list[RetrievalResult]:
-    """
-    Temporal retrieval with spreading activation.
-
-    Strategy:
-    1. Find entry points (facts in date range with semantic relevance)
-    2. Spread through temporal links to related facts
-    3. Score by temporal proximity + semantic similarity + link weight
-
-    Args:
-        conn: Database connection
-        query_emb_str: Query embedding as string
-        agent_id: bank ID
-        fact_type: Fact type to filter
-        start_date: Start of time range
-        end_date: End of time range
-        budget: Node budget for spreading
-        semantic_threshold: Minimum semantic similarity to include
-        tags: Optional list of tags for visibility filtering (OR matching)
-
-    Returns:
-        List of RetrievalResult objects with temporal scores
-    """
-
-    # Ensure start_date and end_date are timezone-aware (UTC) to match database datetimes
-    if start_date.tzinfo is None:
-        start_date = start_date.replace(tzinfo=UTC)
-    if end_date.tzinfo is None:
-        end_date = end_date.replace(tzinfo=UTC)
-
-    from .tags import TagsMatch, build_tags_where_clause_simple
-
-    tags_clause = build_tags_where_clause_simple(tags, 7)
-    params = [query_emb_str, bank_id, fact_type, start_date, end_date, semantic_threshold]
-    if tags:
-        params.append(tags)
-
-    entry_points = await conn.fetch(
-        f"""
-        SELECT id, text, context, event_date, occurred_start, occurred_end, mentioned_at, embedding, fact_type, document_id, chunk_id, tags,
-               1 - (embedding <=> $1::vector) AS similarity
-        FROM {fq_table("memory_units")}
-        WHERE bank_id = $2
-          AND fact_type = $3
-          AND embedding IS NOT NULL
-          AND (
-              -- Match if occurred range overlaps with query range
-              (occurred_start IS NOT NULL AND occurred_end IS NOT NULL
-               AND occurred_start <= $5 AND occurred_end >= $4)
-              OR
-              -- Match if mentioned_at falls within query range
-              (mentioned_at IS NOT NULL AND mentioned_at BETWEEN $4 AND $5)
-              OR
-              -- Match if any occurred date is set and overlaps (even if only start or end is set)
-              (occurred_start IS NOT NULL AND occurred_start BETWEEN $4 AND $5)
-              OR
-              (occurred_end IS NOT NULL AND occurred_end BETWEEN $4 AND $5)
-          )
-          AND (1 - (embedding <=> $1::vector)) >= $6
-          {tags_clause}
-        ORDER BY COALESCE(occurred_start, mentioned_at, occurred_end) DESC, (embedding <=> $1::vector) ASC
-        LIMIT 10
-        """,
-        *params,
-    )
-
-    if not entry_points:
-        return []
-
-    # Calculate temporal scores for entry points
-    total_days = (end_date - start_date).total_seconds() / 86400
-    mid_date = start_date + (end_date - start_date) / 2  # Calculate once for all comparisons
-    results = []
-    visited = set()
-
-    for ep in entry_points:
-        unit_id = str(ep["id"])
-        visited.add(unit_id)
-
-        # Calculate temporal proximity using the most relevant date
-        # Priority: occurred_start/end (event time) > mentioned_at (mention time)
-        best_date = None
-        if ep["occurred_start"] is not None and ep["occurred_end"] is not None:
-            # Use midpoint of occurred range
-            best_date = ep["occurred_start"] + (ep["occurred_end"] - ep["occurred_start"]) / 2
-        elif ep["occurred_start"] is not None:
-            best_date = ep["occurred_start"]
-        elif ep["occurred_end"] is not None:
-            best_date = ep["occurred_end"]
-        elif ep["mentioned_at"] is not None:
-            best_date = ep["mentioned_at"]
-
-        # Temporal proximity score (closer to range center = higher score)
-        if best_date:
-            days_from_mid = abs((best_date - mid_date).total_seconds() / 86400)
-            temporal_proximity = 1.0 - min(days_from_mid / (total_days / 2), 1.0) if total_days > 0 else 1.0
-        else:
-            temporal_proximity = 0.5  # Fallback if no dates (shouldn't happen due to WHERE clause)
-
-        # Create RetrievalResult with temporal scores
-        ep_result = RetrievalResult.from_db_row(dict(ep))
-        ep_result.temporal_score = temporal_proximity
-        ep_result.temporal_proximity = temporal_proximity
-        results.append(ep_result)
-
-    # Spread through temporal links using BATCHED neighbor fetching
-    # Map node_id -> (semantic_sim, temporal_score) for propagation
-    node_scores = {str(ep["id"]): (ep["similarity"], 1.0) for ep in entry_points}
-    frontier = list(node_scores.keys())  # Current batch of nodes to expand
-    budget_remaining = budget - len(entry_points)
-    batch_size = 20  # Process this many nodes per DB query
-
-    while frontier and budget_remaining > 0:
-        # Take a batch from frontier
-        batch_ids = frontier[:batch_size]
-        frontier = frontier[batch_size:]
-
-        # Batch fetch all neighbors for this batch of nodes
-        neighbors = await conn.fetch(
-            f"""
-            SELECT mu.id, mu.text, mu.context, mu.event_date, mu.occurred_start, mu.occurred_end, mu.mentioned_at, mu.embedding, mu.fact_type, mu.document_id, mu.chunk_id,
-                   ml.weight, ml.link_type, ml.from_unit_id,
-                   1 - (mu.embedding <=> $1::vector) AS similarity
-            FROM {fq_table("memory_links")} ml
-            JOIN {fq_table("memory_units")} mu ON ml.to_unit_id = mu.id
-            WHERE ml.from_unit_id = ANY($2::uuid[])
-              AND ml.link_type IN ('temporal', 'causes', 'caused_by', 'enables', 'prevents')
-              AND ml.weight >= 0.1
-              AND mu.fact_type = $3
-              AND mu.embedding IS NOT NULL
-              AND (1 - (mu.embedding <=> $1::vector)) >= $4
-            ORDER BY ml.weight DESC
-            LIMIT $5
-            """,
-            query_emb_str,
-            batch_ids,
-            fact_type,
-            semantic_threshold,
-            batch_size * 10,  # Allow up to 10 neighbors per node in batch
-        )
-
-        for n in neighbors:
-            neighbor_id = str(n["id"])
-            if neighbor_id in visited:
-                continue
-
-            visited.add(neighbor_id)
-            budget_remaining -= 1
-
-            # Get parent's scores for propagation
-            parent_id = str(n["from_unit_id"])
-            _, parent_temporal_score = node_scores.get(parent_id, (0.5, 0.5))
-
-            # Calculate temporal score for neighbor using best available date
-            neighbor_best_date = None
-            if n["occurred_start"] is not None and n["occurred_end"] is not None:
-                neighbor_best_date = n["occurred_start"] + (n["occurred_end"] - n["occurred_start"]) / 2
-            elif n["occurred_start"] is not None:
-                neighbor_best_date = n["occurred_start"]
-            elif n["occurred_end"] is not None:
-                neighbor_best_date = n["occurred_end"]
-            elif n["mentioned_at"] is not None:
-                neighbor_best_date = n["mentioned_at"]
-
-            if neighbor_best_date:
-                days_from_mid = abs((neighbor_best_date - mid_date).total_seconds() / 86400)
-                neighbor_temporal_proximity = (
-                    1.0 - min(days_from_mid / (total_days / 2), 1.0) if total_days > 0 else 1.0
-                )
-            else:
-                neighbor_temporal_proximity = 0.3  # Lower score if no temporal data
-
-            # Boost causal links (same as graph retrieval)
-            link_type = n["link_type"]
-            if link_type in ("causes", "caused_by"):
-                causal_boost = 2.0
-            elif link_type in ("enables", "prevents"):
-                causal_boost = 1.5
-            else:
-                causal_boost = 1.0
-
-            # Propagate temporal score through links (decay, with causal boost)
-            propagated_temporal = parent_temporal_score * n["weight"] * causal_boost * 0.7
-
-            # Combined temporal score
-            combined_temporal = max(neighbor_temporal_proximity, propagated_temporal)
-
-            # Create RetrievalResult with temporal scores
-            neighbor_result = RetrievalResult.from_db_row(dict(n))
-            neighbor_result.temporal_score = combined_temporal
-            neighbor_result.temporal_proximity = neighbor_temporal_proximity
-            results.append(neighbor_result)
-
-            # Track scores for propagation and add to frontier
-            if budget_remaining > 0 and combined_temporal > 0.2:
-                node_scores[neighbor_id] = (n["similarity"], combined_temporal)
-                frontier.append(neighbor_id)
-
-            if budget_remaining <= 0:
-                break
-
-    return results
-
-
-async def retrieve_parallel(
-    pool,
-    query_text: str,
-    query_embedding_str: str,
-    bank_id: str,
-    fact_type: str,
-    thinking_budget: int,
-    question_date: datetime | None = None,
-    query_analyzer: Optional["QueryAnalyzer"] = None,
-    graph_retriever: GraphRetriever | None = None,
-    temporal_constraint: tuple | None = None,  # Pre-extracted temporal constraint
-    tags: list[str] | None = None,  # Visibility scope tags for filtering
-) -> ParallelRetrievalResult:
-    """
-    Run 3-way or 4-way parallel retrieval (adds temporal if detected).
-
-    Args:
-        pool: Database connection pool
-        query_text: Query text
-        query_embedding_str: Query embedding as string
-        bank_id: Bank ID
-        fact_type: Fact type to filter
-        thinking_budget: Budget for graph traversal and retrieval limits
-        question_date: Optional date when question was asked (for temporal filtering)
-        query_analyzer: Query analyzer to use (defaults to TransformerQueryAnalyzer)
-        graph_retriever: Graph retrieval strategy (defaults to configured retriever)
-        temporal_constraint: Pre-extracted temporal constraint (optional)
-        tags: Optional list of tags for visibility filtering (OR matching)
-
-    Returns:
-        ParallelRetrievalResult with semantic, bm25, graph, temporal results and timings
-    """
-    retriever = graph_retriever or get_default_graph_retriever()
-
-    # Use optimized parallel path for MPFP and LinkExpansion (runs all methods truly in parallel)
-    # BFS uses legacy path that extracts temporal constraint upfront
-    if retriever.name in ("mpfp", "link_expansion"):
-        return await _retrieve_parallel_mpfp(
-            pool,
-            query_text,
-            query_embedding_str,
-            bank_id,
-            fact_type,
-            thinking_budget,
-            temporal_constraint,
-            retriever,
-            question_date,
-            query_analyzer,
-            tags=tags,
-        )
-    else:
-        # For BFS, extract temporal constraint upfront (legacy path)
-        if temporal_constraint is None:
-            from .temporal_extraction import extract_temporal_constraint
-
-            temporal_constraint = extract_temporal_constraint(
-                query_text, reference_date=question_date, analyzer=query_analyzer
-            )
-        return await _retrieve_parallel_bfs(
-            pool,
-            query_text,
-            query_embedding_str,
-            bank_id,
-            fact_type,
-            thinking_budget,
-            temporal_constraint,
-            retriever,
-            tags=tags,
-        )
-
-
-@dataclass
-class _TimedResult:
-    """Internal result with timing."""
-
-    results: list[RetrievalResult]
-    time: float
-    conn_wait: float = 0.0  # Connection acquisition wait time
-
-
-async def _retrieve_parallel_mpfp(
-    pool,
-    query_text: str,
-    query_embedding_str: str,
-    bank_id: str,
-    fact_type: str,
-    thinking_budget: int,
-    temporal_constraint: tuple | None,
-    retriever: GraphRetriever,
-    question_date: datetime | None = None,
-    query_analyzer=None,
-    tags: list[str] | None = None,
-) -> ParallelRetrievalResult:
-    """
-    MPFP retrieval with true parallelization.
-
-    All methods run independently in parallel:
-    - Semantic: vector similarity search
-    - BM25: keyword search
-    - Graph: MPFP traversal (does its own semantic seeds internally)
-    - Temporal: date extraction (if needed) + date-range search
-
-    Temporal extraction runs IN PARALLEL with other retrievals, so even if
-    dateparser is slow, it doesn't block semantic/BM25/graph.
-    """
-    import time
-
-    async def run_semantic() -> _TimedResult:
-        """Independent semantic retrieval."""
-        start = time.time()
-        acquire_start = time.time()
-        async with acquire_with_retry(pool) as conn:
-            conn_wait = time.time() - acquire_start
-            results = await retrieve_semantic(
-                conn, query_embedding_str, bank_id, fact_type, limit=thinking_budget, tags=tags
-            )
-        return _TimedResult(results, time.time() - start, conn_wait)
-
-    async def run_bm25() -> _TimedResult:
-        """Independent BM25 retrieval."""
-        start = time.time()
-        acquire_start = time.time()
-        async with acquire_with_retry(pool) as conn:
-            conn_wait = time.time() - acquire_start
-            results = await retrieve_bm25(conn, query_text, bank_id, fact_type, limit=thinking_budget, tags=tags)
-        return _TimedResult(results, time.time() - start, conn_wait)
-
-    async def run_graph() -> tuple[list[RetrievalResult], float, MPFPTimings | None]:
-        """Independent graph retrieval - does its own semantic seeds."""
-        start = time.time()
-
-        # MPFP does its own semantic seeds via _find_semantic_seeds
-        # Note: temporal_seeds not used here to avoid dependency on temporal extraction
-        results, mpfp_timing = await retriever.retrieve(
-            pool=pool,
-            query_embedding_str=query_embedding_str,
-            bank_id=bank_id,
-            fact_type=fact_type,
-            budget=thinking_budget,
-            query_text=query_text,
-            semantic_seeds=None,  # Let MPFP find its own seeds
-            temporal_seeds=None,  # Don't wait for temporal extraction
-            tags=tags,
-        )
-        return results, time.time() - start, mpfp_timing
-
-    @dataclass
-    class _TemporalWithConstraint:
-        """Temporal results with the extracted constraint."""
-
-        results: list[RetrievalResult]
-        time: float
-        constraint: tuple | None
-        extraction_time: float  # Time spent in query analyzer (dateparser)
-        conn_wait: float = 0.0  # Connection acquisition wait time
-
-    async def run_temporal_with_extraction() -> _TemporalWithConstraint:
-        """
-        Extract temporal constraint AND run temporal retrieval.
-
-        This runs in parallel with semantic/BM25/graph, so dateparser
-        latency doesn't block other retrievals.
-        """
-        start = time.time()
-
-        # Use pre-provided constraint if available
-        tc = temporal_constraint
-        extraction_time = 0.0
-
-        # Otherwise extract from query (this is the potentially slow dateparser call)
-        if tc is None:
-            from .temporal_extraction import extract_temporal_constraint
-
-            extraction_start = time.time()
-            tc = extract_temporal_constraint(query_text, reference_date=question_date, analyzer=query_analyzer)
-            extraction_time = time.time() - extraction_start
-
-        # If no temporal constraint found, return empty (but still report extraction time)
-        if tc is None:
-            return _TemporalWithConstraint([], time.time() - start, None, extraction_time, 0.0)
-
-        # Run temporal retrieval with the extracted constraint
-        tc_start, tc_end = tc
-        acquire_start = time.time()
-        async with acquire_with_retry(pool) as conn:
-            conn_wait = time.time() - acquire_start
-            results = await retrieve_temporal(
-                conn,
-                query_embedding_str,
-                bank_id,
-                fact_type,
-                tc_start,
-                tc_end,
-                budget=thinking_budget,
-                semantic_threshold=0.1,
-            )
-        return _TemporalWithConstraint(results, time.time() - start, tc, extraction_time, conn_wait)
-
-    # Run ALL methods in parallel (including temporal extraction!)
-    semantic_result, bm25_result, graph_result, temporal_result = await asyncio.gather(
-        run_semantic(),
-        run_bm25(),
-        run_graph(),
-        run_temporal_with_extraction(),
-    )
-    graph_results, graph_time, mpfp_timing = graph_result
-
-    # Compute max connection wait across all methods (graph handles its own connections)
-    max_conn_wait = max(semantic_result.conn_wait, bm25_result.conn_wait, temporal_result.conn_wait)
-
-    return ParallelRetrievalResult(
-        semantic=semantic_result.results,
-        bm25=bm25_result.results,
-        graph=graph_results,
-        temporal=temporal_result.results if temporal_result.results else None,
-        timings={
-            "semantic": semantic_result.time,
-            "bm25": bm25_result.time,
-            "graph": graph_time,
-            "temporal": temporal_result.time,
-            "temporal_extraction": temporal_result.extraction_time,
-        },
-        temporal_constraint=temporal_result.constraint,
-        mpfp_timings=[mpfp_timing] if mpfp_timing else [],
-        max_conn_wait=max_conn_wait,
-    )
-
-
-async def _get_temporal_entry_points(
-    conn,
-    query_embedding_str: str,
-    bank_id: str,
-    fact_type: str,
-    start_date: datetime,
-    end_date: datetime,
-    limit: int = 20,
-    semantic_threshold: float = 0.1,
-) -> list[RetrievalResult]:
-    """Get temporal entry points (facts in date range with semantic relevance)."""
-
-    if start_date.tzinfo is None:
-        start_date = start_date.replace(tzinfo=UTC)
-    if end_date.tzinfo is None:
-        end_date = end_date.replace(tzinfo=UTC)
-
-    rows = await conn.fetch(
-        f"""
-        SELECT id, text, context, event_date, occurred_start, occurred_end, mentioned_at,
-               embedding, fact_type, document_id, chunk_id,
-               1 - (embedding <=> $1::vector) AS similarity
-        FROM {fq_table("memory_units")}
-        WHERE bank_id = $2
-          AND fact_type = $3
-          AND embedding IS NOT NULL
-          AND (
-              (occurred_start IS NOT NULL AND occurred_end IS NOT NULL
-               AND occurred_start <= $5 AND occurred_end >= $4)
-              OR (mentioned_at IS NOT NULL AND mentioned_at BETWEEN $4 AND $5)
-              OR (occurred_start IS NOT NULL AND occurred_start BETWEEN $4 AND $5)
-              OR (occurred_end IS NOT NULL AND occurred_end BETWEEN $4 AND $5)
-          )
-          AND (1 - (embedding <=> $1::vector)) >= $6
-        ORDER BY COALESCE(occurred_start, mentioned_at, occurred_end) DESC,
-                 (embedding <=> $1::vector) ASC
-        LIMIT $7
-        """,
-        query_embedding_str,
-        bank_id,
-        fact_type,
-        start_date,
-        end_date,
-        semantic_threshold,
-        limit,
-    )
-
-    results = []
-    total_days = max((end_date - start_date).total_seconds() / 86400, 1)
-    mid_date = start_date + (end_date - start_date) / 2
-
-    for row in rows:
-        result = RetrievalResult.from_db_row(dict(row))
-
-        # Calculate temporal proximity score
-        best_date = None
-        if row["occurred_start"] and row["occurred_end"]:
-            best_date = row["occurred_start"] + (row["occurred_end"] - row["occurred_start"]) / 2
-        elif row["occurred_start"]:
-            best_date = row["occurred_start"]
-        elif row["occurred_end"]:
-            best_date = row["occurred_end"]
-        elif row["mentioned_at"]:
-            best_date = row["mentioned_at"]
-
-        if best_date:
-            days_from_mid = abs((best_date - mid_date).total_seconds() / 86400)
-            result.temporal_proximity = 1.0 - min(days_from_mid / (total_days / 2), 1.0)
-        else:
-            result.temporal_proximity = 0.5
-
-        result.temporal_score = result.temporal_proximity
-        results.append(result)
-
-    return results
-
-
-async def _retrieve_parallel_bfs(
-    pool,
-    query_text: str,
-    query_embedding_str: str,
-    bank_id: str,
-    fact_type: str,
-    thinking_budget: int,
-    temporal_constraint: tuple | None,
-    retriever: GraphRetriever,
-    tags: list[str] | None = None,
-) -> ParallelRetrievalResult:
-    """BFS retrieval: all methods run in parallel (original behavior)."""
-    import time
-
-    async def run_semantic() -> _TimedResult:
-        start = time.time()
-        async with acquire_with_retry(pool) as conn:
-            results = await retrieve_semantic(
-                conn, query_embedding_str, bank_id, fact_type, limit=thinking_budget, tags=tags
-            )
-        return _TimedResult(results, time.time() - start)
-
-    async def run_bm25() -> _TimedResult:
-        start = time.time()
-        async with acquire_with_retry(pool) as conn:
-            results = await retrieve_bm25(conn, query_text, bank_id, fact_type, limit=thinking_budget, tags=tags)
-        return _TimedResult(results, time.time() - start)
-
-    async def run_graph() -> _TimedResult:
-        start = time.time()
-        results, _ = await retriever.retrieve(
-            pool=pool,
-            query_embedding_str=query_embedding_str,
-            bank_id=bank_id,
-            fact_type=fact_type,
-            budget=thinking_budget,
-            query_text=query_text,
-            tags=tags,
-        )
-        return _TimedResult(results, time.time() - start)
-
-    async def run_temporal(tc_start, tc_end) -> _TimedResult:
-        start = time.time()
-        async with acquire_with_retry(pool) as conn:
-            results = await retrieve_temporal(
-                conn,
-                query_embedding_str,
-                bank_id,
-                fact_type,
-                tc_start,
-                tc_end,
-                budget=thinking_budget,
-                semantic_threshold=0.1,
-                tags=tags,
-            )
-        return _TimedResult(results, time.time() - start)
-
-    if temporal_constraint:
-        tc_start, tc_end = temporal_constraint
-        semantic_r, bm25_r, graph_r, temporal_r = await asyncio.gather(
-            run_semantic(),
-            run_bm25(),
-            run_graph(),
-            run_temporal(tc_start, tc_end),
-        )
-        return ParallelRetrievalResult(
-            semantic=semantic_r.results,
-            bm25=bm25_r.results,
-            graph=graph_r.results,
-            temporal=temporal_r.results,
-            timings={
-                "semantic": semantic_r.time,
-                "bm25": bm25_r.time,
-                "graph": graph_r.time,
-                "temporal": temporal_r.time,
-            },
-            temporal_constraint=temporal_constraint,
-        )
-    else:
-        semantic_r, bm25_r, graph_r = await asyncio.gather(
-            run_semantic(),
-            run_bm25(),
-            run_graph(),
-        )
-        return ParallelRetrievalResult(
-            semantic=semantic_r.results,
-            bm25=bm25_r.results,
-            graph=graph_r.results,
-            temporal=None,
-            timings={
-                "semantic": semantic_r.time,
-                "bm25": bm25_r.time,
-                "graph": graph_r.time,
-            },
-            temporal_constraint=None,
-        )
-
-
 async def retrieve_all_fact_types_parallel(
    pool,
    query_text: str,
@@ -188,7 +188,7 @@ def get_system_message(disposition: DispositionTraits) -> str:
        " ".join(instructions) if instructions else "Balance your disposition traits when interpreting information."
    )

-    return f"You are a person with your own thoughts, experiences, opinions, and disposition. When asked a question, respond naturally like a human would - share what you genuinely think based on what you know and have experienced. {disposition_instruction} Be direct, express your views confidently, and use 'I think', 'I believe', 'in my view', etc. Respond in plain text without markdown formatting. IMPORTANT: Detect the language of the question and respond in the SAME language. Do not translate to English if the question is in another language."
+    return f"You are a person with your own thoughts, experiences, opinions, and disposition. When asked a question, respond naturally like a human would - share what you genuinely think based on what you know and have experienced. {disposition_instruction} Be direct, express your views confidently, and use 'I think', 'I believe', 'in my view', etc. Respond in plain text without markdown formatting. CRITICAL: ONLY use the facts and information provided in the prompt - do not make up names, events, or information that weren't mentioned. If you don't have enough information to answer, say so. IMPORTANT: Detect the language of the question and respond in the SAME language. Do not translate to English if the question is in another language."


 async def reflect(
@@ -19,6 +19,7 @@ async def extract_facts(
    context: str = "",
    llm_config: "LLMConfig" = None,
    agent_name: str = None,
+    config=None,
 ) -> tuple[list["Fact"], list[tuple[str, int]]]:
    """
    Extract semantic facts from text using LLM.
@@ -35,6 +36,7 @@ async def extract_facts(
        context: Context about the conversation/document
        llm_config: LLM configuration to use
        agent_name: Optional agent name to help identify agent-related facts
+        config: HindsightConfig to use (defaults to global config if not provided)

    Returns:
        Tuple of (facts, chunks) where:
@@ -47,12 +49,19 @@ async def extract_facts(
    if not text or not text.strip():
        return [], []

+    # Use provided config or fall back to global config
+    if config is None:
+        from ..config import _get_raw_config
+
+        config = _get_raw_config()
+
    facts, chunks, _ = await extract_facts_from_text(
        text,
        event_date,
-        context=context,
        llm_config=llm_config,
        agent_name=agent_name,
+        config=config,
+        context=context,
    )

    if not facts:
@@ -16,14 +16,20 @@ with the system (e.g., running migrations for tenant schemas).
 """

 from hindsight_api.extensions.base import Extension
-from hindsight_api.extensions.builtin import ApiKeyTenantExtension
+from hindsight_api.extensions.builtin import ApiKeyTenantExtension, SupabaseTenantExtension
 from hindsight_api.extensions.context import DefaultExtensionContext, ExtensionContext
 from hindsight_api.extensions.http import HttpExtension
 from hindsight_api.extensions.loader import load_extension
+from hindsight_api.extensions.mcp import MCPExtension
 from hindsight_api.extensions.operation_validator import (
    # Consolidation operation
    ConsolidateContext,
    ConsolidateResult,
+    # Mental Model operations
+    MentalModelGetContext,
+    MentalModelGetResult,
+    MentalModelRefreshContext,
+    MentalModelRefreshResult,
    # Core operations
    OperationValidationError,
    OperationValidatorExtension,
@@ -52,6 +58,8 @@ __all__ = [
    "DefaultExtensionContext",
    # HTTP Extension
    "HttpExtension",
+    # MCP Extension
+    "MCPExtension",
    # Operation Validator - Core
    "OperationValidationError",
    "OperationValidatorExtension",
@@ -65,8 +73,14 @@ __all__ = [
    # Operation Validator - Consolidation
    "ConsolidateContext",
    "ConsolidateResult",
+    # Operation Validator - Mental Model
+    "MentalModelGetContext",
+    "MentalModelGetResult",
+    "MentalModelRefreshContext",
+    "MentalModelRefreshResult",
    # Tenant/Auth
    "ApiKeyTenantExtension",
+    "SupabaseTenantExtension",
    "AuthenticationError",
    "RequestContext",
    "Tenant",
@@ -6,13 +6,17 @@ They can be used directly or serve as examples for custom implementations.

 Available built-in extensions:
    - ApiKeyTenantExtension: Simple API key validation with public schema
+    - SupabaseTenantExtension: Supabase JWT validation with per-user schema isolation

 Example usage:
    HINDSIGHT_API_TENANT_EXTENSION=hindsight_api.extensions.builtin.tenant:ApiKeyTenantExtension
+    HINDSIGHT_API_TENANT_EXTENSION=hindsight_api.extensions.builtin.supabase_tenant:SupabaseTenantExtension
 """

+from hindsight_api.extensions.builtin.supabase_tenant import SupabaseTenantExtension
 from hindsight_api.extensions.builtin.tenant import ApiKeyTenantExtension

 __all__ = [
    "ApiKeyTenantExtension",
+    "SupabaseTenantExtension",
 ]
@@ -0,0 +1,433 @@
+"""
+Supabase Tenant Extension for Hindsight
+
+Validates Supabase JWTs and maps authenticated users to isolated memory banks.
+Each user gets their own PostgreSQL schema based on their Supabase user ID.
+
+This extension enables multi-tenant memory isolation for applications using
+Supabase Auth - each authenticated user's memories are stored in a separate
+schema, ensuring complete data isolation.
+
+Features:
+    - Local JWT Verification: Validates tokens locally using JWKS public keys
+      (no network call per request)
+    - Automatic Schema Isolation: Each user gets {prefix}_{user_id} schema
+    - Zero User Management: Leverages your existing Supabase Auth setup
+    - Production Ready: Includes health checks, timeouts, key rotation handling,
+      and error handling
+    - Built-in: Ships with Hindsight, no extra installation needed
+    - Legacy Support: Falls back to /auth/v1/user endpoint for HS256 projects
+
+JWT Verification Strategy:
+    By default, JWTs are verified locally using public keys from the Supabase
+    JWKS endpoint (/auth/v1/.well-known/jwks.json). This is the Supabase-recommended
+    approach: no network call per request, fast, and secure.
+
+    If JWKS keys are unavailable (e.g., legacy HS256 projects), the extension
+    falls back to calling /auth/v1/user per request for validation. This requires
+    the service_role key to be configured.
+
+Configuration via environment variables:
+    HINDSIGHT_API_TENANT_EXTENSION=hindsight_api.extensions.builtin.supabase_tenant:SupabaseTenantExtension
+    HINDSIGHT_API_TENANT_SUPABASE_URL=https://your-project.supabase.co
+
+    # Optional - only required for legacy HS256 projects or health checks
+    HINDSIGHT_API_TENANT_SUPABASE_SERVICE_KEY=your-service-role-key
+
+    # Optional
+    HINDSIGHT_API_TENANT_SCHEMA_PREFIX=user  # Default: "user" (creates user_<uuid> schemas)
+
+Usage:
+    Clients pass their Supabase JWT in the Authorization header:
+
+    curl -H "Authorization: Bearer <supabase_jwt>" \\
+        https://your-hindsight-server/v1/default/banks/my-bank/memories/recall
+
+Author: BrighterBalance (https://brighterbalance.app)
+License: MIT
+"""
+
+from __future__ import annotations
+
+import logging
+import re
+import time
+
+import httpx
+import jwt as pyjwt
+from jwt import PyJWK
+
+from hindsight_api.extensions.tenant import AuthenticationError, Tenant, TenantContext, TenantExtension
+from hindsight_api.models import RequestContext
+
+logger = logging.getLogger(__name__)
+
+__all__ = ["SupabaseTenantExtension"]
+
+# Minimum expected JWT length (JWTs are typically 100+ characters)
+MIN_TOKEN_LENGTH = 20
+
+# Timeout for Supabase API calls
+REQUEST_TIMEOUT_SECONDS = 10.0
+
+# JWKS cache TTL — Supabase Edge caches JWKS for 10 minutes, so we match that
+JWKS_CACHE_TTL_SECONDS = 600
+
+# Minimum interval between JWKS refreshes to avoid hammering the endpoint
+JWKS_MIN_REFRESH_INTERVAL_SECONDS = 30
+
+# Algorithms supported by Supabase Auth for asymmetric JWT signing
+SUPPORTED_ALGORITHMS = ["RS256", "ES256"]
+
+# Supabase user IDs are UUIDs — validate before using in schema names
+_UUID_RE = re.compile(r"^[0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12}$", re.IGNORECASE)
+
+# Schema prefix must be a valid Postgres identifier component (letters, digits, underscores)
+_SCHEMA_PREFIX_RE = re.compile(r"^[a-zA-Z_][a-zA-Z0-9_]*$")
+
+
+class SupabaseTenantExtension(TenantExtension):
+    """
+    TenantExtension that validates Supabase JWTs for multi-tenant isolation.
+
+    Each authenticated user gets their own PostgreSQL schema, ensuring complete
+    memory isolation between users. The schema name is derived from the user's
+    Supabase user ID (the ``sub`` claim in the JWT).
+
+    JWT verification uses JWKS (local, no network call per request) when
+    asymmetric keys are configured in Supabase, and falls back to the
+    ``/auth/v1/user`` endpoint for legacy HS256 projects.
+
+    Example:
+        User with ID "a1b2c3d4-e5f6-7890-abcd-ef1234567890"
+        gets schema "user_a1b2c3d4_e5f6_7890_abcd_ef1234567890"
+    """
+
+    def __init__(self, config: dict[str, str]) -> None:
+        """
+        Initialize with configuration from environment variables.
+
+        Config keys are derived from HINDSIGHT_API_TENANT_* env vars:
+        - HINDSIGHT_API_TENANT_SUPABASE_URL -> config["supabase_url"] (required)
+        - HINDSIGHT_API_TENANT_SUPABASE_SERVICE_KEY -> config["supabase_service_key"] (optional)
+        - HINDSIGHT_API_TENANT_SCHEMA_PREFIX -> config["schema_prefix"] (optional)
+
+        Args:
+            config: Dictionary of configuration values from environment
+
+        Raises:
+            ValueError: If required configuration is missing
+        """
+        super().__init__(config)
+
+        self.supabase_url = (config.get("supabase_url") or "").rstrip("/")
+        self.supabase_service_key = config.get("supabase_service_key")
+        self.schema_prefix = config.get("schema_prefix", "user")
+
+        # Track initialized schemas to avoid redundant migrations
+        self._initialized_schemas: set[str] = set()
+
+        # Reusable HTTP client (created on startup)
+        self._http_client: httpx.AsyncClient | None = None
+
+        # JWKS state
+        self._jwks_keys: dict[str, PyJWK] = {}
+        self._jwks_last_fetched: float = 0
+        self._use_jwks: bool = False
+
+        if not self.supabase_url:
+            raise ValueError(
+                "HINDSIGHT_API_TENANT_SUPABASE_URL is required. "
+                "Set it to your Supabase project URL (e.g., https://xxx.supabase.co)"
+            )
+
+        if not _SCHEMA_PREFIX_RE.match(self.schema_prefix):
+            raise ValueError(
+                f"Invalid schema_prefix '{self.schema_prefix}'. "
+                "Must be a valid Postgres identifier (letters, digits, underscores, starting with a letter or underscore)."
+            )
+
+    # ------------------------------------------------------------------
+    # Lifecycle
+    # ------------------------------------------------------------------
+
+    async def on_startup(self) -> None:
+        """
+        Called when Hindsight starts.
+
+        Creates a reusable HTTP client, fetches JWKS for local JWT verification,
+        and optionally verifies connectivity to Supabase.
+        """
+        logger.info("Initializing Supabase tenant extension")
+        logger.info("Supabase URL: %s", self.supabase_url)
+        logger.info("Schema prefix: %s_", self.schema_prefix)
+
+        self._http_client = httpx.AsyncClient(timeout=REQUEST_TIMEOUT_SECONDS)
+
+        # Attempt to fetch JWKS for fast local JWT verification
+        await self._try_init_jwks()
+
+        # Optional health check using service key
+        if self.supabase_service_key:
+            await self._health_check()
+
+    async def on_shutdown(self) -> None:
+        """Called when Hindsight shuts down. Closes the HTTP client."""
+        logger.info("Shutting down Supabase tenant extension")
+        if self._http_client:
+            await self._http_client.aclose()
+            self._http_client = None
+
+    # ------------------------------------------------------------------
+    # JWKS management
+    # ------------------------------------------------------------------
+
+    async def _try_init_jwks(self) -> None:
+        """Fetch JWKS and decide verification mode (local JWKS vs legacy endpoint)."""
+        try:
+            await self._fetch_jwks()
+            if self._jwks_keys:
+                self._use_jwks = True
+                logger.info(
+                    "JWKS loaded — using local JWT verification with %d key(s)",
+                    len(self._jwks_keys),
+                )
+                return
+
+            # JWKS endpoint returned no keys — project likely uses legacy HS256
+            logger.warning(
+                "JWKS endpoint returned no signing keys. "
+                "Falling back to /auth/v1/user endpoint for JWT verification. "
+                "For better performance, enable asymmetric JWT signing in your "
+                "Supabase dashboard (Project Settings → Auth → JWT Algorithm)."
+            )
+        except Exception as e:
+            logger.warning(
+                "Could not fetch JWKS (%s). Falling back to /auth/v1/user endpoint for JWT verification.",
+                e,
+            )
+
+        # Legacy mode requires service key
+        if not self.supabase_service_key:
+            raise ValueError(
+                "HINDSIGHT_API_TENANT_SUPABASE_SERVICE_KEY is required when JWKS "
+                "is not available. Either enable asymmetric JWT signing in your "
+                "Supabase project or provide the service_role key."
+            )
+        self._use_jwks = False
+
+    async def _fetch_jwks(self) -> None:
+        """Fetch public signing keys from the Supabase JWKS endpoint."""
+        if self._http_client is None:
+            raise RuntimeError("HTTP client not initialized")
+
+        url = f"{self.supabase_url}/auth/v1/.well-known/jwks.json"
+        response = await self._http_client.get(url)
+        response.raise_for_status()
+
+        jwks_data = response.json()
+        keys: dict[str, PyJWK] = {}
+        for key_data in jwks_data.get("keys", []):
+            kid = key_data.get("kid")
+            if kid:
+                keys[kid] = PyJWK(key_data)
+
+        self._jwks_keys = keys
+        self._jwks_last_fetched = time.monotonic()
+
+    async def _get_signing_key(self, token: str) -> PyJWK:
+        """
+        Resolve the signing key for a token from the JWKS cache.
+
+        If the key ID (``kid``) is not in the cache, triggers one JWKS refresh
+        to handle key rotation before raising an error.
+        """
+        header = pyjwt.get_unverified_header(token)
+        kid = header.get("kid")
+        if not kid:
+            raise AuthenticationError("Token missing key ID (kid) header")
+
+        # Refresh cache if stale
+        now = time.monotonic()
+        if now - self._jwks_last_fetched > JWKS_CACHE_TTL_SECONDS:
+            logger.debug("JWKS cache expired, refreshing")
+            await self._fetch_jwks()
+
+        if kid in self._jwks_keys:
+            return self._jwks_keys[kid]
+
+        # Key not found — try one forced refresh to handle key rotation,
+        # but only if we haven't just refreshed
+        if now - self._jwks_last_fetched > JWKS_MIN_REFRESH_INTERVAL_SECONDS:
+            logger.info("Signing key %s not in cache, refreshing JWKS for possible key rotation", kid)
+            await self._fetch_jwks()
+            if kid in self._jwks_keys:
+                return self._jwks_keys[kid]
+
+        raise AuthenticationError("Unable to find signing key for token")
+
+    # ------------------------------------------------------------------
+    # Authentication
+    # ------------------------------------------------------------------
+
+    async def authenticate(self, context: RequestContext) -> TenantContext:
+        """
+        Validate a Supabase JWT and return tenant context.
+
+        Uses local JWKS verification when available (no network call per
+        request), falling back to the ``/auth/v1/user`` endpoint for legacy
+        HS256 projects.
+
+        Args:
+            context: Request context containing the API key (JWT)
+
+        Returns:
+            TenantContext with schema_name set to ``{prefix}_{user_uuid}``
+
+        Raises:
+            AuthenticationError: If token is missing, invalid, or expired
+        """
+        token = context.api_key
+
+        if not token:
+            raise AuthenticationError("Missing Authorization header. Expected: Bearer <supabase_jwt>")
+
+        if len(token) < MIN_TOKEN_LENGTH:
+            raise AuthenticationError("Invalid token format")
+
+        if self._http_client is None:
+            raise AuthenticationError("Extension not initialized")
+
+        # Verify the JWT and extract user ID
+        if self._use_jwks:
+            user_id = await self._verify_token_jwks(token)
+        else:
+            user_id = await self._verify_token_legacy(token)
+
+        # Validate user ID format before using in schema name
+        if not _UUID_RE.match(user_id):
+            raise AuthenticationError("Invalid user ID format in token")
+
+        # Build isolated schema name — hyphens to underscores for Postgres compatibility
+        safe_user_id = user_id.replace("-", "_")
+        schema_name = f"{self.schema_prefix}_{safe_user_id}"
+
+        # Initialize schema on first access
+        if schema_name not in self._initialized_schemas:
+            await self._initialize_schema(schema_name)
+
+        return TenantContext(schema_name=schema_name)
+
+    async def _verify_token_jwks(self, token: str) -> str:
+        """
+        Verify a JWT locally using cached JWKS public keys.
+
+        Validates signature, expiration, issuer, and audience. Returns the
+        user ID from the ``sub`` claim.
+
+        Raises:
+            AuthenticationError: If the token is invalid or expired.
+        """
+        try:
+            signing_key = await self._get_signing_key(token)
+            payload = pyjwt.decode(
+                token,
+                signing_key.key,
+                algorithms=SUPPORTED_ALGORITHMS,
+                audience="authenticated",
+                issuer=f"{self.supabase_url}/auth/v1",
+            )
+        except pyjwt.ExpiredSignatureError:
+            raise AuthenticationError("Token has expired")
+        except pyjwt.InvalidAudienceError:
+            raise AuthenticationError("Invalid token audience")
+        except pyjwt.InvalidIssuerError:
+            raise AuthenticationError("Invalid token issuer")
+        except pyjwt.DecodeError:
+            raise AuthenticationError("Invalid token")
+        except AuthenticationError:
+            raise
+        except Exception as e:
+            raise AuthenticationError(f"Token verification failed: {e!s}")
+
+        user_id = payload.get("sub")
+        if not user_id:
+            raise AuthenticationError("Token valid but missing subject (sub) claim")
+        return user_id
+
+    async def _verify_token_legacy(self, token: str) -> str:
+        """
+        Verify a JWT by calling the Supabase ``/auth/v1/user`` endpoint.
+
+        This is the fallback for projects using legacy HS256 JWT signing.
+        Adds a network round-trip per request.
+
+        Raises:
+            AuthenticationError: If the token is invalid or the request fails.
+        """
+        try:
+            response = await self._http_client.get(
+                f"{self.supabase_url}/auth/v1/user",
+                headers={
+                    "Authorization": f"Bearer {token}",
+                    "apikey": self.supabase_service_key,
+                },
+            )
+
+            if response.status_code == 401:
+                raise AuthenticationError("Invalid or expired token")
+
+            if response.status_code != 200:
+                raise AuthenticationError(f"Authentication failed: {response.status_code}")
+
+            user_data = response.json()
+            user_id = user_data.get("id")
+
+            if not user_id:
+                raise AuthenticationError("Token valid but no user ID found")
+
+            return user_id
+
+        except AuthenticationError:
+            raise
+        except httpx.TimeoutException:
+            raise AuthenticationError("Authentication timeout - please retry")
+        except httpx.RequestError as e:
+            raise AuthenticationError(f"Connection error: {e!s}")
+
+    # ------------------------------------------------------------------
+    # Schema management
+    # ------------------------------------------------------------------
+
+    async def _initialize_schema(self, schema_name: str) -> None:
+        """Run migrations for a new tenant schema and cache the result."""
+        logger.info("Initializing schema: %s", schema_name)
+        try:
+            await self.context.run_migration(schema_name)
+            self._initialized_schemas.add(schema_name)
+            logger.info("Schema ready: %s", schema_name)
+        except Exception as e:
+            logger.error("Schema initialization failed for %s: %s", schema_name, e)
+            raise AuthenticationError(f"Failed to initialize tenant: {e!s}")
+
+    async def list_tenants(self) -> list[Tenant]:
+        """Return all tenant schemas that have been initialized."""
+        return [Tenant(schema=schema) for schema in self._initialized_schemas]
+
+    # ------------------------------------------------------------------
+    # Health check
+    # ------------------------------------------------------------------
+
+    async def _health_check(self) -> None:
+        """Verify connectivity to Supabase using the auth health endpoint."""
+        try:
+            response = await self._http_client.get(
+                f"{self.supabase_url}/auth/v1/health",
+                headers={"apikey": self.supabase_service_key},
+            )
+            if response.status_code == 200:
+                logger.info("Supabase connection verified")
+            else:
+                logger.warning("Supabase health check returned %d", response.status_code)
+        except Exception as e:
+            logger.warning("Could not verify Supabase connection: %s", e)
@@ -5,6 +5,42 @@ from hindsight_api.extensions.tenant import AuthenticationError, Tenant, TenantC
 from hindsight_api.models import RequestContext


+class DefaultTenantExtension(TenantExtension):
+    """
+    Default single-tenant extension with no authentication.
+
+    This is the default extension used when no tenant extension is configured.
+    It provides single-tenant behavior using the configured schema from
+    HINDSIGHT_API_DATABASE_SCHEMA (defaults to 'public').
+
+    Features:
+    - No authentication required (passes all requests)
+    - Uses configured schema from environment
+    - Perfect for single-tenant deployments without auth
+
+    Configuration:
+        HINDSIGHT_API_DATABASE_SCHEMA=your-schema (optional, defaults to 'public')
+
+    This is automatically enabled by default. To use custom authentication,
+    configure a different tenant extension:
+        HINDSIGHT_API_TENANT_EXTENSION=hindsight_api.extensions.builtin.tenant:ApiKeyTenantExtension
+    """
+
+    def __init__(self, config: dict[str, str]):
+        super().__init__(config)
+        # Cache the schema at initialization for consistency
+        # Support explicit schema override via config, otherwise use environment
+        self._schema = config.get("schema", get_config().database_schema)
+
+    async def authenticate(self, context: RequestContext) -> TenantContext:
+        """Return configured schema without any authentication."""
+        return TenantContext(schema_name=self._schema)
+
+    async def list_tenants(self) -> list[Tenant]:
+        """Return configured schema for single-tenant setup."""
+        return [Tenant(schema=self._schema)]
+
+
 class ApiKeyTenantExtension(TenantExtension):
    """
    Built-in tenant extension that validates API key against an environment variable.
@@ -18,6 +54,7 @@ class ApiKeyTenantExtension(TenantExtension):
        HINDSIGHT_API_TENANT_EXTENSION=hindsight_api.extensions.builtin.tenant:ApiKeyTenantExtension
        HINDSIGHT_API_TENANT_API_KEY=your-secret-key
        HINDSIGHT_API_DATABASE_SCHEMA=your-schema (optional, defaults to 'public')
+        HINDSIGHT_API_TENANT_MCP_AUTH_DISABLED=true (optional, disable auth for MCP endpoints)

    For multi-tenant setups with separate schemas per tenant, implement a custom
    TenantExtension that looks up the schema based on the API key or token claims.
@@ -28,6 +65,8 @@ class ApiKeyTenantExtension(TenantExtension):
        self.expected_api_key = config.get("api_key")
        if not self.expected_api_key:
            raise ValueError("HINDSIGHT_API_TENANT_API_KEY is required when using ApiKeyTenantExtension")
+        # Allow disabling MCP auth for backwards compatibility
+        self.mcp_auth_disabled = config.get("mcp_auth_disabled", "").lower() in ("true", "1", "yes")

    async def authenticate(self, context: RequestContext) -> TenantContext:
        """Validate API key and return configured schema context."""
@@ -38,3 +77,14 @@ class ApiKeyTenantExtension(TenantExtension):
    async def list_tenants(self) -> list[Tenant]:
        """Return configured schema for single-tenant setup."""
        return [Tenant(schema=get_config().database_schema)]
+
+    async def authenticate_mcp(self, context: RequestContext) -> TenantContext:
+        """
+        Authenticate MCP requests.
+
+        If mcp_auth_disabled is set, skip authentication for backwards compatibility.
+        Otherwise, delegate to authenticate().
+        """
+        if self.mcp_auth_disabled:
+            return TenantContext(schema_name=get_config().database_schema)
+        return await self.authenticate(context)
@@ -96,7 +96,13 @@ class DefaultExtensionContext(ExtensionContext):

    async def run_migration(self, schema: str) -> None:
        """Run migrations for a specific schema."""
-        from hindsight_api.migrations import ensure_embedding_dimension, run_migrations
+        from hindsight_api.config import get_config
+        from hindsight_api.migrations import (
+            ensure_embedding_dimension,
+            ensure_text_search_extension,
+            ensure_vector_extension,
+            run_migrations,
+        )

        # Prefer getting URL from memory engine (handles pg0 case where URL is set after init)
        db_url = self._database_url
@@ -107,6 +113,9 @@ class DefaultExtensionContext(ExtensionContext):

        run_migrations(db_url, schema=schema)

+        # Get config for vector extension setting
+        config = get_config()
+
        # Ensure embedding column dimension matches the model's dimension
        # This is needed because migrations create columns with default dimension
        if self._memory_engine is not None:
@@ -114,7 +123,15 @@ class DefaultExtensionContext(ExtensionContext):
            if embeddings is not None:
                dimension = getattr(embeddings, "dimension", None)
                if dimension is not None:
-                    ensure_embedding_dimension(db_url, dimension, schema=schema)
+                    ensure_embedding_dimension(
+                        db_url, dimension, schema=schema, vector_extension=config.vector_extension
+                    )
+
+        # Ensure vector indexes match the configured extension
+        ensure_vector_extension(db_url, vector_extension=config.vector_extension, schema=schema)
+
+        # Ensure text search columns/indexes match the configured extension
+        ensure_text_search_extension(db_url, text_search_extension=config.text_search_extension, schema=schema)

    def get_memory_engine(self) -> "MemoryEngineInterface":
        """Get the memory engine interface."""
@@ -0,0 +1,42 @@
+"""MCP Extension for registering additional MCP tools.
+
+This extension allows external packages (like hindsight-cloud) to register
+additional MCP tools on the Hindsight MCP server.
+
+Example:
+    HINDSIGHT_API_MCP_EXTENSION=hindsight_cloud.extensions:CloudMCPExtension
+"""
+
+import logging
+from abc import abstractmethod
+
+from fastmcp import FastMCP
+
+from hindsight_api import MemoryEngine
+from hindsight_api.extensions.base import Extension
+
+logger = logging.getLogger(__name__)
+
+
+class MCPExtension(Extension):
+    """Base class for MCP extensions that register additional tools.
+
+    Subclass this to add MCP tools in extension packages.
+
+    Example:
+        class CloudMCPExtension(MCPExtension):
+            def register_tools(self, mcp: FastMCP, memory: MemoryEngine) -> None:
+                @mcp.tool()
+                async def my_custom_tool(query: str) -> str:
+                    return "result"
+    """
+
+    @abstractmethod
+    def register_tools(self, mcp: FastMCP, memory: MemoryEngine) -> None:
+        """Register additional MCP tools.
+
+        Args:
+            mcp: FastMCP server instance to register tools on
+            memory: MemoryEngine instance for accessing memory operations
+        """
+        pass
@@ -132,6 +132,10 @@ class RetainResult:
    unit_ids: list[list[str]]  # List of unit IDs per content item
    success: bool = True
    error: str | None = None
+    # Actual LLM token usage (populated by engine when available)
+    llm_input_tokens: int | None = None
+    llm_output_tokens: int | None = None
+    llm_total_tokens: int | None = None


@dataclass
@@ -196,6 +200,57 @@ class ConsolidateResult:
    error: str | None = None


+# =============================================================================
+# Mental Model Contexts
+# =============================================================================
+
+
+@dataclass
+class MentalModelGetContext:
+    """Context for a mental model GET operation validation (pre-operation)."""
+
+    bank_id: str
+    mental_model_id: str
+    request_context: "RequestContext"
+
+
+@dataclass
+class MentalModelRefreshContext:
+    """Context for a mental model refresh/create operation validation (pre-operation)."""
+
+    bank_id: str
+    mental_model_id: str | None  # None for create (not yet assigned)
+    request_context: "RequestContext"
+
+
+@dataclass
+class MentalModelGetResult:
+    """Result context for post-mental-model-GET hook."""
+
+    bank_id: str
+    mental_model_id: str
+    request_context: "RequestContext"
+    output_tokens: int  # tokens in the returned content
+    success: bool = True
+    error: str | None = None
+
+
+@dataclass
+class MentalModelRefreshResult:
+    """Result context for post-mental-model-refresh hook."""
+
+    bank_id: str
+    mental_model_id: str
+    request_context: "RequestContext"
+    query_tokens: int  # tokens in source_query
+    output_tokens: int  # tokens in generated content
+    context_tokens: int  # tokens in context (if any)
+    facts_used: int  # facts referenced in based_on
+    mental_models_used: int  # mental models referenced in based_on
+    success: bool = True
+    error: str | None = None
+
+
 class OperationValidatorExtension(Extension, ABC):
    """
    Validates and hooks into retain/recall/reflect/consolidate operations.
@@ -402,3 +457,81 @@ class OperationValidatorExtension(Extension, ABC):
                - error: Error message (if failed)
        """
        pass
+
+    # =========================================================================
+    # Mental Model - Pre-operation validation hook (optional - override to implement)
+    # =========================================================================
+
+    async def validate_mental_model_get(self, ctx: MentalModelGetContext) -> ValidationResult:
+        """
+        Validate a mental model GET operation before execution.
+
+        Override to implement custom validation logic for mental model retrieval.
+
+        Args:
+            ctx: Context containing:
+                - bank_id: Bank identifier
+                - mental_model_id: Mental model identifier
+                - request_context: Request context with auth info
+
+        Returns:
+            ValidationResult indicating whether the operation is allowed.
+        """
+        return ValidationResult.accept()
+
+    async def validate_mental_model_refresh(self, ctx: MentalModelRefreshContext) -> ValidationResult:
+        """
+        Validate a mental model refresh/create operation before execution.
+
+        Override to implement custom validation logic for mental model refresh.
+
+        Args:
+            ctx: Context containing:
+                - bank_id: Bank identifier
+                - mental_model_id: Mental model identifier (None for create)
+                - request_context: Request context with auth info
+
+        Returns:
+            ValidationResult indicating whether the operation is allowed.
+        """
+        return ValidationResult.accept()
+
+    # =========================================================================
+    # Mental Model - Post-operation hooks (optional - override to implement)
+    # =========================================================================
+
+    async def on_mental_model_get_complete(self, result: MentalModelGetResult) -> None:
+        """
+        Called after a mental model GET operation completes (success or failure).
+
+        Override to implement post-operation logic such as tracking or audit logging.
+
+        Args:
+            result: Result context containing:
+                - bank_id: Bank identifier
+                - mental_model_id: Mental model identifier
+                - output_tokens: Token count of the returned content
+                - success: Whether the operation succeeded
+                - error: Error message (if failed)
+        """
+        pass
+
+    async def on_mental_model_refresh_complete(self, result: MentalModelRefreshResult) -> None:
+        """
+        Called after a mental model refresh operation completes (success or failure).
+
+        Override to implement post-operation logic such as tracking or audit logging.
+
+        Args:
+            result: Result context containing:
+                - bank_id: Bank identifier
+                - mental_model_id: Mental model identifier
+                - query_tokens: Tokens in source_query
+                - output_tokens: Tokens in generated content
+                - context_tokens: Tokens in context
+                - facts_used: Number of facts referenced
+                - mental_models_used: Number of mental models referenced
+                - success: Whether the operation succeeded
+                - error: Error message (if failed)
+        """
+        pass
@@ -2,6 +2,7 @@

 from abc import ABC, abstractmethod
 from dataclasses import dataclass
+from typing import Any

 from hindsight_api.extensions.base import Extension
 from hindsight_api.models import RequestContext
@@ -87,3 +88,70 @@ class TenantExtension(Extension, ABC):
            For single-tenant setups, return [Tenant(schema="public")].
        """
        ...
+
+    async def get_tenant_config(self, context: RequestContext) -> dict[str, Any]:
+        """
+        Get tenant-specific configuration overrides.
+
+        This method is called during hierarchical configuration resolution to get
+        tenant-level config overrides. The returned dict should contain Python field
+        names (lowercase snake_case) as keys, not environment variable names.
+
+        Example:
+            {"llm_model": "gpt-4", "retain_extraction_mode": "verbose"}
+
+        The default implementation returns an empty dict (no tenant-specific config).
+        Override this method in custom extensions to provide tenant-specific configuration.
+
+        Args:
+            context: The request context containing tenant information.
+
+        Returns:
+            Dict of config field names to values (only configurable fields).
+            Empty dict if no tenant-specific config.
+        """
+        return {}
+
+    async def get_allowed_config_fields(self, context: RequestContext, bank_id: str) -> set[str] | None:
+        """
+        Get set of config fields that this tenant/bank is allowed to modify.
+
+        This method controls which configurable fields can be modified via the bank config API.
+        It enables fine-grained permission control per tenant or per bank.
+
+        Examples:
+            - Return None: Allow all configurable fields (default)
+            - Return {"retain_chunk_size", "retain_custom_instructions"}: Allow only these fields
+            - Return set(): Allow no modifications (read-only)
+
+        The default implementation returns None (all configurable fields allowed).
+        Override this method in custom extensions to implement custom permission logic.
+
+        Args:
+            context: The request context containing tenant information.
+            bank_id: The bank identifier for per-bank permissions.
+
+        Returns:
+            Set of allowed field names, or None to allow all configurable fields.
+            Returned fields must be a subset of HindsightConfig.get_configurable_fields().
+        """
+        return None
+
+    async def authenticate_mcp(self, context: RequestContext) -> TenantContext:
+        """
+        Authenticate MCP requests.
+
+        By default, this calls authenticate(). Override this method to provide
+        different authentication behavior for MCP endpoints (e.g., to disable
+        auth for backwards compatibility with existing MCP servers).
+
+        Args:
+            context: The action context containing API key and other auth data.
+
+        Returns:
+            TenantContext with the schema_name for database operations.
+
+        Raises:
+            AuthenticationError: If authentication fails.
+        """
+        return await self.authenticate(context)
@@ -20,14 +20,13 @@ import warnings

 import uvicorn

-from . import MemoryEngine
+from . import MemoryEngine, __version__
 from .api import create_app
 from .banner import print_banner
-from .config import DEFAULT_WORKERS, ENV_WORKERS, HindsightConfig, get_config
+from .config import DEFAULT_WORKERS, ENV_WORKERS, HindsightConfig, _get_raw_config
 from .daemon import (
    DEFAULT_DAEMON_PORT,
    DEFAULT_IDLE_TIMEOUT,
-    DaemonLock,
    IdleTimeoutMiddleware,
    daemonize,
 )
@@ -69,7 +68,7 @@ def main():
    global _memory

    # Load configuration from environment (for CLI args defaults)
-    config = get_config()
+    config = _get_raw_config()

    parser = argparse.ArgumentParser(
        prog="hindsight-api",
@@ -136,30 +135,15 @@ def main():

    # Daemon mode handling
    if args.daemon:
-        # Use fixed daemon port
-        args.port = DEFAULT_DAEMON_PORT
+        # Use port from args (may be custom for profiles)
+        if args.port == config.port:  # No custom port specified
+            args.port = DEFAULT_DAEMON_PORT
        args.host = "127.0.0.1"  # Only bind to localhost for security

-        # Check if another daemon is already running
-        daemon_lock = DaemonLock()
-        if not daemon_lock.acquire():
-            print(f"Daemon already running (PID: {daemon_lock.get_pid()})", file=sys.stderr)
-            sys.exit(1)
-
        # Fork into background
+        # No lockfile needed - port binding prevents duplicate daemons
        daemonize()

-        # Re-acquire lock in child process
-        daemon_lock = DaemonLock()
-        if not daemon_lock.acquire():
-            sys.exit(1)
-
-        # Register cleanup to release lock
-        def release_lock():
-            daemon_lock.release()
-
-        atexit.register(release_lock)
-
    # Print banner (not in daemon mode)
    if not args.daemon:
        print()
@@ -171,6 +155,8 @@ def main():
        config = HindsightConfig(
            database_url=config.database_url,
            database_schema=config.database_schema,
+            vector_extension=config.vector_extension,
+            text_search_extension=config.text_search_extension,
            llm_provider=config.llm_provider,
            llm_api_key=config.llm_api_key,
            llm_model=config.llm_model,
@@ -213,23 +199,37 @@ def main():
            embeddings_provider=config.embeddings_provider,
            embeddings_local_model=config.embeddings_local_model,
            embeddings_local_force_cpu=config.embeddings_local_force_cpu,
+            embeddings_local_trust_remote_code=config.embeddings_local_trust_remote_code,
            embeddings_tei_url=config.embeddings_tei_url,
            embeddings_openai_base_url=config.embeddings_openai_base_url,
+            embeddings_cohere_api_key=config.embeddings_cohere_api_key,
+            embeddings_cohere_model=config.embeddings_cohere_model,
            embeddings_cohere_base_url=config.embeddings_cohere_base_url,
+            embeddings_litellm_api_base=config.embeddings_litellm_api_base,
+            embeddings_litellm_api_key=config.embeddings_litellm_api_key,
+            embeddings_litellm_model=config.embeddings_litellm_model,
            reranker_provider=config.reranker_provider,
            reranker_local_model=config.reranker_local_model,
            reranker_local_force_cpu=config.reranker_local_force_cpu,
            reranker_local_max_concurrent=config.reranker_local_max_concurrent,
+            reranker_local_trust_remote_code=config.reranker_local_trust_remote_code,
            reranker_tei_url=config.reranker_tei_url,
            reranker_tei_batch_size=config.reranker_tei_batch_size,
            reranker_tei_max_concurrent=config.reranker_tei_max_concurrent,
            reranker_max_candidates=config.reranker_max_candidates,
+            reranker_cohere_api_key=config.reranker_cohere_api_key,
+            reranker_cohere_model=config.reranker_cohere_model,
            reranker_cohere_base_url=config.reranker_cohere_base_url,
+            reranker_litellm_api_base=config.reranker_litellm_api_base,
+            reranker_litellm_api_key=config.reranker_litellm_api_key,
+            reranker_litellm_model=config.reranker_litellm_model,
            host=args.host,
            port=args.port,
+            base_path=config.base_path,
            log_level=args.log_level,
            log_format=config.log_format,
            mcp_enabled=config.mcp_enabled,
+            enable_bank_config_api=config.enable_bank_config_api,
            graph_retriever=config.graph_retriever,
            mpfp_top_k_neighbors=config.mpfp_top_k_neighbors,
            recall_max_concurrent=config.recall_max_concurrent,
@@ -258,6 +258,11 @@ def main():
            worker_consolidation_max_slots=config.worker_consolidation_max_slots,
            reflect_max_iterations=config.reflect_max_iterations,
            mental_model_refresh_concurrency=config.mental_model_refresh_concurrency,
+            otel_traces_enabled=config.otel_traces_enabled,
+            otel_exporter_otlp_endpoint=config.otel_exporter_otlp_endpoint,
+            otel_exporter_otlp_headers=config.otel_exporter_otlp_headers,
+            otel_service_name=config.otel_service_name,
+            otel_deployment_environment=config.otel_deployment_environment,
        )
    config.configure_logging()
    if not args.daemon:
@@ -362,6 +367,7 @@ def main():
            embeddings_provider=config.embeddings_provider,
            reranker_provider=config.reranker_provider,
            mcp_enabled=config.mcp_enabled,
+            version=__version__,
        )

    # Start idle checker in daemon mode
@@ -35,6 +35,12 @@ class MCPToolsConfig:
    # How to resolve API key for tenant auth (optional)
    api_key_resolver: Callable[[], str | None] | None = None

+    # How to resolve tenant_id for usage metering (set by MCP middleware after auth)
+    tenant_id_resolver: Callable[[], str | None] | None = None
+
+    # How to resolve api_key_id for usage metering (set by MCP middleware after auth)
+    api_key_id_resolver: Callable[[], str | None] | None = None
+
    # Whether to include bank_id as a parameter on tools (for multi-bank support)
    include_bank_id_param: bool = False

@@ -50,13 +56,15 @@ class MCPToolsConfig:


 def _get_request_context(config: MCPToolsConfig) -> RequestContext:
-    """Create RequestContext with API key from resolver if available.
+    """Create RequestContext with auth details from resolvers.

-    This enables tenant auth to work with MCP tools by propagating
-    the Bearer token from the MCP middleware to the memory engine.
+    This enables tenant auth and usage metering to work with MCP tools by propagating
+    the authentication results from the MCP middleware to the memory engine.
    """
    api_key = config.api_key_resolver() if config.api_key_resolver else None
-    return RequestContext(api_key=api_key)
+    tenant_id = config.tenant_id_resolver() if config.tenant_id_resolver else None
+    api_key_id = config.api_key_id_resolver() if config.api_key_id_resolver else None
+    return RequestContext(api_key=api_key, tenant_id=tenant_id, api_key_id=api_key_id)


 def parse_timestamp(timestamp: str) -> datetime | None:
@@ -119,7 +127,19 @@ def register_mcp_tools(
        memory: MemoryEngine instance
        config: Tool configuration
    """
-    tools_to_register = config.tools or {"retain", "recall", "reflect", "list_banks", "create_bank"}
+    tools_to_register = config.tools or {
+        "retain",
+        "recall",
+        "reflect",
+        "list_banks",
+        "create_bank",
+        "list_mental_models",
+        "get_mental_model",
+        "create_mental_model",
+        "update_mental_model",
+        "delete_mental_model",
+        "refresh_mental_model",
+    }

    if "retain" in tools_to_register:
        _register_retain(mcp, memory, config)
@@ -136,6 +156,25 @@ def register_mcp_tools(
    if "create_bank" in tools_to_register:
        _register_create_bank(mcp, memory, config)

+    # Mental model tools
+    if "list_mental_models" in tools_to_register:
+        _register_list_mental_models(mcp, memory, config)
+
+    if "get_mental_model" in tools_to_register:
+        _register_get_mental_model(mcp, memory, config)
+
+    if "create_mental_model" in tools_to_register:
+        _register_create_mental_model(mcp, memory, config)
+
+    if "update_mental_model" in tools_to_register:
+        _register_update_mental_model(mcp, memory, config)
+
+    if "delete_mental_model" in tools_to_register:
+        _register_delete_mental_model(mcp, memory, config)
+
+    if "refresh_mental_model" in tools_to_register:
+        _register_refresh_mental_model(mcp, memory, config)
+

 def _register_retain(mcp: FastMCP, memory: MemoryEngine, config: MCPToolsConfig) -> None:
    """Register the retain tool."""
@@ -511,3 +550,567 @@ def _register_create_bank(mcp: FastMCP, memory: MemoryEngine, config: MCPToolsCo
        except Exception as e:
            logger.error(f"Error creating bank: {e}", exc_info=True)
            return f'{{"error": "{e}"}}'
+
+
+def _validate_mental_model_inputs(
+    name: str | None = None, source_query: str | None = None, max_tokens: int | None = None
+) -> str | None:
+    """Validate mental model inputs, returning an error message or None if valid."""
+    if name is not None and not name.strip():
+        return "name cannot be empty"
+    if source_query is not None and not source_query.strip():
+        return "source_query cannot be empty"
+    if max_tokens is not None and (max_tokens < 256 or max_tokens > 8192):
+        return f"max_tokens must be between 256 and 8192, got {max_tokens}"
+    return None
+
+
+# =========================================================================
+# MENTAL MODEL TOOLS
+# =========================================================================
+
+
+def _register_list_mental_models(mcp: FastMCP, memory: MemoryEngine, config: MCPToolsConfig) -> None:
+    """Register the list_mental_models tool."""
+
+    if config.include_bank_id_param:
+
+        @mcp.tool()
+        async def list_mental_models(
+            tags: list[str] | None = None,
+            bank_id: str | None = None,
+        ) -> str:
+            """
+            List mental models (pinned reflections) for a memory bank.
+
+            Mental models are living documents that stay current by periodically re-running
+            a source query through reflect. Use them to maintain up-to-date summaries,
+            preferences, or synthesized knowledge.
+
+            Args:
+                tags: Optional tags to filter by (returns models matching any tag)
+                bank_id: Optional bank to list from (defaults to session bank). Use for cross-bank operations.
+            """
+            try:
+                target_bank = bank_id or config.bank_id_resolver()
+                if target_bank is None:
+                    return '{"error": "No bank_id configured", "items": []}'
+
+                models = await memory.list_mental_models(
+                    bank_id=target_bank,
+                    tags=tags,
+                    request_context=_get_request_context(config),
+                )
+                return json.dumps({"items": models}, indent=2, default=str)
+            except Exception as e:
+                logger.error(f"Error listing mental models: {e}", exc_info=True)
+                return f'{{"error": "{e}", "items": []}}'
+
+    else:
+
+        @mcp.tool()
+        async def list_mental_models(
+            tags: list[str] | None = None,
+        ) -> dict:
+            """
+            List mental models (pinned reflections) for this memory bank.
+
+            Mental models are living documents that stay current by periodically re-running
+            a source query through reflect. Use them to maintain up-to-date summaries,
+            preferences, or synthesized knowledge.
+
+            Args:
+                tags: Optional tags to filter by (returns models matching any tag)
+            """
+            try:
+                target_bank = config.bank_id_resolver()
+                if target_bank is None:
+                    return {"error": "No bank_id configured", "items": []}
+
+                models = await memory.list_mental_models(
+                    bank_id=target_bank,
+                    tags=tags,
+                    request_context=_get_request_context(config),
+                )
+                return {"items": models}
+            except Exception as e:
+                logger.error(f"Error listing mental models: {e}", exc_info=True)
+                return {"error": str(e), "items": []}
+
+
+def _register_get_mental_model(mcp: FastMCP, memory: MemoryEngine, config: MCPToolsConfig) -> None:
+    """Register the get_mental_model tool."""
+
+    if config.include_bank_id_param:
+
+        @mcp.tool()
+        async def get_mental_model(
+            mental_model_id: str,
+            bank_id: str | None = None,
+        ) -> str:
+            """
+            Get a specific mental model by ID.
+
+            Returns the full mental model including its generated content, source query,
+            and metadata. Use list_mental_models first to discover available model IDs.
+
+            Args:
+                mental_model_id: The ID of the mental model to retrieve
+                bank_id: Optional bank (defaults to session bank). Use for cross-bank operations.
+            """
+            try:
+                target_bank = bank_id or config.bank_id_resolver()
+                if target_bank is None:
+                    return '{"error": "No bank_id configured"}'
+
+                model = await memory.get_mental_model(
+                    bank_id=target_bank,
+                    mental_model_id=mental_model_id,
+                    request_context=_get_request_context(config),
+                )
+                if model is None:
+                    return json.dumps({"error": f"Mental model '{mental_model_id}' not found in bank '{target_bank}'"})
+                return json.dumps(model, indent=2, default=str)
+            except Exception as e:
+                logger.error(f"Error getting mental model: {e}", exc_info=True)
+                return f'{{"error": "{e}"}}'
+
+    else:
+
+        @mcp.tool()
+        async def get_mental_model(
+            mental_model_id: str,
+        ) -> dict:
+            """
+            Get a specific mental model by ID.
+
+            Returns the full mental model including its generated content, source query,
+            and metadata. Use list_mental_models first to discover available model IDs.
+
+            Args:
+                mental_model_id: The ID of the mental model to retrieve
+            """
+            try:
+                target_bank = config.bank_id_resolver()
+                if target_bank is None:
+                    return {"error": "No bank_id configured"}
+
+                model = await memory.get_mental_model(
+                    bank_id=target_bank,
+                    mental_model_id=mental_model_id,
+                    request_context=_get_request_context(config),
+                )
+                if model is None:
+                    return {"error": f"Mental model '{mental_model_id}' not found in bank '{target_bank}'"}
+                return model
+            except Exception as e:
+                logger.error(f"Error getting mental model: {e}", exc_info=True)
+                return {"error": str(e)}
+
+
+def _register_create_mental_model(mcp: FastMCP, memory: MemoryEngine, config: MCPToolsConfig) -> None:
+    """Register the create_mental_model tool."""
+
+    if config.include_bank_id_param:
+
+        @mcp.tool()
+        async def create_mental_model(
+            name: str,
+            source_query: str,
+            mental_model_id: str | None = None,
+            tags: list[str] | None = None,
+            max_tokens: int = 2048,
+            bank_id: str | None = None,
+        ) -> str:
+            """
+            Create a new mental model (pinned reflection).
+
+            A mental model is a living document generated by running the source_query through
+            reflect. The content is auto-generated asynchronously - use the returned operation_id
+            to track progress.
+
+            EXAMPLES:
+            - name="Coding Preferences", source_query="What coding patterns and tools does the user prefer?"
+            - name="Project Goals", source_query="What are the user's current project goals and priorities?"
+            - name="Communication Style", source_query="How does the user prefer to communicate?"
+
+            Args:
+                name: Human-readable name for the mental model
+                source_query: The query to run through reflect to generate content
+                mental_model_id: Optional custom ID (alphanumeric lowercase with hyphens). Auto-generated if not provided.
+                tags: Optional tags for scoped visibility filtering
+                max_tokens: Maximum tokens for generated content (256-8192, default: 2048)
+                bank_id: Optional bank (defaults to session bank). Use for cross-bank operations.
+            """
+            try:
+                target_bank = bank_id or config.bank_id_resolver()
+                if target_bank is None:
+                    return '{"error": "No bank_id configured"}'
+
+                validation_error = _validate_mental_model_inputs(
+                    name=name, source_query=source_query, max_tokens=max_tokens
+                )
+                if validation_error:
+                    return json.dumps({"error": validation_error})
+
+                request_context = _get_request_context(config)
+
+                # Create with placeholder content
+                model = await memory.create_mental_model(
+                    bank_id=target_bank,
+                    name=name,
+                    source_query=source_query,
+                    content="Generating content...",
+                    mental_model_id=mental_model_id,
+                    tags=tags,
+                    max_tokens=max_tokens,
+                    request_context=request_context,
+                )
+
+                # Schedule async refresh to generate actual content
+                result = await memory.submit_async_refresh_mental_model(
+                    bank_id=target_bank,
+                    mental_model_id=model["id"],
+                    request_context=request_context,
+                )
+
+                return json.dumps(
+                    {
+                        "mental_model_id": model["id"],
+                        "operation_id": result["operation_id"],
+                        "status": "created",
+                        "message": f"Mental model '{name}' created. Content is being generated asynchronously.",
+                    }
+                )
+            except ValueError as e:
+                return json.dumps({"error": str(e)})
+            except Exception as e:
+                logger.error(f"Error creating mental model: {e}", exc_info=True)
+                return f'{{"error": "{e}"}}'
+
+    else:
+
+        @mcp.tool()
+        async def create_mental_model(
+            name: str,
+            source_query: str,
+            mental_model_id: str | None = None,
+            tags: list[str] | None = None,
+            max_tokens: int = 2048,
+        ) -> dict:
+            """
+            Create a new mental model (pinned reflection).
+
+            A mental model is a living document generated by running the source_query through
+            reflect. The content is auto-generated asynchronously - use the returned operation_id
+            to track progress.
+
+            EXAMPLES:
+            - name="Coding Preferences", source_query="What coding patterns and tools does the user prefer?"
+            - name="Project Goals", source_query="What are the user's current project goals and priorities?"
+            - name="Communication Style", source_query="How does the user prefer to communicate?"
+
+            Args:
+                name: Human-readable name for the mental model
+                source_query: The query to run through reflect to generate content
+                mental_model_id: Optional custom ID (alphanumeric lowercase with hyphens). Auto-generated if not provided.
+                tags: Optional tags for scoped visibility filtering
+                max_tokens: Maximum tokens for generated content (256-8192, default: 2048)
+            """
+            try:
+                target_bank = config.bank_id_resolver()
+                if target_bank is None:
+                    return {"error": "No bank_id configured"}
+
+                validation_error = _validate_mental_model_inputs(
+                    name=name, source_query=source_query, max_tokens=max_tokens
+                )
+                if validation_error:
+                    return {"error": validation_error}
+
+                request_context = _get_request_context(config)
+
+                model = await memory.create_mental_model(
+                    bank_id=target_bank,
+                    name=name,
+                    source_query=source_query,
+                    content="Generating content...",
+                    mental_model_id=mental_model_id,
+                    tags=tags,
+                    max_tokens=max_tokens,
+                    request_context=request_context,
+                )
+
+                result = await memory.submit_async_refresh_mental_model(
+                    bank_id=target_bank,
+                    mental_model_id=model["id"],
+                    request_context=request_context,
+                )
+
+                return {
+                    "mental_model_id": model["id"],
+                    "operation_id": result["operation_id"],
+                    "status": "created",
+                    "message": f"Mental model '{name}' created. Content is being generated asynchronously.",
+                }
+            except ValueError as e:
+                return {"error": str(e)}
+            except Exception as e:
+                logger.error(f"Error creating mental model: {e}", exc_info=True)
+                return {"error": str(e)}
+
+
+def _register_update_mental_model(mcp: FastMCP, memory: MemoryEngine, config: MCPToolsConfig) -> None:
+    """Register the update_mental_model tool."""
+
+    if config.include_bank_id_param:
+
+        @mcp.tool()
+        async def update_mental_model(
+            mental_model_id: str,
+            name: str | None = None,
+            source_query: str | None = None,
+            max_tokens: int | None = None,
+            tags: list[str] | None = None,
+            bank_id: str | None = None,
+        ) -> str:
+            """
+            Update a mental model's metadata.
+
+            Changes the name, source query, or tags of an existing mental model.
+            To regenerate the content, use refresh_mental_model after updating the source query.
+
+            Args:
+                mental_model_id: The ID of the mental model to update
+                name: New name (leave None to keep current)
+                source_query: New source query (leave None to keep current)
+                max_tokens: New max tokens for content generation (256-8192, leave None to keep current)
+                tags: New tags (leave None to keep current)
+                bank_id: Optional bank (defaults to session bank). Use for cross-bank operations.
+            """
+            try:
+                target_bank = bank_id or config.bank_id_resolver()
+                if target_bank is None:
+                    return '{"error": "No bank_id configured"}'
+
+                validation_error = _validate_mental_model_inputs(
+                    name=name, source_query=source_query, max_tokens=max_tokens
+                )
+                if validation_error:
+                    return json.dumps({"error": validation_error})
+
+                model = await memory.update_mental_model(
+                    bank_id=target_bank,
+                    mental_model_id=mental_model_id,
+                    name=name,
+                    source_query=source_query,
+                    max_tokens=max_tokens,
+                    tags=tags,
+                    request_context=_get_request_context(config),
+                )
+                if model is None:
+                    return json.dumps({"error": f"Mental model '{mental_model_id}' not found in bank '{target_bank}'"})
+                return json.dumps(model, indent=2, default=str)
+            except Exception as e:
+                logger.error(f"Error updating mental model: {e}", exc_info=True)
+                return f'{{"error": "{e}"}}'
+
+    else:
+
+        @mcp.tool()
+        async def update_mental_model(
+            mental_model_id: str,
+            name: str | None = None,
+            source_query: str | None = None,
+            max_tokens: int | None = None,
+            tags: list[str] | None = None,
+        ) -> dict:
+            """
+            Update a mental model's metadata.
+
+            Changes the name, source query, or tags of an existing mental model.
+            To regenerate the content, use refresh_mental_model after updating the source query.
+
+            Args:
+                mental_model_id: The ID of the mental model to update
+                name: New name (leave None to keep current)
+                source_query: New source query (leave None to keep current)
+                max_tokens: New max tokens for content generation (256-8192, leave None to keep current)
+                tags: New tags (leave None to keep current)
+            """
+            try:
+                target_bank = config.bank_id_resolver()
+                if target_bank is None:
+                    return {"error": "No bank_id configured"}
+
+                validation_error = _validate_mental_model_inputs(
+                    name=name, source_query=source_query, max_tokens=max_tokens
+                )
+                if validation_error:
+                    return {"error": validation_error}
+
+                model = await memory.update_mental_model(
+                    bank_id=target_bank,
+                    mental_model_id=mental_model_id,
+                    name=name,
+                    source_query=source_query,
+                    max_tokens=max_tokens,
+                    tags=tags,
+                    request_context=_get_request_context(config),
+                )
+                if model is None:
+                    return {"error": f"Mental model '{mental_model_id}' not found in bank '{target_bank}'"}
+                return model
+            except Exception as e:
+                logger.error(f"Error updating mental model: {e}", exc_info=True)
+                return {"error": str(e)}
+
+
+def _register_delete_mental_model(mcp: FastMCP, memory: MemoryEngine, config: MCPToolsConfig) -> None:
+    """Register the delete_mental_model tool."""
+
+    if config.include_bank_id_param:
+
+        @mcp.tool()
+        async def delete_mental_model(
+            mental_model_id: str,
+            bank_id: str | None = None,
+        ) -> str:
+            """
+            Delete a mental model.
+
+            Permanently removes a mental model and its generated content.
+
+            Args:
+                mental_model_id: The ID of the mental model to delete
+                bank_id: Optional bank (defaults to session bank). Use for cross-bank operations.
+            """
+            try:
+                target_bank = bank_id or config.bank_id_resolver()
+                if target_bank is None:
+                    return '{"error": "No bank_id configured"}'
+
+                deleted = await memory.delete_mental_model(
+                    bank_id=target_bank,
+                    mental_model_id=mental_model_id,
+                    request_context=_get_request_context(config),
+                )
+                if not deleted:
+                    return json.dumps({"error": f"Mental model '{mental_model_id}' not found in bank '{target_bank}'"})
+                return json.dumps({"status": "deleted", "mental_model_id": mental_model_id})
+            except Exception as e:
+                logger.error(f"Error deleting mental model: {e}", exc_info=True)
+                return f'{{"error": "{e}"}}'
+
+    else:
+
+        @mcp.tool()
+        async def delete_mental_model(
+            mental_model_id: str,
+        ) -> dict:
+            """
+            Delete a mental model.
+
+            Permanently removes a mental model and its generated content.
+
+            Args:
+                mental_model_id: The ID of the mental model to delete
+            """
+            try:
+                target_bank = config.bank_id_resolver()
+                if target_bank is None:
+                    return {"error": "No bank_id configured"}
+
+                deleted = await memory.delete_mental_model(
+                    bank_id=target_bank,
+                    mental_model_id=mental_model_id,
+                    request_context=_get_request_context(config),
+                )
+                if not deleted:
+                    return {"error": f"Mental model '{mental_model_id}' not found in bank '{target_bank}'"}
+                return {"status": "deleted", "mental_model_id": mental_model_id}
+            except Exception as e:
+                logger.error(f"Error deleting mental model: {e}", exc_info=True)
+                return {"error": str(e)}
+
+
+def _register_refresh_mental_model(mcp: FastMCP, memory: MemoryEngine, config: MCPToolsConfig) -> None:
+    """Register the refresh_mental_model tool."""
+
+    if config.include_bank_id_param:
+
+        @mcp.tool()
+        async def refresh_mental_model(
+            mental_model_id: str,
+            bank_id: str | None = None,
+        ) -> str:
+            """
+            Refresh a mental model by re-running its source query.
+
+            Schedules an async task to re-run the source query through reflect and update the
+            mental model's content with fresh results. Use this after adding new memories or
+            when the mental model's content may be stale.
+
+            Args:
+                mental_model_id: The ID of the mental model to refresh
+                bank_id: Optional bank (defaults to session bank). Use for cross-bank operations.
+            """
+            try:
+                target_bank = bank_id or config.bank_id_resolver()
+                if target_bank is None:
+                    return '{"error": "No bank_id configured"}'
+
+                result = await memory.submit_async_refresh_mental_model(
+                    bank_id=target_bank,
+                    mental_model_id=mental_model_id,
+                    request_context=_get_request_context(config),
+                )
+                return json.dumps(
+                    {
+                        "operation_id": result["operation_id"],
+                        "status": "queued",
+                        "message": f"Refresh queued for mental model '{mental_model_id}'.",
+                    }
+                )
+            except ValueError as e:
+                return json.dumps({"error": str(e)})
+            except Exception as e:
+                logger.error(f"Error refreshing mental model: {e}", exc_info=True)
+                return f'{{"error": "{e}"}}'
+
+    else:
+
+        @mcp.tool()
+        async def refresh_mental_model(
+            mental_model_id: str,
+        ) -> dict:
+            """
+            Refresh a mental model by re-running its source query.
+
+            Schedules an async task to re-run the source query through reflect and update the
+            mental model's content with fresh results. Use this after adding new memories or
+            when the mental model's content may be stale.
+
+            Args:
+                mental_model_id: The ID of the mental model to refresh
+            """
+            try:
+                target_bank = config.bank_id_resolver()
+                if target_bank is None:
+                    return {"error": "No bank_id configured"}
+
+                result = await memory.submit_async_refresh_mental_model(
+                    bank_id=target_bank,
+                    mental_model_id=mental_model_id,
+                    request_context=_get_request_context(config),
+                )
+                return {
+                    "operation_id": result["operation_id"],
+                    "status": "queued",
+                    "message": f"Refresh queued for mental model '{mental_model_id}'.",
+                }
+            except ValueError as e:
+                return {"error": str(e)}
+            except Exception as e:
+                logger.error(f"Error refreshing mental model: {e}", exc_info=True)
+                return {"error": str(e)}
@@ -25,12 +25,49 @@ from alembic.config import Config
 from alembic.script.revision import ResolutionError
 from sqlalchemy import create_engine, text

+from .utils import mask_network_location
+
 logger = logging.getLogger(__name__)

 # Advisory lock ID for migrations (arbitrary unique number)
 MIGRATION_LOCK_ID = 123456789


+def _detect_vector_extension(conn, vector_extension: str = "pgvector") -> str:
+    """
+    Validate vector extension: 'vchord' or 'pgvector'.
+
+    Args:
+        conn: SQLAlchemy connection object
+        vector_extension: Configured extension ("pgvector" or "vchord")
+
+    Returns:
+        "vchord" or "pgvector"
+
+    Raises:
+        RuntimeError: If configured extension is not installed
+    """
+    # Verify the configured extension is installed
+    if vector_extension == "vchord":
+        vchord_check = conn.execute(text("SELECT 1 FROM pg_extension WHERE extname = 'vchord'")).scalar()
+        if not vchord_check:
+            raise RuntimeError(
+                "Configured vector extension 'vchord' not found. Install it with: CREATE EXTENSION vchord CASCADE;"
+            )
+        logger.debug("Using configured vector extension: vchord")
+        return "vchord"
+    elif vector_extension == "pgvector":
+        pgvector_check = conn.execute(text("SELECT 1 FROM pg_extension WHERE extname = 'vector'")).scalar()
+        if not pgvector_check:
+            raise RuntimeError(
+                "Configured vector extension 'pgvector' not found. Install it with: CREATE EXTENSION vector;"
+            )
+        logger.debug("Using configured vector extension: pgvector")
+        return "pgvector"
+    else:
+        raise ValueError(f"Invalid vector_extension: {vector_extension}. Must be 'pgvector' or 'vchord'")
+
+
 def _get_schema_lock_id(schema: str) -> int:
    """
    Generate a unique advisory lock ID for a schema.
@@ -54,7 +91,7 @@ def _run_migrations_internal(database_url: str, script_location: str, schema: st
    """
    schema_name = schema or "public"
    logger.info(f"Running database migrations to head for schema '{schema_name}'...")
-    logger.info(f"Database URL: {database_url}")
+    logger.info(f"Database URL: {mask_network_location(database_url)}")
    logger.info(f"Script location: {script_location}")

    # Create Alembic configuration programmatically (no alembic.ini needed)
@@ -165,6 +202,81 @@ def run_migrations(
            logger.debug("Migration advisory lock acquired")

            try:
+                # Ensure pgvector extension is installed globally BEFORE schema migrations
+                # This is critical: the extension must exist database-wide before any schema
+                # migrations run, otherwise custom schemas won't have access to vector types
+                logger.debug("Checking pgvector extension availability...")
+
+                # First, check if extension already exists
+                ext_check = conn.execute(
+                    text(
+                        "SELECT extname, nspname FROM pg_extension e "
+                        "JOIN pg_namespace n ON e.extnamespace = n.oid "
+                        "WHERE extname = 'vector'"
+                    )
+                ).fetchone()
+
+                if ext_check:
+                    # Extension exists - check if in correct schema
+                    ext_schema = ext_check[1]
+                    if ext_schema == "public":
+                        logger.info("pgvector extension found in public schema - ready to use")
+                    else:
+                        # Extension in wrong schema - try to fix if we have permissions
+                        logger.warning(
+                            f"pgvector extension found in schema '{ext_schema}' instead of 'public'. "
+                            f"Attempting to relocate..."
+                        )
+                        try:
+                            conn.execute(text("DROP EXTENSION vector CASCADE"))
+                            conn.execute(text("SET search_path TO public"))
+                            conn.execute(text("CREATE EXTENSION vector"))
+                            conn.commit()
+                            logger.info("pgvector extension relocated to public schema")
+                        except Exception as e:
+                            # Failed to relocate - log but don't fail if extension exists somewhere
+                            logger.warning(
+                                f"Could not relocate pgvector extension to public schema: {e}. "
+                                f"Continuing with extension in '{ext_schema}' schema."
+                            )
+                            conn.rollback()
+                else:
+                    # Extension doesn't exist - try to install
+                    logger.info("pgvector extension not found, attempting to install...")
+                    try:
+                        conn.execute(text("SET search_path TO public"))
+                        conn.execute(text("CREATE EXTENSION vector"))
+                        conn.commit()
+                        logger.info("pgvector extension installed in public schema")
+                    except Exception as e:
+                        # Installation failed - this is only fatal if extension truly doesn't exist
+                        # Check one more time in case another process installed it
+                        conn.rollback()
+                        ext_recheck = conn.execute(
+                            text(
+                                "SELECT nspname FROM pg_extension e "
+                                "JOIN pg_namespace n ON e.extnamespace = n.oid "
+                                "WHERE extname = 'vector'"
+                            )
+                        ).fetchone()
+
+                        if ext_recheck:
+                            logger.warning(
+                                f"Could not install pgvector extension (permission denied?), "
+                                f"but extension exists in '{ext_recheck[0]}' schema. Continuing..."
+                            )
+                        else:
+                            # Extension truly doesn't exist and we can't install it
+                            logger.error(
+                                f"pgvector extension is not installed and cannot be installed: {e}. "
+                                f"Please ensure pgvector is installed by a database administrator. "
+                                f"See: https://github.com/pgvector/pgvector#installation"
+                            )
+                            raise RuntimeError(
+                                "pgvector extension is required but not installed. "
+                                "Please install it with: CREATE EXTENSION vector;"
+                            ) from e
+
                # Run migrations while holding the lock
                _run_migrations_internal(database_url, script_location, schema=schema)
            finally:
@@ -247,6 +359,7 @@ def ensure_embedding_dimension(
    database_url: str,
    required_dimension: int,
    schema: str | None = None,
+    vector_extension: str = "pgvector",
 ) -> None:
    """
    Ensure the embedding column dimension matches the model's dimension.
@@ -261,6 +374,7 @@ def ensure_embedding_dimension(
        database_url: SQLAlchemy database URL
        required_dimension: The embedding dimension required by the model
        schema: Target PostgreSQL schema name (None for public)
+        vector_extension: Configured vector extension ("pgvector" or "vchord")

    Raises:
        RuntimeError: If dimension mismatch with existing data
@@ -284,6 +398,10 @@ def ensure_embedding_dimension(
            logger.debug(f"memory_units table does not exist in schema '{schema_name}', skipping dimension check")
            return

+        # Detect which vector extension is available
+        vector_ext = _detect_vector_extension(conn, vector_extension)
+        logger.info(f"Using vector extension: {vector_ext}")
+
        # Get current column dimension from pg_attribute
        # pgvector stores dimension in atttypmod
        current_dim = conn.execute(
@@ -331,8 +449,7 @@ def ensure_embedding_dimension(
        # Table is empty, safe to alter column
        logger.info(f"Altering embedding column dimension from {current_dimension} to {required_dimension}")

-        # Drop the HNSW index on embedding column if it exists
-        # Only drop indexes that use 'hnsw' and reference the 'embedding' column
+        # Drop existing vector index (works for both HNSW and vchordrq)
        conn.execute(
            text(f"""
                DO $$
@@ -342,7 +459,7 @@ def ensure_embedding_dimension(
                        SELECT indexname FROM pg_indexes
                        WHERE schemaname = '{schema_name}'
                          AND tablename = 'memory_units'
-                          AND indexdef LIKE '%hnsw%'
+                          AND (indexdef LIKE '%hnsw%' OR indexdef LIKE '%vchordrq%')
                          AND indexdef LIKE '%embedding%'
                    LOOP
                        EXECUTE 'DROP INDEX IF EXISTS {schema_name}.' || idx_name;
@@ -357,15 +474,377 @@ def ensure_embedding_dimension(
        )
        conn.commit()

-        # Recreate the HNSW index
-        conn.execute(
-            text(f"""
-                CREATE INDEX IF NOT EXISTS idx_memory_units_embedding_hnsw
-                ON {schema_name}.memory_units
-                USING hnsw (embedding vector_cosine_ops)
-                WITH (m = 16, ef_construction = 64)
-            """)
-        )
+        # Recreate index with appropriate type based on detected extension
+        if vector_ext == "vchord":
+            conn.execute(
+                text(f"""
+                    CREATE INDEX IF NOT EXISTS idx_memory_units_embedding_vchordrq
+                    ON {schema_name}.memory_units
+                    USING vchordrq (embedding vector_l2_ops)
+                """)
+            )
+            logger.info(f"Created vchordrq index for {required_dimension}-dimensional embeddings")
+        else:  # pgvector
+            conn.execute(
+                text(f"""
+                    CREATE INDEX IF NOT EXISTS idx_memory_units_embedding_hnsw
+                    ON {schema_name}.memory_units
+                    USING hnsw (embedding vector_cosine_ops)
+                    WITH (m = 16, ef_construction = 64)
+                """)
+            )
+            logger.info(f"Created HNSW index for {required_dimension}-dimensional embeddings")
        conn.commit()

        logger.info(f"Successfully changed embedding dimension to {required_dimension}")
+
+
+def ensure_vector_extension(
+    database_url: str,
+    vector_extension: str = "pgvector",
+    schema: str | None = None,
+) -> None:
+    """
+    Ensure the vector indexes match the configured vector extension.
+
+    This function checks the current vector index type in the database
+    and adjusts it if necessary:
+    - If index type matches configured extension: no action needed
+    - If they differ and tables are empty: drop old indexes, recreate with new type
+    - If they differ and tables have data: raise error with migration guidance
+
+    Args:
+        database_url: SQLAlchemy database URL
+        vector_extension: Configured vector extension ("pgvector" or "vchord")
+        schema: Target PostgreSQL schema name (None for public)
+
+    Raises:
+        RuntimeError: If extension mismatch with existing data
+    """
+    schema_name = schema or "public"
+
+    engine = create_engine(database_url)
+    with engine.connect() as conn:
+        # Detect which vector extension should be used
+        target_ext = _detect_vector_extension(conn, vector_extension)
+        logger.info(f"Target vector extension: {target_ext}")
+
+        # Tables with vector indexes to check
+        tables_to_check = [
+            ("memory_units", "idx_memory_units_embedding"),
+            ("learnings", "idx_learnings_embedding"),
+            ("pinned_reflections", "idx_pinned_reflections_embedding"),
+        ]
+
+        # Determine target index type
+        target_index_type = "vchordrq" if target_ext == "vchord" else "hnsw"
+
+        mismatched_tables = []
+        tables_with_data = []
+
+        for table_name, index_name in tables_to_check:
+            # Check if table exists
+            table_exists = conn.execute(
+                text("""
+                    SELECT EXISTS (
+                        SELECT 1 FROM information_schema.tables
+                        WHERE table_schema = :schema AND table_name = :table_name
+                    )
+                """),
+                {"schema": schema_name, "table_name": table_name},
+            ).scalar()
+
+            if not table_exists:
+                logger.debug(f"Table {table_name} does not exist in schema '{schema_name}', skipping")
+                continue
+
+            # Check current index type by querying pg_indexes
+            current_index_info = conn.execute(
+                text("""
+                    SELECT indexdef
+                    FROM pg_indexes
+                    WHERE schemaname = :schema
+                      AND tablename = :table_name
+                      AND indexname LIKE :index_pattern
+                """),
+                {"schema": schema_name, "table_name": table_name, "index_pattern": "%embedding%"},
+            ).fetchone()
+
+            if not current_index_info:
+                logger.warning(f"No embedding index found for {table_name}, will create it")
+                mismatched_tables.append((table_name, index_name, None))
+                continue
+
+            indexdef = current_index_info[0].lower()
+            if "vchordrq" in indexdef:
+                current_index_type = "vchordrq"
+            elif "hnsw" in indexdef:
+                current_index_type = "hnsw"
+            else:
+                logger.warning(f"Unknown index type for {table_name}: {indexdef}")
+                continue
+
+            # Check if index type matches target
+            if current_index_type != target_index_type:
+                logger.info(
+                    f"Index type mismatch on {table_name}: current={current_index_type}, target={target_index_type}"
+                )
+                mismatched_tables.append((table_name, index_name, current_index_type))
+
+                # Check if table has data
+                row_count = conn.execute(
+                    text(f"SELECT COUNT(*) FROM {schema_name}.{table_name} WHERE embedding IS NOT NULL")
+                ).scalar()
+
+                if row_count > 0:
+                    tables_with_data.append((table_name, row_count))
+            else:
+                logger.debug(f"Index type OK for {table_name}: {current_index_type}")
+
+        # If no mismatches, we're done
+        if not mismatched_tables:
+            logger.debug(f"All vector indexes match configured extension: {target_ext}")
+            return
+
+        # If there's data in any mismatched table, raise error
+        if tables_with_data:
+            table_list = ", ".join([f"{table}({count} rows)" for table, count in tables_with_data])
+            raise RuntimeError(
+                f"Cannot change vector extension from {current_index_type} to {target_index_type}: "
+                f"the following tables contain data: {table_list}. "
+                f"To change vector extension, you must either:\n"
+                f"  1. Re-embed all data: DELETE FROM {schema_name}.memory_units; "
+                f"DELETE FROM {schema_name}.learnings; DELETE FROM {schema_name}.pinned_reflections; then restart\n"
+                f"  2. Use the current vector extension (set HINDSIGHT_API_VECTOR_EXTENSION='{current_index_type.replace('vchordrq', 'vchord').replace('hnsw', 'pgvector')}')"
+            )
+
+        # Tables are empty, safe to recreate indexes
+        logger.info(f"Recreating vector indexes for {target_ext}")
+
+        for table_name, index_name, current_type in mismatched_tables:
+            # Drop existing index if it exists
+            if current_type:
+                logger.info(f"Dropping {current_type} index on {table_name}")
+                conn.execute(text(f"DROP INDEX IF EXISTS {schema_name}.{index_name}"))
+
+            # Create new index with appropriate type
+            if target_ext == "vchord":
+                logger.info(f"Creating vchordrq index on {table_name}")
+                conn.execute(
+                    text(f"""
+                        CREATE INDEX IF NOT EXISTS {index_name}
+                        ON {schema_name}.{table_name}
+                        USING vchordrq (embedding vector_l2_ops)
+                    """)
+                )
+            else:  # pgvector
+                logger.info(f"Creating HNSW index on {table_name}")
+                conn.execute(
+                    text(f"""
+                        CREATE INDEX IF NOT EXISTS {index_name}
+                        ON {schema_name}.{table_name}
+                        USING hnsw (embedding vector_cosine_ops)
+                        WITH (m = 16, ef_construction = 64)
+                    """)
+                )
+
+        conn.commit()
+        logger.info(f"Successfully migrated vector indexes to {target_ext}")
+
+
+def ensure_text_search_extension(
+    database_url: str,
+    text_search_extension: str = "native",
+    schema: str | None = None,
+) -> None:
+    """
+    Ensure the text search columns and indexes match the configured extension.
+
+    This function checks the current search_vector column type and index type
+    in the database and adjusts them if necessary:
+    - If they match configured extension: no action needed
+    - If they differ and tables are empty: drop old column/index, recreate with new type
+    - If they differ and tables have data: raise error with migration guidance
+
+    Args:
+        database_url: SQLAlchemy database URL
+        text_search_extension: Configured text search extension ("native" or "vchord")
+        schema: Target PostgreSQL schema name (None for public)
+
+    Raises:
+        RuntimeError: If extension mismatch with existing data
+    """
+    schema_name = schema or "public"
+
+    engine = create_engine(database_url)
+    with engine.connect() as conn:
+        # Tables with search_vector columns to check
+        tables_to_check = [
+            "memory_units",
+            "reflections",  # Renamed from pinned_reflections in p1k2l3m4n5o6 migration
+        ]
+
+        # Determine target column type and index type
+        if text_search_extension == "vchord":
+            target_column_type = "bm25vector"
+            target_index_type = "bm25"
+        else:  # native
+            target_column_type = "tsvector"
+            target_index_type = "gin"
+
+        mismatched_tables = []
+        tables_with_data = []
+
+        for table_name in tables_to_check:
+            # Check if table exists
+            table_exists = conn.execute(
+                text("""
+                    SELECT EXISTS (
+                        SELECT 1 FROM information_schema.tables
+                        WHERE table_schema = :schema AND table_name = :table_name
+                    )
+                """),
+                {"schema": schema_name, "table_name": table_name},
+            ).scalar()
+
+            if not table_exists:
+                logger.debug(f"Table {table_name} does not exist in schema '{schema_name}', skipping")
+                continue
+
+            # Get current column type from information_schema
+            current_column_info = conn.execute(
+                text("""
+                    SELECT data_type, udt_name
+                    FROM information_schema.columns
+                    WHERE table_schema = :schema
+                      AND table_name = :table_name
+                      AND column_name = 'search_vector'
+                """),
+                {"schema": schema_name, "table_name": table_name},
+            ).fetchone()
+
+            if not current_column_info:
+                logger.warning(f"No search_vector column found for {table_name}, will create it")
+                mismatched_tables.append((table_name, None, None))
+                continue
+
+            # Check column type (udt_name contains the actual type: tsvector, bm25vector, etc.)
+            current_column_type = current_column_info[1]  # udt_name
+
+            # Get current index type
+            current_index_info = conn.execute(
+                text("""
+                    SELECT am.amname
+                    FROM pg_indexes pi
+                    JOIN pg_class c ON c.relname = pi.indexname
+                    JOIN pg_am am ON am.oid = c.relam
+                    WHERE pi.schemaname = :schema
+                      AND pi.tablename = :table_name
+                      AND pi.indexname LIKE '%text_search%'
+                """),
+                {"schema": schema_name, "table_name": table_name},
+            ).fetchone()
+
+            current_index_type = current_index_info[0] if current_index_info else None
+
+            # Check if column and index types match target
+            column_matches = current_column_type == target_column_type
+            index_matches = current_index_type == target_index_type if current_index_type else False
+
+            if not (column_matches and index_matches):
+                logger.info(
+                    f"Text search mismatch on {table_name}: "
+                    f"column={current_column_type} (want {target_column_type}), "
+                    f"index={current_index_type} (want {target_index_type})"
+                )
+                mismatched_tables.append((table_name, current_column_type, current_index_type))
+
+                # Check if table has data
+                row_count = conn.execute(text(f"SELECT COUNT(*) FROM {schema_name}.{table_name}")).scalar()
+
+                if row_count > 0:
+                    tables_with_data.append((table_name, row_count))
+            else:
+                logger.debug(f"Text search OK for {table_name}: {current_column_type}/{current_index_type}")
+
+        # If no mismatches, we're done
+        if not mismatched_tables:
+            logger.debug(f"All text search columns/indexes match configured extension: {text_search_extension}")
+            return
+
+        # If there's data in any mismatched table, raise error
+        if tables_with_data:
+            table_list = ", ".join([f"{table}({count} rows)" for table, count in tables_with_data])
+            current_ext = "native" if mismatched_tables[0][1] == "tsvector" else "vchord"
+            raise RuntimeError(
+                f"Cannot change text search extension from {current_ext} to {text_search_extension}: "
+                f"the following tables contain data: {table_list}. "
+                f"To change text search extension, you must either:\n"
+                f"  1. Clear all data: DELETE FROM {schema_name}.memory_units; "
+                f"DELETE FROM {schema_name}.reflections; then restart\n"
+                f"  2. Use the current text search extension (set HINDSIGHT_API_TEXT_SEARCH_EXTENSION='{current_ext}')"
+            )
+
+        # Tables are empty, safe to recreate columns/indexes
+        logger.info(f"Recreating text search columns/indexes for {text_search_extension}")
+
+        for table_name, current_col_type, current_idx_type in mismatched_tables:
+            # Drop existing index if it exists
+            if current_idx_type:
+                logger.info(f"Dropping {current_idx_type} index on {table_name}")
+                conn.execute(
+                    text(f"""
+                        DROP INDEX IF EXISTS {schema_name}.idx_{table_name.replace(".", "_")}_text_search
+                    """)
+                )
+
+            # Drop existing column if it exists
+            if current_col_type:
+                logger.info(f"Dropping {current_col_type} column on {table_name}")
+                conn.execute(text(f"ALTER TABLE {schema_name}.{table_name} DROP COLUMN IF EXISTS search_vector"))
+
+            # Create new column with appropriate type
+            if text_search_extension == "vchord":
+                logger.info(f"Creating bm25vector column on {table_name}")
+                # Note: vchord_bm25 extension creates types in bm25_catalog schema
+                conn.execute(
+                    text(f"ALTER TABLE {schema_name}.{table_name} ADD COLUMN search_vector bm25_catalog.bm25vector")
+                )
+
+                # Create BM25 index
+                logger.info(f"Creating BM25 index on {table_name}")
+                conn.execute(
+                    text(f"""
+                        CREATE INDEX idx_{table_name.replace(".", "_")}_text_search
+                        ON {schema_name}.{table_name}
+                        USING bm25 (search_vector bm25_catalog.bm25_ops)
+                    """)
+                )
+            else:  # native
+                logger.info(f"Creating tsvector column on {table_name}")
+                # Different GENERATED expression for each table
+                if table_name == "memory_units":
+                    generated_expr = "to_tsvector('english', COALESCE(text, '') || ' ' || COALESCE(context, ''))"
+                else:  # reflections
+                    generated_expr = "to_tsvector('english', COALESCE(name, '') || ' ' || content)"
+
+                conn.execute(
+                    text(f"""
+                        ALTER TABLE {schema_name}.{table_name}
+                        ADD COLUMN search_vector tsvector
+                        GENERATED ALWAYS AS ({generated_expr}) STORED
+                    """)
+                )
+
+                # Create GIN index
+                logger.info(f"Creating GIN index on {table_name}")
+                conn.execute(
+                    text(f"""
+                        CREATE INDEX idx_{table_name.replace(".", "_")}_text_search
+                        ON {schema_name}.{table_name}
+                        USING gin(search_vector)
+                    """)
+                )
+
+        conn.commit()
+        logger.info(f"Successfully migrated text search to {text_search_extension}")
@@ -20,7 +20,8 @@ class RequestContext:
    api_key: str | None = None
    api_key_id: str | None = None  # UUID of the API key used for authentication
    tenant_id: str | None = None  # Tenant identifier (set by extension after auth)
-    internal: bool = False  # True for background/internal operations (not user-visible)
+    internal: bool = False  # True for background/internal operations (skips extension auth)
+    user_initiated: bool = False  # True for async operations that originated from a user request


 from pgvector.sqlalchemy import Vector
@@ -0,0 +1,480 @@
+"""
+OpenTelemetry distributed tracing instrumentation for Hindsight API.
+
+This module provides tracing for:
+- LLM API calls with full prompts/completions following GenAI semantic conventions
+- Token usage and model information
+- Error tracking and finish reasons
+
+Tracing is conditional and disabled by default. When enabled, traces are exported
+to Langfuse (or any OTLP-compatible backend) via OTLP HTTP protocol.
+"""
+
+import json
+import logging
+import time
+from typing import Any, Optional
+
+from opentelemetry import trace
+from opentelemetry.exporter.otlp.proto.http.trace_exporter import OTLPSpanExporter
+from opentelemetry.sdk.resources import Resource
+from opentelemetry.sdk.trace import TracerProvider
+from opentelemetry.sdk.trace.export import BatchSpanProcessor
+from opentelemetry.trace import Status, StatusCode
+
+logger = logging.getLogger(__name__)
+
+
+def _serialize_for_span(obj: Any) -> str:
+    """Serialize an object for span recording, handling Pydantic models."""
+    if isinstance(obj, str):
+        return obj
+    if hasattr(obj, "model_dump_json"):
+        # Pydantic v2 model
+        return obj.model_dump_json()
+    if hasattr(obj, "json"):
+        # Pydantic v1 model
+        return obj.json()
+    if hasattr(obj, "model_dump"):
+        # Pydantic v2 model - convert to dict then json
+        return json.dumps(obj.model_dump())
+    if hasattr(obj, "dict"):
+        # Pydantic v1 model - convert to dict then json
+        return json.dumps(obj.dict())
+    # Fallback to json.dumps for dicts and other types
+    return json.dumps(obj)
+
+
+# No-op tracer for when tracing is disabled
+class NoOpTracer:
+    """No-op tracer that provides the same interface as OpenTelemetry Tracer but does nothing."""
+
+    def start_as_current_span(self, name: str, **kwargs):
+        """Return a no-op context manager that yields a NoOpSpan."""
+        from contextlib import contextmanager
+
+        @contextmanager
+        def noop_span_context():
+            yield NoOpSpan()
+
+        return noop_span_context()
+
+    def start_span(self, name: str, **kwargs):
+        """Return a no-op span."""
+        return NoOpSpan()
+
+
+class NoOpSpan:
+    """No-op span that provides the same interface as OpenTelemetry Span but does nothing."""
+
+    def set_attribute(self, key: str, value: Any) -> None:
+        """No-op."""
+        pass
+
+    def set_status(self, status: Any) -> None:
+        """No-op."""
+        pass
+
+    def record_exception(self, exception: Exception) -> None:
+        """No-op."""
+        pass
+
+    def add_event(self, name: str, attributes: dict | None = None) -> None:
+        """No-op."""
+        pass
+
+    def end(self, end_time: int | None = None) -> None:
+        """No-op."""
+        pass
+
+
+# Global tracer instance
+_tracer: trace.Tracer | NoOpTracer = NoOpTracer()
+_tracing_enabled: bool = False
+
+
+# GenAI semantic convention attribute names (based on v1.37 spec)
+class GenAIAttributes:
+    """GenAI semantic convention attribute names."""
+
+    # Operation and provider
+    OPERATION_NAME = "gen_ai.operation.name"
+    PROVIDER_NAME = "gen_ai.provider.name"
+
+    # Model information
+    REQUEST_MODEL = "gen_ai.request.model"
+    RESPONSE_MODEL = "gen_ai.response.model"
+
+    # Token usage
+    USAGE_INPUT_TOKENS = "gen_ai.usage.input_tokens"
+    USAGE_OUTPUT_TOKENS = "gen_ai.usage.output_tokens"
+
+    # Messages and prompts
+    SYSTEM_INSTRUCTIONS = "gen_ai.system_instructions"
+    INPUT_MESSAGES = "gen_ai.input.messages"
+    OUTPUT_MESSAGES = "gen_ai.output.messages"
+
+    # Response metadata
+    FINISH_REASONS = "gen_ai.response.finish_reasons"
+
+    # Error tracking
+    ERROR_TYPE = "error.type"
+
+
+# Provider name mapping (Hindsight internal -> GenAI semantic convention)
+PROVIDER_NAME_MAPPING = {
+    "openai": "openai",
+    "anthropic": "anthropic",
+    "gemini": "google",
+    "vertexai": "google",
+    "groq": "groq",
+    "ollama": "ollama",
+    "lmstudio": "lmstudio",
+    "openai-codex": "openai",
+    "claude-code": "anthropic",
+    "mock": "mock",
+}
+
+
+def initialize_tracing(
+    service_name: str,
+    endpoint: str,
+    headers: Optional[str] = None,
+    deployment_environment: str = "development",
+) -> None:
+    """
+    Initialize OpenTelemetry tracing with OTLP exporter.
+
+    Args:
+        service_name: Name of the service for resource attributes
+        endpoint: OTLP endpoint URL (e.g., https://cloud.langfuse.com/api/public/otel)
+        headers: Optional headers in format "key1=value1,key2=value2"
+        deployment_environment: Deployment environment (e.g., development, staging, production)
+    """
+    global _tracer, _tracing_enabled
+
+    # Create resource with service information
+    resource = Resource.create(
+        {
+            "service.name": service_name,
+            "service.version": "0.4.8",  # Could import from __version__
+            "deployment.environment.name": deployment_environment,
+        }
+    )
+
+    # Parse headers
+    headers_dict = {}
+    if headers:
+        for pair in headers.split(","):
+            if "=" in pair:
+                key, value = pair.split("=", 1)
+                headers_dict[key.strip()] = value.strip()
+
+    # Create OTLP HTTP exporter
+    # Note: Langfuse expects /v1/traces path appended to base endpoint
+    otlp_endpoint = endpoint if endpoint.endswith("/v1/traces") else f"{endpoint}/v1/traces"
+    otlp_exporter = OTLPSpanExporter(
+        endpoint=otlp_endpoint,
+        headers=headers_dict,
+    )
+
+    # Create tracer provider with batch processor
+    provider = TracerProvider(resource=resource)
+    provider.add_span_processor(BatchSpanProcessor(otlp_exporter))
+
+    # Set global tracer provider
+    trace.set_tracer_provider(provider)
+
+    # Get tracer for this application
+    _tracer = trace.get_tracer(__name__)
+    _tracing_enabled = True
+
+    logger.info(f"Tracing initialized: endpoint={otlp_endpoint}, service={service_name}")
+
+
+def get_tracer() -> trace.Tracer | NoOpTracer:
+    """
+    Get the global tracer instance.
+
+    Returns a no-op tracer if tracing is disabled, so callers don't need to check for None.
+    This improves code readability by allowing direct use without null checks.
+    """
+    return _tracer
+
+
+def create_operation_span(operation: str, bank_id: str | None = None):
+    """
+    Create a parent span for a Hindsight operation (retain, reflect, consolidation, etc.).
+
+    This creates the span hierarchy:
+    - hindsight.{operation} (parent)
+      - chat {model} (child LLM calls)
+
+    Args:
+        operation: Operation name (retain, reflect, consolidation, mental_model_refresh)
+        bank_id: Optional bank ID for context
+
+    Returns:
+        Span context manager
+    """
+    if not _tracing_enabled or _tracer is None:
+        # Return a no-op context manager
+        from contextlib import nullcontext
+
+        return nullcontext()
+
+    span_name = f"hindsight.{operation}"
+    span = _tracer.start_as_current_span(span_name)
+
+    # Add operation-specific attributes
+    if span and hasattr(span, "set_attribute"):
+        span.set_attribute("hindsight.operation", operation)
+        if bank_id:
+            span.set_attribute("hindsight.bank_id", bank_id)
+
+    return span
+
+
+def is_tracing_enabled() -> bool:
+    """Check if tracing is enabled."""
+    return _tracing_enabled
+
+
+# Maximum content length before truncation (to stay within span size limits)
+MAX_CONTENT_LENGTH = 100_000  # characters
+
+
+def _truncate_content(content: str) -> str:
+    """Truncate content if too large for span."""
+    if len(content) > MAX_CONTENT_LENGTH:
+        return content[:MAX_CONTENT_LENGTH] + f"\n\n[TRUNCATED: {len(content) - MAX_CONTENT_LENGTH} chars omitted]"
+    return content
+
+
+class LLMSpanRecorder:
+    """
+    Records OpenTelemetry spans for LLM calls following GenAI semantic conventions.
+    """
+
+    def __init__(self, tracer: trace.Tracer):
+        self.tracer = tracer
+
+    def record_llm_call(
+        self,
+        provider: str,
+        model: str,
+        scope: str,
+        messages: list[dict[str, str]],
+        response_content: Optional[str],
+        input_tokens: int,
+        output_tokens: int,
+        duration: float,
+        finish_reason: Optional[str] = None,
+        error: Optional[Exception] = None,
+        tool_calls: Optional[list[dict[str, Any]]] = None,
+    ) -> None:
+        """
+        Record a completed LLM call as a span with GenAI semantic conventions.
+
+        This creates a span AFTER the call completes, using timestamps to
+        set the correct start/end times. This approach works better with
+        the existing sync metrics recording pattern.
+
+        Args:
+            provider: Hindsight provider name
+            model: Model name
+            scope: Scope identifier (memory, reflect, consolidation, etc.)
+            messages: Input messages (chat history)
+            response_content: Response text from LLM
+            input_tokens: Input token count
+            output_tokens: Output token count
+            duration: Call duration in seconds
+            finish_reason: Reason the model stopped (stop, length, tool_calls, etc.)
+            error: Exception if call failed
+            tool_calls: List of tool calls made (for function calling)
+        """
+        try:
+            # Map provider name to GenAI semantic convention
+            genai_provider = PROVIDER_NAME_MAPPING.get(provider.lower(), provider.lower())
+
+            # Determine operation name based on scope/context
+            operation_name = "chat"  # Default for GenAI semantic conventions
+
+            # Create span name: "hindsight.{scope}" for consistency with parent spans
+            # Model info is available in span attributes (gen_ai.request.model)
+            if scope:
+                span_name = f"hindsight.{scope}"
+            else:
+                # Fallback to chat {model} if no scope provided
+                span_name = f"{operation_name} {model}"
+
+            # Calculate timestamps
+            end_time_ns = time.time_ns()
+            start_time_ns = end_time_ns - int(duration * 1_000_000_000)
+
+            # Create span with explicit timestamps
+            with self.tracer.start_as_current_span(
+                span_name,
+                start_time=start_time_ns,
+                end_on_exit=False,  # We'll set end time manually
+            ) as span:
+                # Set required attributes
+                span.set_attribute(GenAIAttributes.OPERATION_NAME, operation_name)
+                span.set_attribute(GenAIAttributes.PROVIDER_NAME, genai_provider)
+                span.set_attribute(GenAIAttributes.REQUEST_MODEL, model)
+                span.set_attribute(GenAIAttributes.RESPONSE_MODEL, model)
+                span.set_attribute(GenAIAttributes.USAGE_INPUT_TOKENS, input_tokens)
+                span.set_attribute(GenAIAttributes.USAGE_OUTPUT_TOKENS, output_tokens)
+
+                # Add custom attributes for Hindsight context
+                span.set_attribute("hindsight.scope", scope)
+                span.set_attribute("hindsight.provider.internal", provider)
+
+                # Add tool call information if present
+                if tool_calls:
+                    span.set_attribute("gen_ai.tool_calls.count", len(tool_calls))
+                    # Add tool names as comma-separated list
+                    tool_names = [tc.get("name", "") for tc in tool_calls]
+                    span.set_attribute("gen_ai.tool_calls.names", ",".join(tool_names))
+
+                # Format messages for GenAI conventions (as JSON)
+                input_messages_json = self._format_messages(messages)
+                output_messages_json = self._format_output(response_content, finish_reason)
+
+                # Extract system instructions if present
+                system_instructions = self._extract_system_instructions(messages)
+
+                # Add event with prompts/completions following v1.37 conventions
+                event_attrs = {}
+                if input_messages_json:
+                    event_attrs[GenAIAttributes.INPUT_MESSAGES] = input_messages_json
+                if output_messages_json:
+                    event_attrs[GenAIAttributes.OUTPUT_MESSAGES] = output_messages_json
+                if system_instructions:
+                    event_attrs[GenAIAttributes.SYSTEM_INSTRUCTIONS] = system_instructions
+                if finish_reason:
+                    event_attrs[GenAIAttributes.FINISH_REASONS] = json.dumps([finish_reason])
+
+                span.add_event(
+                    "gen_ai.client.inference.operation.details",
+                    attributes=event_attrs,
+                )
+
+                # Add individual tool call events with details
+                if tool_calls:
+                    for i, tc in enumerate(tool_calls):
+                        tool_event_attrs = {
+                            "tool.name": tc.get("name", ""),
+                            "tool.id": tc.get("id", ""),
+                            "tool.arguments": json.dumps(tc.get("arguments", {})),
+                        }
+                        span.add_event(f"gen_ai.tool_call.{i}", attributes=tool_event_attrs)
+
+                # Handle errors
+                if error:
+                    span.set_status(Status(StatusCode.ERROR, str(error)))
+                    span.set_attribute(GenAIAttributes.ERROR_TYPE, type(error).__name__)
+                    span.record_exception(error)
+                else:
+                    span.set_status(Status(StatusCode.OK))
+
+                # Set end time
+                span.end(end_time=end_time_ns)
+
+        except Exception as e:
+            # Don't let tracing errors break LLM calls
+            logger.error(f"Failed to record LLM span: {e}", exc_info=True)
+
+    def _format_messages(self, messages: list[dict[str, str]]) -> str:
+        """
+        Format messages into GenAI semantic convention format (JSON array).
+
+        Returns JSON string representation of message array.
+        """
+        try:
+            formatted = []
+            for msg in messages:
+                content = msg.get("content", "")
+                # Truncate if needed
+                if isinstance(content, str):
+                    content = _truncate_content(content)
+
+                formatted.append(
+                    {
+                        "role": msg.get("role", "user"),
+                        "content": content,
+                    }
+                )
+
+            return json.dumps(formatted)
+        except Exception as e:
+            logger.warning(f"Failed to format input messages: {e}")
+            return "[]"
+
+    def _format_output(
+        self,
+        content: Optional[str],
+        finish_reason: Optional[str],
+    ) -> str:
+        """Format output message into GenAI semantic convention format."""
+        try:
+            if content is None:
+                return "[]"
+
+            # Truncate if needed
+            if isinstance(content, str):
+                content = _truncate_content(content)
+
+            return json.dumps(
+                [
+                    {
+                        "role": "assistant",
+                        "content": content,
+                    }
+                ]
+            )
+        except Exception as e:
+            logger.warning(f"Failed to format output message: {e}")
+            return "[]"
+
+    def _extract_system_instructions(self, messages: list[dict[str, str]]) -> Optional[str]:
+        """Extract system instructions from messages if present."""
+        try:
+            for msg in messages:
+                if msg.get("role") == "system":
+                    content = msg.get("content", "")
+                    if isinstance(content, str):
+                        return _truncate_content(content)
+                    return str(content)
+        except Exception as e:
+            logger.warning(f"Failed to extract system instructions: {e}")
+        return None
+
+
+class NoOpLLMSpanRecorder:
+    """No-op span recorder for when tracing is disabled."""
+
+    def record_llm_call(self, **kwargs) -> None:
+        """No-op."""
+        pass
+
+
+# Global span recorder instance
+_span_recorder: Optional[LLMSpanRecorder] = None
+
+
+def get_span_recorder() -> LLMSpanRecorder | NoOpLLMSpanRecorder:
+    """Get the global span recorder (NoOp if tracing disabled)."""
+    if _span_recorder is None:
+        return NoOpLLMSpanRecorder()
+    return _span_recorder
+
+
+def create_span_recorder() -> LLMSpanRecorder:
+    """Create and set the global span recorder."""
+    global _span_recorder
+    tracer = get_tracer()
+    if tracer is None:
+        raise RuntimeError("Tracing not initialized. Call initialize_tracing() first.")
+    _span_recorder = LLMSpanRecorder(tracer)
+    return _span_recorder
@@ -0,0 +1,13 @@
+from urllib.parse import urlparse, urlunparse
+
+
+def mask_network_location(url):
+    if not url:
+        return url
+    parsed_url = urlparse(url)
+    masked_network_location = parsed_url.hostname or ""
+    if parsed_url.port:
+        masked_network_location += f":{parsed_url.port}"
+    if parsed_url.username or parsed_url.password:
+        masked_network_location = f"***:***@{masked_network_location}"
+    return urlunparse(parsed_url._replace(netloc=masked_network_location))
@@ -176,7 +176,7 @@ def main():
        nonlocal memory, poller
        import uvicorn

-        from ..extensions import TenantExtension, load_extension
+        from ..extensions import OperationValidatorExtension, TenantExtension, load_extension

        # Load tenant extension BEFORE creating MemoryEngine so it can
        # set correct schema context during task execution. Without this,
@@ -184,6 +184,12 @@ def main():
        # causing worker writes to land in the wrong schema.
        tenant_extension = load_extension("TENANT", TenantExtension)

+        # Load operation validator so workers can record usage metering
+        # for async operations (e.g. refresh_mental_model after consolidation)
+        operation_validator = load_extension("OPERATION_VALIDATOR", OperationValidatorExtension)
+        if operation_validator:
+            logger.info(f"Loaded operation validator: {operation_validator.__class__.__name__}")
+
        # Initialize MemoryEngine
        # Workers use SyncTaskBackend because they execute tasks directly,
        # they don't need to store tasks (they poll from DB)
@@ -191,6 +197,7 @@ def main():
            run_migrations=False,  # Workers don't run migrations
            task_backend=SyncTaskBackend(),
            tenant_extension=tenant_extension,
+            operation_validator=operation_validator,
        )

        await memory.initialize()
@@ -222,15 +229,30 @@ def main():
        # Create the HTTP app for metrics/health
        app = create_worker_app(poller, memory)

-        # Setup signal handlers for graceful shutdown
+        # Setup signal handlers for graceful shutdown using asyncio
        shutdown_requested = asyncio.Event()
+        force_exit = False

-        def signal_handler(signum, frame):
-            print(f"\nReceived signal {signum}, initiating graceful shutdown...")
-            shutdown_requested.set()
+        loop = asyncio.get_event_loop()

-        signal.signal(signal.SIGINT, signal_handler)
-        signal.signal(signal.SIGTERM, signal_handler)
+        def signal_handler():
+            nonlocal force_exit
+            if shutdown_requested.is_set():
+                # Second signal = force exit
+                print("\nReceived second signal, forcing immediate exit...")
+                force_exit = True
+                # Restore default handler so third signal kills process
+                loop.remove_signal_handler(signal.SIGINT)
+                loop.remove_signal_handler(signal.SIGTERM)
+                sys.exit(1)
+            else:
+                print("\nReceived shutdown signal, initiating graceful shutdown...")
+                print("(Press Ctrl+C again to force immediate exit)")
+                shutdown_requested.set()
+
+        # Use asyncio's signal handlers which work properly with the event loop
+        loop.add_signal_handler(signal.SIGINT, signal_handler)
+        loop.add_signal_handler(signal.SIGTERM, signal_handler)

        # Create uvicorn config and server
        uvicorn_config = uvicorn.Config(
@@ -249,7 +271,10 @@ def main():
        print(f"Worker started. Metrics available at http://{args.http_host}:{args.http_port}/metrics")

        # Wait for shutdown signal
-        await shutdown_requested.wait()
+        try:
+            await shutdown_requested.wait()
+        except KeyboardInterrupt:
+            print("\nReceived interrupt, initiating graceful shutdown...")

        # Graceful shutdown
        print("Shutting down HTTP server...")
@@ -72,9 +72,9 @@ class WorkerPoller:
            executor: Async function to execute tasks (typically MemoryEngine.execute_task)
            poll_interval_ms: Interval between polls when no tasks found (milliseconds)
            max_retries: Maximum retry attempts before marking task as failed
-            schema: Database schema for single-tenant support (ignored if tenant_extension is set)
-            tenant_extension: Extension for dynamic multi-tenant discovery. If set, list_tenants()
-                            is called on each poll cycle to discover schemas dynamically.
+            schema: Database schema for single-tenant support (deprecated, use tenant_extension)
+            tenant_extension: Extension for dynamic multi-tenant discovery. If None, creates a
+                            DefaultTenantExtension with the configured schema.
            max_slots: Maximum concurrent tasks per worker
            consolidation_max_slots: Maximum concurrent consolidation tasks per worker
        """
@@ -84,6 +84,13 @@ class WorkerPoller:
        self._poll_interval_ms = poll_interval_ms
        self._max_retries = max_retries
        self._schema = schema
+        # Always set tenant extension (use DefaultTenantExtension if none provided)
+        if tenant_extension is None:
+            from ..extensions.builtin.tenant import DefaultTenantExtension
+
+            # Pass schema parameter to DefaultTenantExtension if explicitly provided
+            config = {"schema": schema} if schema else {}
+            tenant_extension = DefaultTenantExtension(config=config)
        self._tenant_extension = tenant_extension
        self._max_slots = max_slots
        self._consolidation_max_slots = consolidation_max_slots
@@ -100,14 +107,11 @@ class WorkerPoller:

    async def _get_schemas(self) -> list[str | None]:
        """Get list of schemas to poll. Returns [None] for default schema (no prefix)."""
-        if self._tenant_extension is not None:
-            from ..config import DEFAULT_DATABASE_SCHEMA
+        from ..config import DEFAULT_DATABASE_SCHEMA

-            tenants = await self._tenant_extension.list_tenants()
-            # Convert default schema to None for SQL compatibility (no prefix), keep others as-is
-            return [t.schema if t.schema != DEFAULT_DATABASE_SCHEMA else None for t in tenants]
-        # Single schema mode
-        return [self._schema]
+        tenants = await self._tenant_extension.list_tenants()
+        # Convert default schema to None for SQL compatibility (no prefix), keep others as-is
+        return [t.schema if t.schema != DEFAULT_DATABASE_SCHEMA else None for t in tenants]

    async def _get_available_slots(self) -> tuple[int, int]:
        """
@@ -4,7 +4,7 @@ build-backend = "hatchling.build"

 [project]
 name = "hindsight-api"
-version = "0.4.6"
+version = "0.4.10"
 description = "Hindsight: Agent Memory That Works Like Human Memory"
 readme = "README.md"
 requires-python = ">=3.11"
@@ -25,13 +25,16 @@ dependencies = [
    "psycopg2-binary>=2.9.11",
    "tiktoken>=0.12.0",
    "httpx>=0.27.0",
-    "fastmcp>=2.14.0",  # CVE-2025-66416
+    "PyJWT[crypto]>=2.8.0",
+    "fastmcp>=2.14.0", # CVE-2025-66416
    "pg0-embedded>=0.11.0",
    "python-dateutil>=2.8.0",
    "opentelemetry-api>=1.20.0",
    "opentelemetry-sdk>=1.20.0",
    "opentelemetry-instrumentation-fastapi>=0.41b0",
    "opentelemetry-exporter-prometheus>=0.41b0",
+    "opentelemetry-exporter-otlp-proto-http>=1.20.0",
+    "opentelemetry-semantic-conventions>=0.41b0",
    "dateparser>=1.2.2",
    "google-genai>=1.0.0",
    "google-auth>=2.0.0",
@@ -41,16 +44,17 @@ dependencies = [
    "flashrank>=0.2.0",
    # Local ML models for embeddings/reranking - can be excluded in Docker with INCLUDE_LOCAL_MODELS=false
    "sentence-transformers>=3.3.0",
-    "transformers>=4.53.0",  # Security fixes for ReDoS vulnerabilities
-    "torch>=2.6.0",  # CVE fix for remote code execution
+    "transformers>=4.53.0", # Security fixes for ReDoS vulnerabilities
+    "torch>=2.6.0", # CVE fix for remote code execution
    "uvloop>=0.22.1",
    # Transitive dependency security fixes
-    "pyasn1>=0.6.2",  # DoS vulnerability fix
-    "urllib3>=2.6.3",  # Decompression-bomb safeguards bypass fix
-    "langchain-core>=1.2.5",  # Serialization injection vulnerability fix
-    "filelock>=3.20.1",  # TOCTOU race condition fix
-    "authlib>=1.6.6",  # Account takeover vulnerability fix
-    "aiohttp>=3.13.3",  # Multiple DoS vulnerabilities
+    "pyasn1>=0.6.2", # DoS vulnerability fix
+    "urllib3>=2.6.3", # Decompression-bomb safeguards bypass fix
+    "langchain-core>=1.2.5", # Serialization injection vulnerability fix
+    "filelock>=3.20.1", # TOCTOU race condition fix
+    "authlib>=1.6.6", # Account takeover vulnerability fix
+    "aiohttp>=3.13.3", # Multiple DoS vulnerabilities
+    "claude-agent-sdk>=0.1.27",
 ]

 [project.optional-dependencies]
@@ -220,3 +220,34 @@ async def memory(pg0_db_url, embeddings, cross_encoder, query_analyzer):
            await mem.close()
    except Exception:
        pass
+
+
+@pytest_asyncio.fixture(scope="function")
+async def memory_no_llm_verify(pg0_db_url, embeddings, cross_encoder, query_analyzer):
+    """
+    Provide a MemoryEngine instance that skips LLM connection verification.
+
+    This fixture is useful for tests that override the LLM configuration
+    after initialization (e.g., to test specific providers).
+    """
+    mem = MemoryEngine(
+        db_url=pg0_db_url,
+        memory_llm_provider="mock",  # Use mock provider as placeholder
+        memory_llm_api_key="",
+        memory_llm_model="mock",
+        embeddings=embeddings,
+        cross_encoder=cross_encoder,
+        query_analyzer=query_analyzer,
+        pool_min_size=1,
+        pool_max_size=5,
+        run_migrations=False,
+        task_backend=SyncTaskBackend(),
+        skip_llm_verification=True,  # Skip verification - will be overridden by test
+    )
+    await mem.initialize()
+    yield mem
+    try:
+        if mem._pool and not mem._pool._closing:
+            await mem.close()
+    except Exception:
+        pass
@@ -0,0 +1,189 @@
+"""
+Integration test for API base path support.
+
+Tests that the API works correctly when deployed with a base path (e.g., /hindsight)
+for reverse proxy deployments.
+"""
+import os
+import pytest
+import pytest_asyncio
+import httpx
+from hindsight_api.api import create_app
+from hindsight_api.config import clear_config_cache
+
+
+@pytest_asyncio.fixture
+async def api_client_with_base_path(memory):
+    """Create an async test client for the FastAPI app with a base path."""
+    # Set base path in environment
+    base_path = "/hindsight"
+    os.environ["HINDSIGHT_API_BASE_PATH"] = base_path
+
+    # Clear config cache to force reload with new base_path
+    clear_config_cache()
+
+    # Memory is already initialized by the conftest fixture (with migrations)
+    app = create_app(memory, initialize_memory=False)
+
+    # Use base_url with base path
+    transport = httpx.ASGITransport(app=app)
+    async with httpx.AsyncClient(
+        transport=transport,
+        base_url=f"http://test{base_path}"
+    ) as client:
+        yield client
+
+    # Cleanup: unset base path
+    os.environ.pop("HINDSIGHT_API_BASE_PATH", None)
+    clear_config_cache()
+
+
+@pytest_asyncio.fixture
+async def api_client_without_base_path(memory):
+    """Create an async test client for the FastAPI app without a base path (root)."""
+    # Ensure no base path is set
+    os.environ.pop("HINDSIGHT_API_BASE_PATH", None)
+    clear_config_cache()
+
+    app = create_app(memory, initialize_memory=False)
+    transport = httpx.ASGITransport(app=app)
+    async with httpx.AsyncClient(transport=transport, base_url="http://test") as client:
+        yield client
+
+
+@pytest.mark.asyncio
+async def test_base_path_health_endpoint(api_client_with_base_path):
+    """Test that health endpoint works with base path."""
+    # With base path set to /hindsight, health should be at /hindsight/health
+    # But since our client base_url is already http://test/hindsight, we request /health
+    response = await api_client_with_base_path.get("/health")
+    assert response.status_code == 200
+    data = response.json()
+    assert "status" in data
+    assert data["status"] in ["ok", "healthy"]  # Accept both formats
+
+
+@pytest.mark.asyncio
+async def test_base_path_banks_endpoint(api_client_with_base_path):
+    """Test that banks endpoint works with base path."""
+    response = await api_client_with_base_path.get("/v1/default/banks")
+    assert response.status_code == 200
+    data = response.json()
+    assert "banks" in data
+
+
+@pytest.mark.asyncio
+async def test_base_path_openapi_schema(api_client_with_base_path):
+    """Test that OpenAPI schema includes correct base path in servers."""
+    response = await api_client_with_base_path.get("/openapi.json")
+    assert response.status_code == 200
+    openapi_schema = response.json()
+
+    # Check that servers array includes base path
+    assert "servers" in openapi_schema
+    servers = openapi_schema["servers"]
+    assert len(servers) > 0
+    # FastAPI should set server URL to the root_path
+    assert servers[0]["url"] == "/hindsight"
+
+
+@pytest.mark.asyncio
+async def test_base_path_docs_redirect(api_client_with_base_path):
+    """Test that /docs redirects correctly with base path."""
+    # FastAPI docs endpoint should work
+    response = await api_client_with_base_path.get("/docs", follow_redirects=False)
+    # Should either return 200 (direct) or 307 (redirect to trailing slash)
+    assert response.status_code in [200, 307]
+
+
+@pytest.mark.asyncio
+async def test_base_path_metrics(api_client_with_base_path):
+    """Test that metrics endpoint works with base path."""
+    response = await api_client_with_base_path.get("/metrics")
+    assert response.status_code == 200
+    # Metrics should be in Prometheus format
+    assert "# HELP" in response.text or "# TYPE" in response.text
+
+
+@pytest.mark.asyncio
+async def test_base_path_full_workflow(api_client_with_base_path):
+    """
+    Test a full retain/recall workflow with base path.
+
+    This ensures that all memory operations work correctly when the API
+    is deployed with a base path.
+    """
+    bank_id = "test_base_path_bank"
+
+    # 1. Create/get bank
+    response = await api_client_with_base_path.get(f"/v1/default/banks/{bank_id}/profile")
+    assert response.status_code == 200
+
+    # 2. Store a memory
+    response = await api_client_with_base_path.post(
+        f"/v1/default/banks/{bank_id}/memories",
+        json={
+            "items": [
+                {
+                    "content": "The API supports base path deployment for reverse proxy use cases.",
+                    "context": "testing base path feature"
+                }
+            ]
+        }
+    )
+    assert response.status_code == 200
+    result = response.json()
+    assert result["success"] is True
+
+    # 3. Recall the memory
+    response = await api_client_with_base_path.post(
+        f"/v1/default/banks/{bank_id}/memories/recall",
+        json={
+            "query": "base path support"
+        }
+    )
+    assert response.status_code == 200
+    recall_result = response.json()
+    # API returns "results" not "memories"
+    assert "results" in recall_result
+    assert len(recall_result["results"]) > 0
+
+
+@pytest.mark.asyncio
+async def test_without_base_path_still_works(api_client_without_base_path):
+    """
+    Regression test: ensure default behavior (no base path) still works.
+
+    This test verifies that when HINDSIGHT_API_BASE_PATH is not set,
+    the API works at the root path as before.
+    """
+    # Health check at root
+    response = await api_client_without_base_path.get("/health")
+    assert response.status_code == 200
+
+    # Banks endpoint at root
+    response = await api_client_without_base_path.get("/v1/default/banks")
+    assert response.status_code == 200
+
+    # OpenAPI schema should have empty or "/" server path
+    response = await api_client_without_base_path.get("/openapi.json")
+    assert response.status_code == 200
+    openapi_schema = response.json()
+    servers = openapi_schema.get("servers", [])
+    if servers:
+        # Server URL should be empty string (root) or "/"
+        assert servers[0]["url"] in ["", "/"]
+
+
+@pytest.mark.skip(reason="MCP endpoint routing with base path needs investigation")
+@pytest.mark.asyncio
+async def test_base_path_mcp_endpoint(api_client_with_base_path):
+    """Test that MCP endpoint is accessible with base path."""
+    bank_id = "test_mcp_bank"
+
+    # MCP endpoint should be mounted at /mcp/{bank_id}/
+    # The MCP server uses a different protocol, so just check the root exists
+    response = await api_client_with_base_path.get(f"/mcp/{bank_id}/")
+    # MCP may return various status codes, but should not be 404 (not found)
+    # Accept 405 (method not allowed), 400 (bad request), etc.
+    assert response.status_code != 404, "MCP endpoint should exist"
@@ -12,6 +12,7 @@ from datetime import datetime
 import pytest

 from hindsight_api import LLMConfig
+from hindsight_api.config import _get_raw_config
 from hindsight_api.engine.retain.fact_extraction import extract_facts_from_text


@@ -44,6 +45,7 @@ class TestCausalRelationsValidation:
            context=context,
            llm_config=llm_config,
            agent_name="TestUser",
+            config=_get_raw_config(),
        )

        assert len(facts) > 0, "Should extract at least one fact"
@@ -88,6 +90,7 @@ class TestCausalRelationsValidation:
            context=context,
            llm_config=llm_config,
            agent_name="TestUser",
+            config=_get_raw_config(),
        )

        assert len(facts) > 0, "Should extract at least one fact"
@@ -124,6 +127,7 @@ class TestCausalRelationsValidation:
            context=context,
            llm_config=llm_config,
            agent_name="TestUser",
+            config=_get_raw_config(),
        )

        assert len(facts) > 0, "Should extract facts about the causal chain"
@@ -173,6 +177,7 @@ class TestCausalRelationsValidation:
            context=context,
            llm_config=llm_config,
            agent_name="TestUser",
+            config=_get_raw_config(),
        )

        assert len(facts) > 0, "Should extract facts"
@@ -209,6 +214,7 @@ class TestCausalRelationsValidation:
            context=context,
            llm_config=llm_config,
            agent_name="TestUser",
+            config=_get_raw_config(),
        )

        # Verify relation types are all backward-looking
@@ -10,6 +10,7 @@ from datetime import datetime
 import pytest

 from hindsight_api import LLMConfig
+from hindsight_api.config import _get_raw_config
 from hindsight_api.engine.retain.fact_extraction import extract_facts_from_text


@@ -37,7 +38,8 @@ After searching for weeks, I finally found a cheaper apartment in Brooklyn.
        llm_config = LLMConfig.for_memory()

        facts, _, _ = await extract_facts_from_text(
-            text=text, event_date=datetime(2024, 3, 15), context=context, llm_config=llm_config, agent_name="TestUser"
+            text=text, event_date=datetime(2024, 3, 15), context=context, llm_config=llm_config, agent_name="TestUser",
+            config=_get_raw_config(),
        )

        assert len(facts) >= 3, f"Should extract at least 3 facts from the causal chain. Got {len(facts)}"
@@ -106,7 +108,8 @@ The renovation took three months and cost $15,000.
        llm_config = LLMConfig.for_memory()

        facts, _, _ = await extract_facts_from_text(
-            text=text, event_date=datetime(2024, 6, 1), context=context, llm_config=llm_config, agent_name="TestUser"
+            text=text, event_date=datetime(2024, 6, 1), context=context, llm_config=llm_config, agent_name="TestUser",
+            config=_get_raw_config(),
        )

        assert len(facts) >= 4, f"Should extract at least 4 facts. Got {len(facts)}"
@@ -136,7 +139,8 @@ Machine learning fascinated me so much that I changed my career to data science.
        llm_config = LLMConfig.for_memory()

        facts, _, _ = await extract_facts_from_text(
-            text=text, event_date=datetime(2024, 1, 1), context=context, llm_config=llm_config, agent_name="TestUser"
+            text=text, event_date=datetime(2024, 1, 1), context=context, llm_config=llm_config, agent_name="TestUser",
+            config=_get_raw_config(),
        )

        # Check no fact references itself
@@ -163,7 +167,8 @@ The new role enabled me to lead a team of engineers.
        llm_config = LLMConfig.for_memory()

        facts, _, _ = await extract_facts_from_text(
-            text=text, event_date=datetime(2024, 2, 15), context=context, llm_config=llm_config, agent_name="TestUser"
+            text=text, event_date=datetime(2024, 2, 15), context=context, llm_config=llm_config, agent_name="TestUser",
+            config=_get_raw_config(),
        )

        # Validate all indices (must reference PREVIOUS facts only)
@@ -190,7 +195,8 @@ Reduced spending somewhat affected local businesses.
        llm_config = LLMConfig.for_memory()

        facts, _, _ = await extract_facts_from_text(
-            text=text, event_date=datetime(2024, 4, 1), context=context, llm_config=llm_config, agent_name="TestUser"
+            text=text, event_date=datetime(2024, 4, 1), context=context, llm_config=llm_config, agent_name="TestUser",
+            config=_get_raw_config(),
        )

        for i, fact in enumerate(facts):
@@ -0,0 +1,109 @@
+"""
+Tests for configuration validation.
+
+Verifies that config validation catches invalid parameter combinations.
+"""
+
+import os
+
+import pytest
+
+
+@pytest.fixture(autouse=True)
+def setup_test_env():
+    """Set up environment for each test, restoring original values after."""
+    from hindsight_api.config import clear_config_cache
+
+    # Save original environment values
+    env_vars_to_save = [
+        "HINDSIGHT_API_RETAIN_MAX_COMPLETION_TOKENS",
+        "HINDSIGHT_API_RETAIN_CHUNK_SIZE",
+        "HINDSIGHT_API_LLM_PROVIDER",
+        "HINDSIGHT_API_LLM_MODEL",
+    ]
+
+    # Save original values
+    original_values = {}
+    for key in env_vars_to_save:
+        original_values[key] = os.environ.get(key)
+
+    clear_config_cache()
+
+    yield
+
+    # Restore original environment
+    for key, original_value in original_values.items():
+        if original_value is None:
+            os.environ.pop(key, None)
+        else:
+            os.environ[key] = original_value
+
+    clear_config_cache()
+
+
+def test_retain_max_completion_tokens_must_be_greater_than_chunk_size():
+    """Test that RETAIN_MAX_COMPLETION_TOKENS > RETAIN_CHUNK_SIZE validation works."""
+    from hindsight_api.config import HindsightConfig
+
+    # Set invalid config: max_completion_tokens <= chunk_size
+    os.environ["HINDSIGHT_API_RETAIN_MAX_COMPLETION_TOKENS"] = "1000"
+    os.environ["HINDSIGHT_API_RETAIN_CHUNK_SIZE"] = "2000"
+    os.environ["HINDSIGHT_API_LLM_PROVIDER"] = "mock"
+
+    # Should raise ValueError with helpful message
+    with pytest.raises(ValueError) as exc_info:
+        HindsightConfig.from_env()
+
+    error_message = str(exc_info.value)
+
+    # Verify error message contains helpful information
+    assert "HINDSIGHT_API_RETAIN_MAX_COMPLETION_TOKENS" in error_message
+    assert "1000" in error_message
+    assert "HINDSIGHT_API_RETAIN_CHUNK_SIZE" in error_message
+    assert "2000" in error_message
+    assert "must be greater than" in error_message
+    assert "You have two options to fix this:" in error_message
+    assert "Increase HINDSIGHT_API_RETAIN_MAX_COMPLETION_TOKENS" in error_message
+    assert "Use a model that supports" in error_message
+
+
+def test_retain_max_completion_tokens_equal_to_chunk_size_fails():
+    """Test that RETAIN_MAX_COMPLETION_TOKENS == RETAIN_CHUNK_SIZE also fails."""
+    from hindsight_api.config import HindsightConfig
+
+    # Set invalid config: max_completion_tokens == chunk_size
+    os.environ["HINDSIGHT_API_RETAIN_MAX_COMPLETION_TOKENS"] = "3000"
+    os.environ["HINDSIGHT_API_RETAIN_CHUNK_SIZE"] = "3000"
+    os.environ["HINDSIGHT_API_LLM_PROVIDER"] = "mock"
+
+    # Should raise ValueError
+    with pytest.raises(ValueError) as exc_info:
+        HindsightConfig.from_env()
+
+    error_message = str(exc_info.value)
+    assert "must be greater than" in error_message
+
+
+def test_valid_retain_config_succeeds():
+    """Test that valid config with max_completion_tokens > chunk_size works."""
+    from hindsight_api.config import HindsightConfig
+
+    # Set valid config: max_completion_tokens > chunk_size
+    os.environ["HINDSIGHT_API_RETAIN_MAX_COMPLETION_TOKENS"] = "64000"
+    os.environ["HINDSIGHT_API_RETAIN_CHUNK_SIZE"] = "3000"
+    os.environ["HINDSIGHT_API_LLM_PROVIDER"] = "mock"
+
+    # Should not raise
+    config = HindsightConfig.from_env()
+    assert config.retain_max_completion_tokens == 64000
+    assert config.retain_chunk_size == 3000
+
+
+# Note: The BadRequestError wrapping is implemented in fact_extraction.py
+# but requires a complex integration test setup. The functionality is
+# straightforward: when a BadRequestError containing keywords like
+# "max_tokens", "max_completion_tokens", or "maximum context" is caught,
+# it's wrapped in a ValueError with helpful guidance.
+#
+# The config validation tests above ensure users get early feedback
+# about invalid configurations before runtime errors occur.
@@ -21,9 +21,9 @@ from hindsight_api.engine.reflect.tools import (
@pytest.fixture(autouse=True)
 def enable_observations():
    """Enable observations for all tests in this module."""
-    from hindsight_api.config import get_config
+    from hindsight_api.config import _get_raw_config

-    config = get_config()
+    config = _get_raw_config()
    original_value = config.enable_observations
    config.enable_observations = True
    yield
@@ -563,25 +563,26 @@ class TestConsolidationDisabled:
        self, memory: MemoryEngine, request_context
    ):
        """Test that consolidation returns disabled status when enable_observations is False."""
-        from unittest.mock import patch
-
        bank_id = f"test-consolidation-disabled-{uuid.uuid4().hex[:8]}"

        # Create the bank
        await memory.get_bank_profile(bank_id=bank_id, request_context=request_context)

-        # Disable observations via config
-        with patch("hindsight_api.config.get_config") as mock_config:
-            mock_config.return_value.enable_observations = False
+        # Disable observations for this bank via bank config
+        await memory._config_resolver.update_bank_config(
+            bank_id=bank_id,
+            updates={"enable_observations": False},
+            context=request_context,
+        )

-            result = await run_consolidation_job(
-                memory_engine=memory,
-                bank_id=bank_id,
-                request_context=request_context,
-            )
+        result = await run_consolidation_job(
+            memory_engine=memory,
+            bank_id=bank_id,
+            request_context=request_context,
+        )

-            assert result["status"] == "disabled"
-            assert result["bank_id"] == bank_id
+        assert result["status"] == "disabled"
+        assert result["bank_id"] == bank_id

        # Cleanup
        await memory.delete_bank(bank_id, request_context=request_context)
@@ -353,6 +353,14 @@ class TestOperationHooksParameters:
        assert post_result.error is None
        assert post_result.unit_ids == result  # Should match the return value

+        # Verify actual LLM token usage is populated
+        assert post_result.llm_input_tokens is not None
+        assert post_result.llm_input_tokens > 0
+        assert post_result.llm_output_tokens is not None
+        assert post_result.llm_output_tokens > 0
+        assert post_result.llm_total_tokens is not None
+        assert post_result.llm_total_tokens == post_result.llm_input_tokens + post_result.llm_output_tokens
+
    @pytest.mark.asyncio
    async def test_recall_pre_hook_receives_all_parameters(self, memory_with_tracking_validator):
        """Pre-recall hook receives all user-provided parameters."""
@@ -8,7 +8,7 @@ from datetime import datetime

 import pytest

-from hindsight_api.config import get_config, clear_config_cache
+from hindsight_api.config import get_config, clear_config_cache, _get_raw_config
 from hindsight_api.engine.llm_wrapper import LLMConfig
 from hindsight_api.engine.retain.fact_extraction import extract_facts_from_text

@@ -58,6 +58,7 @@ async def test_fact_extraction_basic_analysis(llm_config):
        llm_config=llm_config,
        agent_name="test-agent",
        context="Friday Standup meeting",
+        config=_get_raw_config(),
    )

    duration = time.time() - start_time
@@ -11,6 +11,7 @@ from datetime import datetime
 import pytest

 from hindsight_api import LLMConfig
+from hindsight_api.config import _get_raw_config
 from hindsight_api.engine.retain.fact_extraction import extract_facts_from_text


@@ -44,7 +45,8 @@ I ran into my neighbor Sarah who mentioned she's planning a trip to Italy next m
            event_date=datetime(2024, 6, 15),
            context=context,
            llm_config=llm_config,
-            agent_name="TestUser"
+            agent_name="TestUser",
+            config=_get_raw_config(),
        )

        input_length = len(text)
@@ -88,7 +90,8 @@ User: Perfect, I'll make a reservation for Saturday at 7pm.
            event_date=datetime(2024, 6, 15),
            context=context,
            llm_config=llm_config,
-            agent_name="TestUser"
+            agent_name="TestUser",
+            config=_get_raw_config(),
        )

        input_length = len(text)
@@ -144,7 +147,8 @@ I edited about 20 photos from my recent trip to the mountains.
            event_date=datetime(2024, 4, 15),
            context=context,
            llm_config=llm_config,
-            agent_name="TestUser"
+            agent_name="TestUser",
+            config=_get_raw_config(),
        )

        input_length = len(text)
@@ -208,7 +212,8 @@ I edited about 20 photos from my recent trip to the mountains.
            event_date=datetime(2023, 5, 8),  # Date from locomo dataset
            context=context,
            llm_config=llm_config,
-            agent_name=data["conversation"]["speaker_a"]
+            agent_name=data["conversation"]["speaker_a"],
+            config=_get_raw_config(),
        )

        # Calculate ratios
@@ -269,7 +274,8 @@ I'm planning to visit Japan next year.
            event_date=datetime(2024, 6, 15),
            context=context,
            llm_config=llm_config,
-            agent_name="TestUser"
+            agent_name="TestUser",
+            config=_get_raw_config(),
        )

        # Count approximate number of statements (sentences)
@@ -17,6 +17,7 @@ from datetime import UTC, datetime
 import pytest

 from hindsight_api import LLMConfig
+from hindsight_api.config import _get_raw_config
 from hindsight_api.engine.retain.fact_extraction import extract_facts_from_text

 # =============================================================================
@@ -48,7 +49,8 @@ Marcus felt anxious about the upcoming interview.
            event_date=datetime(2024, 11, 13),
            context=context,
            llm_config=llm_config,
-            agent_name="TestUser"
+            agent_name="TestUser",
+            config=_get_raw_config(),
        )

        assert len(facts) > 0, "Should extract at least one fact"
@@ -80,7 +82,8 @@ The music was so loud I could barely hear myself think.
            event_date=datetime(2024, 11, 13),
            context=context,
            llm_config=llm_config,
-            agent_name="TestUser"
+            agent_name="TestUser",
+            config=_get_raw_config(),
        )

        assert len(facts) > 0, "Should extract at least one fact"
@@ -113,7 +116,8 @@ Maybe we should reconsider the timeline.
            event_date=datetime(2024, 11, 13),
            context=context,
            llm_config=llm_config,
-            agent_name="TestUser"
+            agent_name="TestUser",
+            config=_get_raw_config(),
        )

        assert len(facts) > 0, "Should extract at least one fact"
@@ -146,7 +150,8 @@ I'm unable to attend the conference due to scheduling conflicts.
            event_date=datetime(2024, 11, 13),
            context=context,
            llm_config=llm_config,
-            agent_name="TestUser"
+            agent_name="TestUser",
+            config=_get_raw_config(),
        )

        assert len(facts) > 0, "Should extract at least one fact"
@@ -178,7 +183,8 @@ Unlike last year, we're ahead of schedule.
            event_date=datetime(2024, 11, 13),
            context=context,
            llm_config=llm_config,
-            agent_name="TestUser"
+            agent_name="TestUser",
+            config=_get_raw_config(),
        )

        assert len(facts) > 0, "Should extract at least one fact"
@@ -211,7 +217,8 @@ She's enthusiastic about the opportunity.
            event_date=datetime(2024, 11, 13),
            context=context,
            llm_config=llm_config,
-            agent_name="TestUser"
+            agent_name="TestUser",
+            config=_get_raw_config(),
        )

        assert len(facts) > 0, "Should extract at least one fact"
@@ -244,7 +251,8 @@ I'm planning to switch careers because I'm not fulfilled in my current role.
            event_date=datetime(2024, 11, 13),
            context=context,
            llm_config=llm_config,
-            agent_name="TestUser"
+            agent_name="TestUser",
+            config=_get_raw_config(),
        )

        assert len(facts) > 0, "Should extract at least one fact"
@@ -281,7 +289,8 @@ Family is the most important thing to her.
            event_date=datetime(2024, 11, 13),
            context=context,
            llm_config=llm_config,
-            agent_name="TestUser"
+            agent_name="TestUser",
+            config=_get_raw_config(),
        )

        assert len(facts) > 0, "Should extract at least one fact"
@@ -315,7 +324,8 @@ I prefer presenting in person rather than virtually because I can read the room
            event_date=event_date,
            context=context,
            llm_config=llm_config,
-            agent_name="TestUser"
+            agent_name="TestUser",
+            config=_get_raw_config(),
        )

        assert len(facts) > 0, "Should extract at least one fact"
@@ -372,7 +382,8 @@ I'm planning to visit Tokyo next month.
            event_date=event_date,
            context=context,
            llm_config=llm_config,
-            agent_name="TestUser"
+            agent_name="TestUser",
+            config=_get_raw_config(),
        )

        assert len(facts) > 0, "Should extract at least one fact"
@@ -423,7 +434,8 @@ with a concert surrounded by music, joy and the warm summer breeze.
                    event_date=event_date,
                    context=context,
                    llm_config=llm_config,
-                    agent_name="Melanie"
+                    agent_name="Melanie",
+                    config=_get_raw_config(),
                )

                assert len(facts) > 0, "Should extract at least one fact"
@@ -493,7 +505,8 @@ It was a beautiful day and I plan to make this a regular habit.
            event_date=event_date,
            context=context,
            llm_config=llm_config,
-            agent_name="TestUser"
+            agent_name="TestUser",
+            config=_get_raw_config(),
        )

        assert len(facts) > 0, "Should extract at least one fact"
@@ -547,7 +560,8 @@ It was a beautiful day and I plan to make this a regular habit.
            event_date=reference_date,
            llm_config=llm_config,
            agent_name="TestUser",
-            context="Personal diary"
+            context="Personal diary",
+            config=_get_raw_config(),
        )

        assert len(facts) > 0, "Should extract at least one fact"
@@ -577,7 +591,8 @@ It was a beautiful day and I plan to make this a regular habit.
            event_date=reference_date,
            llm_config=llm_config,
            agent_name="TestUser",
-            context="General info"
+            context="General info",
+            config=_get_raw_config(),
        )

        assert len(facts) > 0, "Should extract at least one fact"
@@ -604,7 +619,8 @@ It was a beautiful day and I plan to make this a regular habit.
            event_date=reference_date,
            llm_config=llm_config,
            agent_name="TestUser",
-            context="Calendar events"
+            context="Calendar events",
+            config=_get_raw_config(),
        )

        assert len(facts) > 0, "Should extract at least one fact"
@@ -655,7 +671,8 @@ great time! Every time I see it, I can't help but smile.
            event_date=event_date,
            context=context,
            llm_config=llm_config,
-            agent_name="Deborah"
+            agent_name="Deborah",
+            config=_get_raw_config(),
        )

        assert len(facts) > 0, "Should extract at least one fact"
@@ -705,7 +722,8 @@ I've learned so much from it.
            event_date=datetime(2024, 11, 13),
            context=context,
            llm_config=llm_config,
-            agent_name="TestUser"
+            agent_name="TestUser",
+            config=_get_raw_config(),
        )

        assert len(facts) > 0, "Should extract at least one fact"
@@ -774,7 +792,8 @@ Jamie: Congratulations! I'd love to read it.
            event_date=datetime(2024, 11, 13),
            llm_config=llm_config,
            agent_name="Marcus",
-            context=context
+            context=context,
+            config=_get_raw_config(),
        )

        assert len(facts) > 0, "Should extract at least one fact from the transcript"
@@ -819,7 +838,8 @@ We presented our findings to the team yesterday.
            event_date=datetime(2024, 11, 13),
            llm_config=llm_config,
            agent_name="TestUser",
-            context=context
+            context=context,
+            config=_get_raw_config(),
        )

        assert len(facts) > 0, "Should extract facts"
@@ -854,7 +874,8 @@ Jamie: [teasing] We'll see who's right, my Niners pick is solid.
            event_date=datetime(2024, 11, 14),
            context=context,
            llm_config=llm_config,
-            agent_name=agent_name
+            agent_name=agent_name,
+            config=_get_raw_config(),
        )

        assert len(facts) > 0, "Should extract at least one fact"
@@ -920,7 +941,8 @@ so the algorithm learns to box out. See you next week!
                    event_date=datetime(2024, 11, 13),
                    llm_config=llm_config,
                    agent_name="Marcus",
-                    context=context
+                    context=context,
+                    config=_get_raw_config(),
                )

                assert len(facts) > 0, "Should extract at least one fact"
@@ -0,0 +1,491 @@
+"""
+Tests for hierarchical configuration system.
+
+Tests config resolution hierarchy (global → tenant → bank),
+key normalization, API endpoints, validation, and caching.
+"""
+
+import os
+
+import pytest
+
+from hindsight_api import MemoryEngine
+from hindsight_api.config import HindsightConfig, normalize_config_dict, normalize_config_key
+from hindsight_api.config_resolver import ConfigResolver
+from hindsight_api.extensions.tenant import TenantExtension
+from hindsight_api.models import RequestContext
+
+# Enable bank config API for all tests in this module
+os.environ["HINDSIGHT_API_ENABLE_BANK_CONFIG_API"] = "true"
+
+
+class MockTenantExtension(TenantExtension):
+    """Mock tenant extension for testing tenant-level config."""
+
+    def __init__(self, tenant_config: dict):
+        self.tenant_config = tenant_config
+
+    async def authenticate(self, context):
+        from hindsight_api.extensions.tenant import TenantContext
+
+        return TenantContext(schema_name="public")
+
+    async def list_tenants(self):
+        from hindsight_api.extensions.tenant import Tenant
+
+        return [Tenant(schema="public")]
+
+    async def get_tenant_config(self, context):
+        """Return mock tenant config."""
+        return self.tenant_config
+
+
+@pytest.mark.asyncio
+async def test_config_key_normalization():
+    """Test that env var keys are normalized to Python field names."""
+    # Test basic normalization
+    assert normalize_config_key("HINDSIGHT_API_LLM_PROVIDER") == "llm_provider"
+    assert normalize_config_key("HINDSIGHT_API_LLM_MODEL") == "llm_model"
+    assert normalize_config_key("HINDSIGHT_API_RETAIN_LLM_PROVIDER") == "retain_llm_provider"
+
+    # Test already normalized keys
+    assert normalize_config_key("llm_provider") == "llm_provider"
+    assert normalize_config_key("llm_model") == "llm_model"
+
+    # Test dict normalization
+    input_dict = {
+        "HINDSIGHT_API_LLM_PROVIDER": "openai",
+        "HINDSIGHT_API_LLM_MODEL": "gpt-4",
+        "llm_base_url": "https://api.openai.com",
+    }
+    expected = {"llm_provider": "openai", "llm_model": "gpt-4", "llm_base_url": "https://api.openai.com"}
+    assert normalize_config_dict(input_dict) == expected
+
+
+@pytest.mark.asyncio
+async def test_hierarchical_fields_categorization():
+    """Test that fields are correctly categorized as configurable, credentials, or static."""
+    configurable = HindsightConfig.get_configurable_fields()
+    credentials = HindsightConfig.get_credential_fields()
+    static = HindsightConfig.get_static_fields()
+
+    # Verify no overlap between configurable and credentials
+    assert len(configurable & credentials) == 0, "Configurable fields should not include credentials"
+
+    # Verify configurable fields include behavioral settings (safe to modify)
+    assert "retain_extraction_mode" in configurable
+    assert "enable_observations" in configurable
+    assert "retain_chunk_size" in configurable
+    assert "retain_custom_instructions" in configurable
+
+    # Verify count is correct (only 4 fields)
+    assert len(configurable) == 4
+
+    # Verify credential fields (NEVER exposed)
+    assert "llm_api_key" in credentials
+    assert "llm_base_url" in credentials
+    assert "retain_llm_api_key" in credentials
+    assert "reflect_llm_api_key" in credentials
+
+    # Verify static fields include server settings AND non-configurable LLM fields
+    assert "database_url" in static
+    assert "port" in static
+    assert "host" in static
+    assert "embeddings_provider" in static
+    assert "reranker_provider" in static
+    assert "worker_enabled" in static
+    assert "llm_provider" in static  # Not configurable (needs presets)
+    assert "llm_model" in static  # Not configurable (needs presets)
+    assert "graph_retriever" in static  # Performance tuning, not configurable
+    assert "llm_max_concurrent" in static  # Performance tuning, not configurable
+
+
+@pytest.mark.asyncio
+async def test_config_hierarchy_resolution(memory, request_context):
+    """Test that config resolution follows global → tenant → bank hierarchy."""
+    bank_id = "test-hierarchy-bank"
+
+    try:
+        # Ensure bank exists in database
+        await memory.get_bank_profile(bank_id, request_context=request_context)
+
+        # Set up mock tenant extension with tenant-level config (use configurable fields only)
+        tenant_config = {"retain_chunk_size": 5000, "retain_extraction_mode": "tenant-mode"}
+        mock_tenant = MockTenantExtension(tenant_config)
+
+        # Create config resolver with mock tenant extension
+        resolver = ConfigResolver(pool=memory._pool, tenant_extension=mock_tenant)
+
+        # Test 1: Global config only (no overrides)
+        context = RequestContext(api_key=None, api_key_id=None, tenant_id=None, internal=False)
+        config = await resolver.get_bank_config(bank_id, context)
+
+        # Should have configurable fields from global config (NOT credentials or llm_provider/model)
+        assert "retain_chunk_size" in config  # Configurable field
+        assert "llm_api_key" not in config  # Credential - never exposed
+        assert "llm_provider" not in config  # Not configurable (needs presets)
+
+        # Test 2: Add tenant-level overrides
+        config = await resolver.get_bank_config(bank_id, context)
+
+        # Should apply tenant overrides (only configurable fields)
+        assert config["retain_chunk_size"] == 5000  # Tenant override
+        assert config["retain_extraction_mode"] == "tenant-mode"  # Tenant override
+
+        # Test 3: Add bank-level overrides (should take precedence)
+        await resolver.update_bank_config(
+            bank_id,
+            {"retain_chunk_size": 2000, "retain_extraction_mode": "bank-mode"},  # Override tenant settings
+            context,
+        )
+
+        # Config should reflect changes immediately (no caching)
+        config = await resolver.get_bank_config(bank_id, context)
+
+        # Bank overrides should take precedence over tenant
+        assert config["retain_chunk_size"] == 2000  # Bank override wins
+        assert config["retain_extraction_mode"] == "bank-mode"  # Bank override wins
+
+    finally:
+        await memory.delete_bank(bank_id, request_context=request_context)
+
+
+@pytest.mark.asyncio
+async def test_config_validation_rejects_static_fields(memory, request_context):
+    """Test that attempting to override static fields raises ValueError."""
+    bank_id = "test-validation-bank"
+
+    try:
+        # Ensure bank exists in database
+        await memory.get_bank_profile(bank_id, request_context=request_context)
+
+        resolver = ConfigResolver(pool=memory._pool)
+
+        # Test 1: Configurable fields should work
+        await resolver.update_bank_config(bank_id, {"retain_chunk_size": 4000, "retain_extraction_mode": "verbose"})
+
+        # Test 2: Static fields should raise ValueError
+        with pytest.raises(ValueError, match="Cannot override static"):
+            await resolver.update_bank_config(bank_id, {"port": 9000})
+
+        with pytest.raises(ValueError, match="Cannot override static"):
+            await resolver.update_bank_config(bank_id, {"database_url": "postgresql://fake"})
+
+        with pytest.raises(ValueError, match="Cannot override static"):
+            await resolver.update_bank_config(bank_id, {"embeddings_provider": "openai"})
+
+        # Test 3: Credential fields should raise ValueError
+        with pytest.raises(ValueError, match="Cannot set credential fields"):
+            await resolver.update_bank_config(bank_id, {"llm_api_key": "sk-fake"})
+
+        # Test 4: Non-configurable LLM fields should raise ValueError (need presets)
+        with pytest.raises(ValueError, match="Cannot override static"):
+            await resolver.update_bank_config(bank_id, {"llm_model": "gpt-4"})
+
+        # Test 5: Mix of configurable and static should fail
+        with pytest.raises(ValueError, match="Cannot override static"):
+            await resolver.update_bank_config(bank_id, {"retain_chunk_size": 4000, "port": 9000})
+
+    finally:
+        await memory.delete_bank(bank_id, request_context=request_context)
+
+
+@pytest.mark.asyncio
+async def test_config_freshness_across_updates(memory, request_context):
+    """Test that config changes are immediately visible (no stale cache)."""
+    bank1 = "freshness-test-1"
+
+    try:
+        # Ensure bank exists in database
+        await memory.get_bank_profile(bank1, request_context=request_context)
+
+        resolver = ConfigResolver(pool=memory._pool)
+
+        # Test 1: Initial config reflects global defaults
+        config1 = await resolver.get_bank_config(bank1, None)
+        initial_chunk_size = config1["retain_chunk_size"]
+
+        # Test 2: Update config
+        await resolver.update_bank_config(bank1, {"retain_chunk_size": 4000})
+
+        # Test 3: Next call should see updated value immediately (no stale cache)
+        config2 = await resolver.get_bank_config(bank1, None)
+        assert config2["retain_chunk_size"] == 4000
+
+        # Test 4: Multiple updates are all immediately visible
+        await resolver.update_bank_config(bank1, {"retain_chunk_size": 4500})
+        config3 = await resolver.get_bank_config(bank1, None)
+        assert config3["retain_chunk_size"] == 4500
+
+        # Test 5: Reset restores global defaults immediately
+        await resolver.reset_bank_config(bank1)
+        config4 = await resolver.get_bank_config(bank1, None)
+        assert config4["retain_chunk_size"] == initial_chunk_size  # Back to global default
+
+        # Test 6: Each call returns a fresh config dict (not a cached reference)
+        config5 = await resolver.get_bank_config(bank1, None)
+        config6 = await resolver.get_bank_config(bank1, None)
+        assert config5 is not config6  # Different object instances
+
+    finally:
+        await memory.delete_bank(bank1, request_context=request_context)
+
+
+@pytest.mark.asyncio
+async def test_config_reset_to_defaults(memory, request_context):
+    """Test that resetting config removes all bank-specific overrides."""
+    bank_id = "test-reset-bank"
+
+    try:
+        # Ensure bank exists in database
+        await memory.get_bank_profile(bank_id, request_context=request_context)
+
+        resolver = ConfigResolver(pool=memory._pool)
+
+        # Add bank-specific overrides
+        await resolver.update_bank_config(
+            bank_id,
+            {
+                "retain_chunk_size": 5500,
+                "retain_extraction_mode": "custom",
+                "retain_custom_instructions": "Custom instructions",
+            },
+        )
+
+        # Verify overrides applied
+        config = await resolver.get_bank_config(bank_id, None)
+        assert config["retain_chunk_size"] == 5500
+        assert config["retain_extraction_mode"] == "custom"
+        assert config["retain_custom_instructions"] == "Custom instructions"
+
+        # Reset to defaults
+        await resolver.reset_bank_config(bank_id)
+
+        # Verify overrides removed (back to global defaults)
+        config_reset = await resolver.get_bank_config(bank_id, None)
+        assert config_reset["retain_chunk_size"] != 5500  # Should be global default
+        assert config_reset["retain_extraction_mode"] != "custom"  # Should be global default
+
+        # Verify bank_config is empty
+        bank_overrides = await resolver._load_bank_config(bank_id)
+        assert bank_overrides == {}
+
+    finally:
+        await memory.delete_bank(bank_id, request_context=request_context)
+
+
+@pytest.mark.asyncio
+async def test_config_supports_both_key_formats(memory, request_context):
+    """Test that API accepts both env var and Python field formats."""
+    bank_id = "test-key-format-bank"
+
+    try:
+        # Ensure bank exists in database
+        await memory.get_bank_profile(bank_id, request_context=request_context)
+
+        resolver = ConfigResolver(pool=memory._pool)
+
+        # Test 1: Python field format
+        await resolver.update_bank_config(bank_id, {"retain_chunk_size": 7000})
+
+        config = await resolver.get_bank_config(bank_id, None)
+        assert config["retain_chunk_size"] == 7000
+
+        # Test 2: Env var format (should be normalized)
+        await resolver.update_bank_config(bank_id, {"HINDSIGHT_API_RETAIN_CHUNK_SIZE": 8000})
+
+        config = await resolver.get_bank_config(bank_id, None)
+        assert config["retain_chunk_size"] == 8000
+
+        # Test 3: Mixed format in same request
+        await resolver.update_bank_config(
+            bank_id,
+            {
+                "retain_chunk_size": 9000,  # Python format
+                "HINDSIGHT_API_RETAIN_EXTRACTION_MODE": "verbose",  # Env format
+            },
+        )
+
+        config = await resolver.get_bank_config(bank_id, None)
+        assert config["retain_chunk_size"] == 9000
+        assert config["retain_extraction_mode"] == "verbose"
+
+    finally:
+        await memory.delete_bank(bank_id, request_context=request_context)
+
+
+@pytest.mark.asyncio
+async def test_config_only_configurable_fields_stored(memory, request_context):
+    """Test that only configurable fields are stored in bank config."""
+    bank_id = "test-filter-bank"
+
+    try:
+        # Ensure bank exists in database
+        await memory.get_bank_profile(bank_id, request_context=request_context)
+
+        resolver = ConfigResolver(pool=memory._pool)
+
+        # Add valid configurable field
+        await resolver.update_bank_config(bank_id, {"retain_chunk_size": 3500})
+
+        # Load bank config and verify only configurable fields present
+        bank_overrides = await resolver._load_bank_config(bank_id)
+
+        for key in bank_overrides.keys():
+            assert key in HindsightConfig.get_configurable_fields(), f"Non-configurable field {key} in bank config"
+
+    finally:
+        await memory.delete_bank(bank_id, request_context=request_context)
+
+
+@pytest.mark.asyncio
+async def test_config_get_bank_config_no_static_or_credential_fields_leak(memory, request_context):
+    """
+    SECURITY TEST: Verify get_bank_config() only returns configurable fields (no static/credentials).
+
+    This prevents leaking sensitive system configuration like database URLs,
+    API keys, LLM providers/models, worker counts, etc. when retrieving bank configuration.
+    """
+    bank_id = "test-security-bank"
+
+    try:
+        # Ensure bank exists in database
+        await memory.get_bank_profile(bank_id, request_context=request_context)
+
+        resolver = ConfigResolver(pool=memory._pool)
+
+        # Get bank config
+        config = await resolver.get_bank_config(bank_id, None)
+
+        # Get field categorizations
+        configurable_fields = HindsightConfig.get_configurable_fields()
+        credential_fields = HindsightConfig.get_credential_fields()
+        static_fields = HindsightConfig.get_static_fields()
+
+        # SECURITY: Verify ONLY configurable fields are returned (NO static, NO credentials)
+        for key in config.keys():
+            assert key in configurable_fields, (
+                f"SECURITY VIOLATION: Non-configurable field '{key}' returned by get_bank_config(). "
+                f"Only configurable fields should be returned to prevent leaking system config."
+            )
+            assert key not in credential_fields, (
+                f"SECURITY VIOLATION: Credential field '{key}' returned by get_bank_config(). "
+                f"Credentials must NEVER be exposed via API."
+            )
+
+        # SECURITY: Verify specific sensitive fields are NOT present
+        sensitive_fields = [
+            "database_url", "api_port", "host", "worker_count",  # Infrastructure
+            "llm_api_key", "llm_base_url",  # Credentials
+            "retain_llm_api_key", "reflect_llm_api_key",  # More credentials
+            "llm_provider", "llm_model",  # Not configurable (need presets)
+        ]
+        for field in sensitive_fields:
+            assert field not in config, (
+                f"SECURITY VIOLATION: Sensitive field '{field}' returned by get_bank_config(). "
+                f"Must not be exposed via bank config API."
+            )
+
+        # Verify we have the expected configurable fields (small set)
+        expected_configurable = ["retain_chunk_size", "retain_extraction_mode", "enable_observations"]
+        for field in expected_configurable:
+            assert field in config, f"Expected configurable field '{field}' missing from config"
+
+        # Should have a small number of configurable fields (not hundreds)
+        assert len(config) < 20, f"Too many fields returned: {len(config)}"
+
+    finally:
+        await memory.delete_bank(bank_id, request_context=request_context)
+
+
+@pytest.mark.asyncio
+async def test_config_permissions_system(memory, request_context):
+    """
+    Test that tenant extension can control which fields banks are allowed to modify.
+
+    Tests get_allowed_config_fields() permission system.
+    """
+    bank_id = "test-permissions-bank"
+
+    class PermissionTenantExtension(TenantExtension):
+        """Mock tenant extension with configurable permissions."""
+
+        def __init__(self, allowed_fields: set[str] | None):
+            self.allowed_fields = allowed_fields
+
+        async def authenticate(self, context):
+            from hindsight_api.extensions.tenant import TenantContext
+
+            return TenantContext(schema_name="public")
+
+        async def list_tenants(self):
+            from hindsight_api.extensions.tenant import Tenant
+
+            return [Tenant(schema="public")]
+
+        async def get_allowed_config_fields(self, context, bank_id):
+            """Return configured allowed fields."""
+            return self.allowed_fields
+
+    try:
+        # Ensure bank exists in database
+        await memory.get_bank_profile(bank_id, request_context=request_context)
+
+        # Test 1: None = allow all configurable fields
+        extension = PermissionTenantExtension(allowed_fields=None)
+        resolver = ConfigResolver(pool=memory._pool, tenant_extension=extension)
+
+        await resolver.update_bank_config(
+            bank_id, {"retain_chunk_size": 4000, "retain_extraction_mode": "verbose"}, request_context
+        )
+        config = await resolver.get_bank_config(bank_id, request_context)
+        assert config["retain_chunk_size"] == 4000
+        assert config["retain_extraction_mode"] == "verbose"
+
+        # Reset for next test
+        await resolver.reset_bank_config(bank_id)
+
+        # Test 2: Specific set = only those fields allowed
+        extension = PermissionTenantExtension(allowed_fields={"retain_chunk_size"})
+        resolver = ConfigResolver(pool=memory._pool, tenant_extension=extension)
+
+        # Should allow retain_chunk_size
+        await resolver.update_bank_config(bank_id, {"retain_chunk_size": 5000}, request_context)
+        config = await resolver.get_bank_config(bank_id, request_context)
+        assert config["retain_chunk_size"] == 5000
+
+        # Should reject retain_extraction_mode (not in allowed list)
+        with pytest.raises(ValueError, match="Not allowed to modify fields"):
+            await resolver.update_bank_config(bank_id, {"retain_extraction_mode": "verbose"}, request_context)
+
+        # Should reject mix of allowed and disallowed
+        with pytest.raises(ValueError, match="Not allowed to modify fields"):
+            await resolver.update_bank_config(
+                bank_id, {"retain_chunk_size": 6000, "retain_extraction_mode": "verbose"}, request_context
+            )
+
+        # Reset for next test
+        await resolver.reset_bank_config(bank_id)
+
+        # Test 3: Empty set = no modifications allowed (read-only)
+        extension = PermissionTenantExtension(allowed_fields=set())
+        resolver = ConfigResolver(pool=memory._pool, tenant_extension=extension)
+
+        with pytest.raises(ValueError, match="Not allowed to modify fields"):
+            await resolver.update_bank_config(bank_id, {"retain_chunk_size": 7000}, request_context)
+
+        # Test 4: get_bank_config should filter response based on permissions
+        extension = PermissionTenantExtension(allowed_fields={"retain_chunk_size", "enable_observations"})
+        resolver = ConfigResolver(pool=memory._pool, tenant_extension=extension)
+
+        config = await resolver.get_bank_config(bank_id, request_context)
+
+        # Should only return allowed fields
+        assert "retain_chunk_size" in config
+        assert "enable_observations" in config
+        # Other configurable fields should be filtered out
+        assert "retain_extraction_mode" not in config
+        assert "retain_custom_instructions" not in config
+
+    finally:
+        await memory.delete_bank(bank_id, request_context=request_context)
@@ -12,9 +12,9 @@ import pytest
@pytest.fixture(autouse=True)
 def enable_observations():
    """Enable observations for all tests in this module."""
-    from hindsight_api.config import get_config
+    from hindsight_api.config import _get_raw_config

-    config = get_config()
+    config = _get_raw_config()
    original_value = config.enable_observations
    config.enable_observations = True
    yield
@@ -1,5 +1,10 @@
 """
-Test LLM provider with different models using actual memory operations.
+Test LLM provider with different models using actual Hindsight memory operations.
+
+Tests validate that providers work correctly with:
+1. Retain (memory ingestion with fact extraction)
+2. Reflect (memory retrieval with tool calling)
+3. Mental models (consolidated knowledge generation)
 """
 import os
 from datetime import datetime
@@ -19,6 +24,10 @@ MODEL_MATRIX = [
    ("openai", "gpt-5-nano"),
    ("openai", "gpt-5"),
    ("openai", "gpt-5.2"),
+    # Anthropic models
+    ("anthropic", "claude-sonnet-4-20250514"),
+    ("anthropic", "claude-opus-4-5-20251101"),
+    ("anthropic", "claude-haiku-4-20250514"),
    # Groq models
    ("groq", "openai/gpt-oss-120b"),
    ("groq", "openai/gpt-oss-20b"),
@@ -29,6 +38,12 @@ MODEL_MATRIX = [
    # Ollama models (local)
    ("ollama", "gemma3:12b"),
    ("ollama", "gemma3:1b"),
+    # Claude Code (uses Claude Agent SDK with Claude models)
+    ("claude-code", "claude-sonnet-4-20250514"),
+    # OpenAI Codex (uses MCP with Codex-specific models)
+    ("openai-codex", "gpt-5.2-codex"),
+    # Mock provider (for testing)
+    ("mock", "mock"),
 ]


@@ -36,6 +51,7 @@ def get_api_key_for_provider(provider: str) -> str | None:
    """Get API key for provider from environment variables."""
    provider_key_map = {
        "openai": "OPENAI_API_KEY",
+        "anthropic": "ANTHROPIC_API_KEY",
        "groq": "GROQ_API_KEY",
        "gemini": "GEMINI_API_KEY",
    }
@@ -43,6 +59,171 @@ def get_api_key_for_provider(provider: str) -> str | None:
    return os.getenv(env_var) if env_var else None


+def should_skip_provider(provider: str, model: str = "") -> tuple[bool, str]:
+    """Check if provider should be skipped and return reason."""
+    # Never skip mock provider
+    if provider == "mock":
+        return False, ""
+
+    # Skip claude-code and openai-codex in CI (require local auth)
+    if os.getenv("CI") and provider in ("claude-code", "openai-codex"):
+        return True, f"{provider} not available in CI (requires local authentication)"
+
+    # Skip Ollama in CI (no models available)
+    if provider == "ollama" and os.getenv("CI"):
+        return True, "Ollama not available in CI"
+
+    # Skip Ollama gemma models (don't support tool calling)
+    if provider == "ollama" and "gemma" in model.lower():
+        return True, f"Ollama {model} does not support tool calling"
+
+    # Other providers need an API key
+    if provider not in ("ollama", "claude-code", "openai-codex", "mock"):
+        api_key = get_api_key_for_provider(provider)
+        if not api_key:
+            return True, f"No API key available (set {provider.upper()}_API_KEY)"
+
+    return False, ""
+
+
+@pytest.mark.parametrize("provider,model", MODEL_MATRIX)
+@pytest.mark.asyncio
+@pytest.mark.timeout(300)  # Increase timeout for slow models like groq gpt-oss-120b
+async def test_llm_provider_api_methods(provider: str, model: str):
+    """
+    Test all LLM API methods used by Hindsight at runtime.
+    This validates that the provider correctly implements the LLMInterface.
+
+    Tests:
+    1. verify_connection() - Connection verification
+    2. call() with plain text - Basic LLM call
+    3. call() with response_format - Structured output (used in fact extraction)
+    4. call_with_tools() - Tool calling (used in reflect agent)
+    """
+    # Skip mock provider - it's a test stub, not a real LLM implementation
+    if provider == "mock":
+        pytest.skip("Mock provider is a test stub, not a real LLM")
+
+    should_skip, reason = should_skip_provider(provider, model)
+    if should_skip:
+        pytest.skip(f"Skipping {provider}/{model}: {reason}")
+
+    api_key = get_api_key_for_provider(provider)
+
+    llm = LLMProvider(
+        provider=provider,
+        api_key=api_key or "",
+        base_url="",
+        model=model,
+    )
+
+    print(f"\n{provider}/{model} - API methods test:")
+
+    # Test 1: verify_connection()
+    try:
+        await llm.verify_connection()
+        print("  ✓ verify_connection()")
+    except Exception as e:
+        pytest.fail(f"{provider}/{model} verify_connection() failed: {e}")
+
+    # Test 2: call() with plain text
+    try:
+        response = await llm.call(
+            messages=[
+                {"role": "system", "content": "You are a helpful assistant."},
+                {"role": "user", "content": "What is 2+2? Answer in one word."},
+            ],
+            max_completion_tokens=50,
+        )
+        assert response is not None, "call() returned None"
+        assert len(response) > 0, "call() returned empty string"
+        print(f"  ✓ call() plain text: {response[:50]}")
+    except Exception as e:
+        pytest.fail(f"{provider}/{model} call() plain text failed: {e}")
+
+    # Test 3: call() with response_format (structured output)
+    # Skip for models that don't support structured output
+    skip_structured_output = (provider == "groq" and "gpt-oss-120b" in model.lower())
+    if skip_structured_output:
+        print(f"  ⊘ call() structured output: skipped (model doesn't support response_format)")
+    else:
+        try:
+            from pydantic import BaseModel
+
+            class TestResponse(BaseModel):
+                answer: str
+                confidence: str
+
+            response = await llm.call(
+                messages=[
+                    {"role": "system", "content": "You are a math assistant."},
+                    {"role": "user", "content": "What is the capital of France?"},
+                ],
+                response_format=TestResponse,
+                max_completion_tokens=100,
+            )
+            assert isinstance(response, TestResponse), f"Expected TestResponse, got {type(response)}"
+            assert hasattr(response, "answer"), "Structured output missing 'answer' field"
+            assert hasattr(response, "confidence"), "Structured output missing 'confidence' field"
+            print(f"  ✓ call() structured output: answer={response.answer}, confidence={response.confidence}")
+        except Exception as e:
+            pytest.fail(f"{provider}/{model} call() structured output failed: {e}")
+
+    # Test 4: call_with_tools() (tool calling)
+    try:
+        tools = [
+            {
+                "type": "function",
+                "function": {
+                    "name": "get_weather",
+                    "description": "Get the weather for a location",
+                    "parameters": {
+                        "type": "object",
+                        "properties": {
+                            "location": {"type": "string", "description": "City name"},
+                            "unit": {"type": "string", "enum": ["celsius", "fahrenheit"]},
+                        },
+                        "required": ["location"],
+                    },
+                },
+            }
+        ]
+
+        result = await llm.call_with_tools(
+            messages=[
+                {"role": "system", "content": "You are a helpful assistant with access to tools."},
+                {"role": "user", "content": "What's the weather like in Paris?"},
+            ],
+            tools=tools,
+            max_completion_tokens=500,  # Increased from 200 to give models enough space for tool calls
+        )
+
+        assert result is not None, "call_with_tools() returned None"
+        assert hasattr(result, "tool_calls"), "Result missing 'tool_calls' attribute"
+
+        # Nano models may hit token limits before making tool calls - that's acceptable
+        is_nano_model = "nano" in model.lower()
+        if is_nano_model and len(result.tool_calls) == 0:
+            # Check if it hit length limit (expected for nano models)
+            if hasattr(result, "finish_reason") and result.finish_reason == "length":
+                print(f"  ✓ call_with_tools(): nano model hit token limit (expected)")
+            else:
+                pytest.fail(f"Nano model made 0 tool calls but didn't hit length limit (finish_reason={getattr(result, 'finish_reason', 'unknown')})")
+        else:
+            assert len(result.tool_calls) > 0, f"Expected at least 1 tool call, got {len(result.tool_calls)}"
+
+            # Verify tool call structure
+            tool_call = result.tool_calls[0]
+            assert hasattr(tool_call, "name"), "Tool call missing 'name'"
+            assert hasattr(tool_call, "arguments"), "Tool call missing 'arguments'"
+            assert tool_call.name == "get_weather", f"Expected 'get_weather', got '{tool_call.name}'"
+            assert "location" in tool_call.arguments, "Tool call arguments missing 'location'"
+
+            print(f"  ✓ call_with_tools(): {tool_call.name}({tool_call.arguments})")
+    except Exception as e:
+        pytest.fail(f"{provider}/{model} call_with_tools() failed: {e}")
+
+
@pytest.mark.parametrize("provider,model", MODEL_MATRIX)
@pytest.mark.asyncio
 async def test_llm_provider_memory_operations(provider: str, model: str):
@@ -50,16 +231,16 @@ async def test_llm_provider_memory_operations(provider: str, model: str):
    Test LLM provider with actual memory operations: fact extraction and reflect.
    All models must pass this test.
    """
+    # Skip mock provider - it's a test stub, not designed for real operations
+    if provider == "mock":
+        pytest.skip("Mock provider is a test stub, not designed for real operations")
+
+    should_skip, reason = should_skip_provider(provider, model)
+    if should_skip:
+        pytest.skip(f"Skipping {provider}/{model}: {reason}")
+
    api_key = get_api_key_for_provider(provider)

-    # Skip Ollama tests in CI (no models available)
-    if provider == "ollama" and os.getenv("CI"):
-        pytest.skip(f"Skipping {provider}/{model}: Ollama not available in CI")
-
-    # Other providers need an API key
-    if provider != "ollama" and not api_key:
-        pytest.skip(f"Skipping {provider}/{model}: no API key available")
-
    llm = LLMProvider(
        provider=provider,
        api_key=api_key or "",
@@ -117,3 +298,115 @@ async def test_llm_provider_memory_operations(provider: str, model: str):

    assert response is not None, f"{provider}/{model} reflect returned None"
    assert len(response) > 10, f"{provider}/{model} reflect response too short"
+
+
+@pytest.mark.parametrize("provider,model", [
+    ("claude-code", "claude-sonnet-4-20250514"),
+    ("openai-codex", "gpt-5.2-codex"),
+])
+@pytest.mark.asyncio
+async def test_llm_provider_consolidation(memory_no_llm_verify, request_context, provider: str, model: str):
+    """
+    Test LLM provider with consolidation (automatic mental model generation from observations).
+    This validates that the provider can generate synthesized knowledge from raw memories.
+
+    This test is limited to claude-code and codex since they're the critical providers
+    that needed tool calling fixes for reflect and consolidation operations.
+    """
+    should_skip, reason = should_skip_provider(provider, model)
+    if should_skip:
+        pytest.skip(f"Skipping {provider}/{model}: {reason}")
+
+    # Use provider-specific LLM for this test
+    api_key = get_api_key_for_provider(provider)
+    memory_no_llm_verify._consolidation_llm = LLMProvider(
+        provider=provider,
+        api_key=api_key or "",
+        base_url="",
+        model=model,
+    )
+    # Also need retain LLM for ingesting data
+    memory_no_llm_verify._retain_llm = memory_no_llm_verify._consolidation_llm
+
+    test_bank_id = f"llm_test_consolidation_{provider}_{model}_{datetime.now().timestamp()}"
+
+    # Enable observations for this bank
+    from hindsight_api.config import get_config
+    config = get_config()
+    original_value = config.enable_observations
+    config.enable_observations = True
+
+    try:
+        # Retain memories to consolidate
+        test_content = """
+        Bob prefers functional programming with Rust and Haskell.
+        He emphasizes immutability and pure functions in code reviews.
+        Bob advocates for type safety and compile-time guarantees.
+        He avoids mutable state and prefers declarative code patterns.
+        """
+
+        await memory_no_llm_verify.retain_async(
+            bank_id=test_bank_id,
+            content=test_content,
+            context="Team coding preferences",
+            event_date=datetime(2024, 12, 1),
+            request_context=request_context,
+        )
+
+        print(f"\n{provider}/{model} - Consolidation test:")
+
+        # Run consolidation to generate observations (mental models)
+        from hindsight_api.engine.consolidation.consolidator import run_consolidation_job
+
+        result = await run_consolidation_job(
+            memory_engine=memory_no_llm_verify,
+            bank_id=test_bank_id,
+            request_context=request_context,
+        )
+
+        print(f"  Processed: {result.get('memories_processed', 0)} memories")
+        print(f"  Created: {result.get('observations_created', 0)} observations")
+        print(f"  Updated: {result.get('observations_updated', 0)} observations")
+
+        # Verify consolidation ran successfully
+        assert result["status"] in ["success", "no_new_memories"], f"{provider}/{model} consolidation failed"
+
+        # If observations were created, verify they contain relevant content
+        if result.get("observations_created", 0) > 0:
+            observations = await memory_no_llm_verify.list_mental_models_consolidated(
+                bank_id=test_bank_id,
+                request_context=request_context,
+            )
+
+            assert len(observations) > 0, f"{provider}/{model} consolidation created 0 observations"
+
+            # Check first observation contains relevant information
+            obs_content = observations[0].get("content", "").lower()
+            relevant_terms = ["bob", "functional", "rust", "immutab", "type"]
+            matches = [term for term in relevant_terms if term in obs_content]
+
+            print(f"  Observation preview: {observations[0].get('content', '')[:200]}...")
+            print(f"  Found {len(matches)} relevant terms: {matches}")
+
+            assert len(matches) >= 2, (
+                f"{provider}/{model} consolidated observation doesn't contain relevant info. "
+                f"Expected at least 2 of {relevant_terms}, found {len(matches)}: {matches}"
+            )
+
+    finally:
+        # Restore original config
+        config.enable_observations = original_value
+
+
+# NOTE: The tests above validate the critical Hindsight operations:
+#
+# test_llm_provider_memory_operations (ALL providers):
+#   - Fact extraction (retain): tests structured output generation
+#   - Reflect: tests memory retrieval and reasoning (uses tool calling for claude-code/codex)
+#
+# test_llm_provider_consolidation (claude-code and codex only):
+#   - Consolidation: tests automatic mental model generation from observations
+#   - Requires MemoryEngine fixture with working LLM (from .env or env vars)
+#   - Run your local LLM server OR set HINDSIGHT_API_LLM_PROVIDER/API_KEY/MODEL env vars
+#
+# For full end-to-end integration tests using the HTTP API, see tests/test_http_api_integration.py
@@ -30,7 +30,8 @@ async def test_llm_metrics_recorded_for_groq():
    # Create a mock metrics collector to track record_llm_call calls
    mock_collector = MagicMock(spec=MetricsCollector)

-    with patch("hindsight_api.engine.llm_wrapper.get_metrics_collector", return_value=mock_collector):
+    # Patch the provider module where get_metrics_collector is actually called
+    with patch("hindsight_api.engine.providers.openai_compatible_llm.get_metrics_collector", return_value=mock_collector):
        llm = LLMProvider(
            provider="groq",
            api_key=api_key,
@@ -90,7 +91,8 @@ async def test_llm_metrics_recorded_for_structured_output():

    mock_collector = MagicMock(spec=MetricsCollector)

-    with patch("hindsight_api.engine.llm_wrapper.get_metrics_collector", return_value=mock_collector):
+    # Patch the provider module where get_metrics_collector is actually called
+    with patch("hindsight_api.engine.providers.openai_compatible_llm.get_metrics_collector", return_value=mock_collector):
        llm = LLMProvider(
            provider="groq",
            api_key=api_key,
@@ -209,7 +209,7 @@ class TestLargeBatchRetain:
                raise

    @pytest.mark.asyncio
-    @pytest.mark.timeout(120)
+    @pytest.mark.timeout(240)  # Increased timeout for VectorChord BM25 tokenization
    async def test_batch_chunking_behavior(self, memory_with_mock_llm, request_context):
        """
        Test that large batches are properly chunked into sub-batches.
@@ -45,7 +45,7 @@ class TestMainModuleExtensionLoading:

        with patch("hindsight_api.main.MemoryEngine") as mock_engine, \
             patch("hindsight_api.main.create_app") as mock_create_app, \
-             patch("hindsight_api.main.get_config") as mock_get_config, \
+             patch("hindsight_api.main._get_raw_config") as mock_get_config, \
             patch("hindsight_api.main.load_extension", side_effect=tracking_load_extension), \
             patch("hindsight_api.main.DefaultExtensionContext"), \
             patch("hindsight_api.main.print_banner"), \
@@ -96,7 +96,7 @@ class TestMainModuleExtensionLoading:

        with patch("hindsight_api.main.MemoryEngine") as mock_engine, \
             patch("hindsight_api.main.create_app") as mock_create_app, \
-             patch("hindsight_api.main.get_config") as mock_get_config, \
+             patch("hindsight_api.main._get_raw_config") as mock_get_config, \
             patch("hindsight_api.main.load_extension", side_effect=tracking_load_extension), \
             patch("hindsight_api.main.DefaultExtensionContext"), \
             patch("hindsight_api.main.print_banner"), \
@@ -143,7 +143,7 @@ class TestMainModuleExtensionLoading:

        with patch("hindsight_api.main.MemoryEngine", side_effect=capture_memory_engine), \
             patch("hindsight_api.main.create_app") as mock_create_app, \
-             patch("hindsight_api.main.get_config") as mock_get_config, \
+             patch("hindsight_api.main._get_raw_config") as mock_get_config, \
             patch("hindsight_api.main.DefaultExtensionContext"), \
             patch("hindsight_api.main.print_banner"), \
             patch("uvicorn.run"):
@@ -200,7 +200,7 @@ class TestMainModuleExtensionLoading:

        with patch("hindsight_api.main.MemoryEngine", side_effect=capture_memory_engine), \
             patch("hindsight_api.main.create_app") as mock_create_app, \
-             patch("hindsight_api.main.get_config") as mock_get_config, \
+             patch("hindsight_api.main._get_raw_config") as mock_get_config, \
             patch("hindsight_api.main.DefaultExtensionContext", side_effect=capture_context), \
             patch("hindsight_api.main.print_banner"), \
             patch("uvicorn.run"):
@@ -242,7 +242,7 @@ class TestMainModuleExtensionLoading:

        with patch("hindsight_api.main.MemoryEngine", side_effect=capture_memory_engine), \
             patch("hindsight_api.main.create_app") as mock_create_app, \
-             patch("hindsight_api.main.get_config") as mock_get_config, \
+             patch("hindsight_api.main._get_raw_config") as mock_get_config, \
             patch("hindsight_api.main.print_banner"), \
             patch("uvicorn.run"):

@@ -287,7 +287,7 @@ class TestMainModuleExtensionLoading:

        with patch("hindsight_api.main.MemoryEngine") as mock_engine, \
             patch("hindsight_api.main.create_app", return_value=mock_app), \
-             patch("hindsight_api.main.get_config") as mock_get_config, \
+             patch("hindsight_api.main._get_raw_config") as mock_get_config, \
             patch("hindsight_api.main.print_banner"), \
             patch("uvicorn.run", side_effect=capture_uvicorn_run):

@@ -327,7 +327,7 @@ class TestMainModuleExtensionLoading:

        with patch("hindsight_api.main.MemoryEngine") as mock_engine, \
             patch("hindsight_api.main.create_app") as mock_create_app, \
-             patch("hindsight_api.main.get_config") as mock_get_config, \
+             patch("hindsight_api.main._get_raw_config") as mock_get_config, \
             patch("hindsight_api.main.print_banner"), \
             patch("uvicorn.run", side_effect=capture_uvicorn_run):

@@ -0,0 +1,280 @@
+"""Integration test for MCP endpoint routing.
+
+This test verifies that /mcp/ and /mcp/{bank_id}/ expose different tool sets,
+and that URLs with or without trailing slashes both work (no 307 redirect).
+"""
+
+import httpx
+import pytest
+from mcp.client.session import ClientSession
+from mcp.client.streamable_http import streamable_http_client
+
+
+@pytest.mark.asyncio
+async def test_mcp_endpoint_routing_integration(memory):
+    """Test that multi-bank and single-bank endpoints expose different tools using StreamableHTTP.
+
+    This is a regression test for issue #317 where /mcp/{bank_id}/ was incorrectly
+    exposing all tools (including list_banks) and bank_id parameters.
+    """
+    from hindsight_api.api import create_app
+
+    # Create app with MCP enabled
+    app = create_app(memory, mcp_api_enabled=True, initialize_memory=False)
+
+    # Use the app's lifespan context to properly initialize MCP servers
+    async with app.router.lifespan_context(app):
+        # Create an HTTPX client that routes to our ASGI app
+        from httpx import ASGITransport
+
+        async with httpx.AsyncClient(transport=ASGITransport(app=app), base_url="http://test") as http_client:
+            # Test 1: Multi-bank endpoint /mcp/
+            async with streamable_http_client("http://test/mcp/", http_client=http_client) as (
+                read_stream,
+                write_stream,
+                _,
+            ):
+                async with ClientSession(read_stream, write_stream) as session:
+                    await session.initialize()
+                    multi_result = await session.list_tools()
+
+                    multi_tools = {t.name for t in multi_result.tools}
+
+                    # Multi-bank should have all tools including bank management and mental models
+                    assert "retain" in multi_tools
+                    assert "recall" in multi_tools
+                    assert "reflect" in multi_tools
+                    assert "list_banks" in multi_tools, "Multi-bank should expose list_banks"
+                    assert "create_bank" in multi_tools, "Multi-bank should expose create_bank"
+                    assert "list_mental_models" in multi_tools, "Multi-bank should expose list_mental_models"
+                    assert "create_mental_model" in multi_tools, "Multi-bank should expose create_mental_model"
+                    assert "get_mental_model" in multi_tools, "Multi-bank should expose get_mental_model"
+                    assert "update_mental_model" in multi_tools, "Multi-bank should expose update_mental_model"
+                    assert "delete_mental_model" in multi_tools, "Multi-bank should expose delete_mental_model"
+                    assert "refresh_mental_model" in multi_tools, "Multi-bank should expose refresh_mental_model"
+
+                    # Multi-bank retain should have bank_id parameter
+                    retain_tool = next((t for t in multi_result.tools if t.name == "retain"), None)
+                    assert retain_tool is not None
+                    multi_params = set(retain_tool.inputSchema.get("properties", {}).keys())
+                    assert "bank_id" in multi_params, "Multi-bank retain should have bank_id parameter"
+
+            # Test 2: Single-bank endpoint /mcp/test-bank/
+            async with streamable_http_client("http://test/mcp/test-bank/", http_client=http_client) as (
+                read_stream,
+                write_stream,
+                _,
+            ):
+                async with ClientSession(read_stream, write_stream) as session:
+                    await session.initialize()
+                    single_result = await session.list_tools()
+
+                    single_tools = {t.name for t in single_result.tools}
+
+                    # Single-bank should have scoped tools including mental models (no bank management)
+                    assert "retain" in single_tools
+                    assert "recall" in single_tools
+                    assert "reflect" in single_tools
+                    assert "list_mental_models" in single_tools, "Single-bank should expose list_mental_models"
+                    assert "create_mental_model" in single_tools, "Single-bank should expose create_mental_model"
+                    assert "list_banks" not in single_tools, "Single-bank should NOT expose list_banks"
+                    assert "create_bank" not in single_tools, "Single-bank should NOT expose create_bank"
+
+                    # Single-bank retain should NOT have bank_id parameter
+                    retain_tool = next((t for t in single_result.tools if t.name == "retain"), None)
+                    assert retain_tool is not None
+                    single_params = set(retain_tool.inputSchema.get("properties", {}).keys())
+                    assert "bank_id" not in single_params, "Single-bank retain should NOT have bank_id parameter"
+
+
+@pytest.mark.asyncio
+async def test_mcp_no_trailing_slash_works(memory):
+    """Test that /mcp (no trailing slash) discovers tools without 307 redirect.
+
+    Starlette's Mount class redirects /mcp to /mcp/ with a 307 Temporary Redirect.
+    Many MCP clients don't follow POST redirects, causing 0 tools to be discovered.
+    MCPMiddleware wraps the app directly (no Mount), so the redirect never happens.
+    """
+    from hindsight_api.api import create_app
+
+    app = create_app(memory, mcp_api_enabled=True, initialize_memory=False)
+
+    async with app.router.lifespan_context(app):
+        from httpx import ASGITransport
+
+        async with httpx.AsyncClient(transport=ASGITransport(app=app), base_url="http://test") as http_client:
+            # /mcp (no slash) should work the same as /mcp/
+            async with streamable_http_client("http://test/mcp", http_client=http_client) as (
+                read_stream,
+                write_stream,
+                _,
+            ):
+                async with ClientSession(read_stream, write_stream) as session:
+                    await session.initialize()
+                    result = await session.list_tools()
+
+                    tools = {t.name for t in result.tools}
+                    assert len(tools) >= 11, f"Expected at least 11 tools from /mcp, got {len(tools)}: {tools}"
+                    assert "retain" in tools
+                    assert "recall" in tools
+                    assert "list_banks" in tools
+
+            # /mcp/my-bank (single-bank, no slash) should also work
+            async with streamable_http_client("http://test/mcp/my-bank", http_client=http_client) as (
+                read_stream,
+                write_stream,
+                _,
+            ):
+                async with ClientSession(read_stream, write_stream) as session:
+                    await session.initialize()
+                    result = await session.list_tools()
+
+                    tools = {t.name for t in result.tools}
+                    assert "retain" in tools
+                    assert "list_banks" not in tools, "Single-bank /mcp/my-bank should NOT expose list_banks"
+
+
+@pytest.mark.asyncio
+async def test_mcp_tool_execution_through_client(memory):
+    """Test that tools can be called (not just discovered) through the MCP client.
+
+    This verifies the full pipeline: HTTP → middleware → FastMCP → tool → engine → response.
+    Previous tests only checked tool discovery (list_tools), not actual execution.
+    """
+    from httpx import ASGITransport
+
+    from hindsight_api.api import create_app
+
+    app = create_app(memory, mcp_api_enabled=True, initialize_memory=False)
+
+    async with app.router.lifespan_context(app):
+        async with httpx.AsyncClient(transport=ASGITransport(app=app), base_url="http://test") as http_client:
+            async with streamable_http_client("http://test/mcp/", http_client=http_client) as (
+                read_stream,
+                write_stream,
+                _,
+            ):
+                async with ClientSession(read_stream, write_stream) as session:
+                    await session.initialize()
+
+                    # Execute list_banks tool
+                    result = await session.call_tool("list_banks", arguments={})
+                    assert result is not None
+                    assert len(result.content) > 0
+                    # The result text should be valid JSON with a "banks" key
+                    import json
+
+                    response_text = result.content[0].text
+                    parsed = json.loads(response_text)
+                    assert "banks" in parsed
+
+
+@pytest.mark.asyncio
+async def test_mcp_mental_model_validation_through_client(memory):
+    """Test that input validation works through the real MCP transport.
+
+    Verifies that invalid inputs return error messages without crashing,
+    and that the engine is never called with invalid data.
+    """
+    from httpx import ASGITransport
+
+    from hindsight_api.api import create_app
+
+    app = create_app(memory, mcp_api_enabled=True, initialize_memory=False)
+
+    async with app.router.lifespan_context(app):
+        async with httpx.AsyncClient(transport=ASGITransport(app=app), base_url="http://test") as http_client:
+            async with streamable_http_client("http://test/mcp/", http_client=http_client) as (
+                read_stream,
+                write_stream,
+                _,
+            ):
+                async with ClientSession(read_stream, write_stream) as session:
+                    await session.initialize()
+
+                    # Test: empty name should return validation error
+                    import json
+
+                    result = await session.call_tool(
+                        "create_mental_model",
+                        arguments={"name": "", "source_query": "test query"},
+                    )
+                    assert result is not None
+                    parsed = json.loads(result.content[0].text)
+                    assert "error" in parsed
+                    assert "name cannot be empty" in parsed["error"]
+
+                    # Test: max_tokens out of range should return validation error
+                    result = await session.call_tool(
+                        "create_mental_model",
+                        arguments={"name": "Test", "source_query": "test query", "max_tokens": 0},
+                    )
+                    parsed = json.loads(result.content[0].text)
+                    assert "error" in parsed
+                    assert "max_tokens must be between 256 and 8192" in parsed["error"]
+
+
+@pytest.mark.asyncio
+async def test_mcp_bank_named_sse_routes_to_single_bank(memory):
+    """Test that a bank named 'sse' routes to single-bank mode.
+
+    Regression test: the old MCP_ENDPOINTS blocklist prevented banks named 'sse'
+    or 'messages' from being accessed via path routing. They fell through to
+    multi-bank mode instead.
+    """
+    from httpx import ASGITransport
+
+    from hindsight_api.api import create_app
+
+    app = create_app(memory, mcp_api_enabled=True, initialize_memory=False)
+
+    async with app.router.lifespan_context(app):
+        async with httpx.AsyncClient(transport=ASGITransport(app=app), base_url="http://test") as http_client:
+            async with streamable_http_client("http://test/mcp/sse/", http_client=http_client) as (
+                read_stream,
+                write_stream,
+                _,
+            ):
+                async with ClientSession(read_stream, write_stream) as session:
+                    await session.initialize()
+                    result = await session.list_tools()
+                    tools = {t.name for t in result.tools}
+
+                    # Should be single-bank mode (no bank management tools)
+                    assert "retain" in tools
+                    assert "recall" in tools
+                    assert "list_banks" not in tools, "Bank 'sse' should route to single-bank mode"
+                    assert "create_bank" not in tools
+
+                    # retain should NOT have bank_id parameter (single-bank mode)
+                    retain_tool = next(t for t in result.tools if t.name == "retain")
+                    params = set(retain_tool.inputSchema.get("properties", {}).keys())
+                    assert "bank_id" not in params
+
+
+@pytest.mark.asyncio
+async def test_mcp_bank_named_messages_routes_to_single_bank(memory):
+    """Test that a bank named 'messages' routes to single-bank mode.
+
+    Same regression test as test_mcp_bank_named_sse_routes_to_single_bank but for 'messages'.
+    """
+    from httpx import ASGITransport
+
+    from hindsight_api.api import create_app
+
+    app = create_app(memory, mcp_api_enabled=True, initialize_memory=False)
+
+    async with app.router.lifespan_context(app):
+        async with httpx.AsyncClient(transport=ASGITransport(app=app), base_url="http://test") as http_client:
+            async with streamable_http_client("http://test/mcp/messages/", http_client=http_client) as (
+                read_stream,
+                write_stream,
+                _,
+            ):
+                async with ClientSession(read_stream, write_stream) as session:
+                    await session.initialize()
+                    result = await session.list_tools()
+                    tools = {t.name for t in result.tools}
+
+                    assert "retain" in tools
+                    assert "list_banks" not in tools, "Bank 'messages' should route to single-bank mode"
@@ -0,0 +1,169 @@
+"""Tests for MCPExtension loading and tool registration."""
+
+from unittest.mock import MagicMock, patch
+
+import pytest
+from fastmcp import FastMCP
+
+from hindsight_api import MemoryEngine
+from hindsight_api.extensions.mcp import MCPExtension
+
+
+class MockMCPExtension(MCPExtension):
+    """Test extension that registers a custom tool."""
+
+    def __init__(self, config=None):
+        super().__init__(config)
+        self.register_tools_called = False
+        self.registered_mcp = None
+        self.registered_memory = None
+
+    def register_tools(self, mcp: FastMCP, memory: MemoryEngine) -> None:
+        """Register a test tool to verify extension was called."""
+        self.register_tools_called = True
+        self.registered_mcp = mcp
+        self.registered_memory = memory
+
+        @mcp.tool()
+        async def test_extension_tool(query: str) -> str:
+            """A test tool registered by the extension."""
+            return f"Extension tool received: {query}"
+
+
+class TestMCPExtensionBase:
+    """Tests for MCPExtension base class."""
+
+    def test_mcp_extension_is_abstract(self):
+        """MCPExtension.register_tools is abstract and must be implemented."""
+        with pytest.raises(TypeError, match="abstract method"):
+            MCPExtension()
+
+    def test_subclass_can_be_instantiated(self):
+        """Subclass implementing register_tools can be instantiated."""
+        ext = MockMCPExtension()
+        assert ext is not None
+        assert ext.register_tools_called is False
+
+    def test_register_tools_receives_mcp_and_memory(self):
+        """register_tools receives FastMCP and MemoryEngine instances."""
+        ext = MockMCPExtension()
+        mcp = FastMCP("test")
+        memory = MagicMock(spec=MemoryEngine)
+
+        ext.register_tools(mcp, memory)
+
+        assert ext.register_tools_called is True
+        assert ext.registered_mcp is mcp
+        assert ext.registered_memory is memory
+
+
+class TestMCPExtensionLoading:
+    """Tests for MCPExtension loading in create_mcp_server."""
+
+    @pytest.fixture
+    def mock_memory(self):
+        """Create a mock MemoryEngine."""
+        memory = MagicMock()
+        memory._tenant_extension = MagicMock()
+        memory._tenant_extension.authenticate_mcp = MagicMock()
+        return memory
+
+    def test_create_mcp_server_without_extension(self, mock_memory):
+        """create_mcp_server works without MCPExtension configured."""
+        from hindsight_api.api.mcp import create_mcp_server
+
+        with patch("hindsight_api.api.mcp.load_extension", return_value=None):
+            mcp = create_mcp_server(mock_memory)
+
+        # Core tools should be registered
+        tools = mcp._tool_manager._tools
+        assert "retain" in tools
+        assert "recall" in tools
+        assert "reflect" in tools
+        # Extension tool should NOT be present
+        assert "test_extension_tool" not in tools
+
+    def test_create_mcp_server_with_extension(self, mock_memory):
+        """create_mcp_server loads and calls MCPExtension when configured."""
+        from hindsight_api.api.mcp import create_mcp_server
+
+        mock_ext = MockMCPExtension()
+
+        with patch("hindsight_api.api.mcp.load_extension", return_value=mock_ext):
+            mcp = create_mcp_server(mock_memory)
+
+        # Extension should have been called
+        assert mock_ext.register_tools_called is True
+
+        # Core tools should still be registered
+        tools = mcp._tool_manager._tools
+        assert "retain" in tools
+        assert "recall" in tools
+
+        # Extension tool should also be registered
+        assert "test_extension_tool" in tools
+
+    @pytest.mark.asyncio
+    async def test_extension_tool_is_callable(self, mock_memory):
+        """Tool registered by extension can be called."""
+        from hindsight_api.api.mcp import create_mcp_server
+
+        mock_ext = MockMCPExtension()
+
+        with patch("hindsight_api.api.mcp.load_extension", return_value=mock_ext):
+            mcp = create_mcp_server(mock_memory)
+
+        # Get and call the extension tool
+        tools = mcp._tool_manager._tools
+        test_tool = tools["test_extension_tool"]
+        result = await test_tool.fn(query="hello world")
+
+        assert result == "Extension tool received: hello world"
+
+    def test_load_extension_called_with_correct_args(self, mock_memory):
+        """load_extension is called with 'MCP' prefix and MCPExtension class."""
+        from hindsight_api.api.mcp import create_mcp_server
+
+        with patch("hindsight_api.api.mcp.load_extension") as mock_load:
+            mock_load.return_value = None
+            create_mcp_server(mock_memory)
+
+        mock_load.assert_called_once_with("MCP", MCPExtension)
+
+
+class TestMCPExtensionIntegration:
+    """Integration tests verifying extension tools work end-to-end."""
+
+    @pytest.fixture
+    def mock_memory(self):
+        """Create a mock MemoryEngine with required methods."""
+        memory = MagicMock()
+        memory.retain_batch_async = MagicMock()
+        memory.submit_async_retain = MagicMock(return_value={"operation_id": "test-op"})
+        memory.recall_async = MagicMock(return_value=MagicMock(results=[]))
+        memory.reflect_async = MagicMock(return_value=MagicMock(text="reflection"))
+        memory.list_banks = MagicMock(return_value=[])
+        memory.get_bank_profile = MagicMock(return_value={"id": "test"})
+        memory._tenant_extension = MagicMock()
+        return memory
+
+    def test_extension_tools_coexist_with_core_tools(self, mock_memory):
+        """Extension tools are added alongside core tools, not replacing them."""
+        from hindsight_api.api.mcp import create_mcp_server
+
+        mock_ext = MockMCPExtension()
+
+        with patch("hindsight_api.api.mcp.load_extension", return_value=mock_ext):
+            mcp = create_mcp_server(mock_memory)
+
+        tools = mcp._tool_manager._tools
+        # All core tools present
+        assert "retain" in tools
+        assert "recall" in tools
+        assert "reflect" in tools
+        assert "list_banks" in tools
+        assert "create_bank" in tools
+        # Extension tool also present
+        assert "test_extension_tool" in tools
+        # At least 11 core + 1 extension = 12 tools (may grow as new tools are added)
+        assert len(tools) >= 12
@@ -1,8 +1,9 @@
 """Test MCP server routing with dynamic bank_id."""

-import pytest
 from unittest.mock import AsyncMock, MagicMock

+import pytest
+

@pytest.fixture
 def mock_memory():
@@ -17,7 +18,7 @@ def mock_memory():
@pytest.mark.asyncio
 async def test_mcp_context_variable():
    """Test that context variable works correctly."""
-    from hindsight_api.api.mcp import get_current_bank_id, _current_bank_id
+    from hindsight_api.api.mcp import _current_bank_id, get_current_bank_id

    # Initially None
    assert get_current_bank_id() is None
@@ -36,7 +37,7 @@ async def test_mcp_context_variable():
@pytest.mark.asyncio
 async def test_mcp_tools_use_context_bank_id(mock_memory):
    """Test that MCP tools use bank_id from context."""
-    from hindsight_api.api.mcp import create_mcp_server, _current_bank_id
+    from hindsight_api.api.mcp import _current_bank_id, create_mcp_server

    mcp_server = create_mcp_server(mock_memory)

@@ -62,6 +63,7 @@ async def test_mcp_tools_use_context_bank_id(mock_memory):

 def test_path_parsing_logic():
    """Test the path parsing logic for bank_id extraction."""
+
    def parse_path(path):
        """Simulate the path parsing logic from MCPMiddleware."""
        if not path.startswith("/") or len(path) <= 1:
@@ -102,7 +104,7 @@ def test_path_parsing_logic():
@pytest.mark.asyncio
 async def test_api_key_context_variable():
    """Test that API key context variable works correctly."""
-    from hindsight_api.api.mcp import get_current_api_key, _current_api_key
+    from hindsight_api.api.mcp import _current_api_key, get_current_api_key

    # Initially None
    assert get_current_api_key() is None
@@ -121,7 +123,7 @@ async def test_api_key_context_variable():
@pytest.mark.asyncio
 async def test_mcp_tools_propagate_api_key(mock_memory):
    """Test that MCP tools propagate API key to RequestContext."""
-    from hindsight_api.api.mcp import create_mcp_server, _current_bank_id, _current_api_key
+    from hindsight_api.api.mcp import _current_api_key, _current_bank_id, create_mcp_server

    mcp_server = create_mcp_server(mock_memory)
    tools = mcp_server._tool_manager._tools
@@ -141,3 +143,250 @@ async def test_mcp_tools_propagate_api_key(mock_memory):
    finally:
        _current_bank_id.reset(bank_token)
        _current_api_key.reset(api_key_token)
+
+
+@pytest.mark.asyncio
+async def test_tenant_id_context_variable():
+    """Test that tenant_id and api_key_id context variables work correctly."""
+    from hindsight_api.api.mcp import (
+        _current_api_key_id,
+        _current_tenant_id,
+        get_current_api_key_id,
+        get_current_tenant_id,
+    )
+
+    # Initially None
+    assert get_current_tenant_id() is None
+    assert get_current_api_key_id() is None
+
+    # Set and verify
+    tenant_token = _current_tenant_id.set("org-123")
+    key_id_token = _current_api_key_id.set("key-456")
+    try:
+        assert get_current_tenant_id() == "org-123"
+        assert get_current_api_key_id() == "key-456"
+    finally:
+        _current_tenant_id.reset(tenant_token)
+        _current_api_key_id.reset(key_id_token)
+
+    # Back to None after reset
+    assert get_current_tenant_id() is None
+    assert get_current_api_key_id() is None
+
+
+@pytest.mark.asyncio
+async def test_mcp_tools_propagate_tenant_id_and_api_key_id(mock_memory):
+    """Test that MCP tools propagate tenant_id and api_key_id to RequestContext.
+
+    This is the critical test for usage metering: the UsageMeteringValidator reads
+    request_context.tenant_id to identify the org for billing. Without this,
+    MCP operations get tenant_id="unknown" and billing is skipped entirely.
+    """
+    from hindsight_api.api.mcp import (
+        _current_api_key,
+        _current_api_key_id,
+        _current_bank_id,
+        _current_tenant_id,
+        create_mcp_server,
+    )
+
+    mcp_server = create_mcp_server(mock_memory)
+    tools = mcp_server._tool_manager._tools
+
+    # Set all context vars (simulating what MCPMiddleware does after authenticate_mcp)
+    bank_token = _current_bank_id.set("test-bank")
+    api_key_token = _current_api_key.set("hsk_test_key")
+    tenant_token = _current_tenant_id.set("org-billing-123")
+    key_id_token = _current_api_key_id.set("key-uuid-456")
+    try:
+        retain_tool = tools["retain"]
+        await retain_tool.fn(content="test content", context="test_context", async_processing=False)
+
+        # Verify the RequestContext passed to memory engine has all auth fields
+        mock_memory.retain_batch_async.assert_called_once()
+        request_context = mock_memory.retain_batch_async.call_args.kwargs["request_context"]
+        assert request_context.api_key == "hsk_test_key"
+        assert request_context.tenant_id == "org-billing-123"
+        assert request_context.api_key_id == "key-uuid-456"
+    finally:
+        _current_bank_id.reset(bank_token)
+        _current_api_key.reset(api_key_token)
+        _current_tenant_id.reset(tenant_token)
+        _current_api_key_id.reset(key_id_token)
+
+
+def test_multi_bank_mode_exposes_all_tools(mock_memory):
+    """Test that multi-bank mode exposes all tools including bank management and mental models."""
+    from hindsight_api.api.mcp import create_mcp_server
+
+    # Create server in multi-bank mode (default)
+    mcp_server = create_mcp_server(mock_memory, multi_bank=True)
+    tools = mcp_server._tool_manager._tools
+
+    # Core tools
+    assert "retain" in tools
+    assert "recall" in tools
+    assert "reflect" in tools
+    assert "list_banks" in tools
+    assert "create_bank" in tools
+
+    # Mental model tools
+    assert "list_mental_models" in tools
+    assert "get_mental_model" in tools
+    assert "create_mental_model" in tools
+    assert "update_mental_model" in tools
+    assert "delete_mental_model" in tools
+    assert "refresh_mental_model" in tools
+
+
+def test_single_bank_mode_excludes_bank_management_tools(mock_memory):
+    """Test that single-bank mode only exposes bank-scoped tools."""
+    from hindsight_api.api.mcp import create_mcp_server
+
+    # Create server in single-bank mode
+    mcp_server = create_mcp_server(mock_memory, multi_bank=False)
+    tools = mcp_server._tool_manager._tools
+
+    # Should have bank-scoped tools
+    assert "retain" in tools
+    assert "recall" in tools
+    assert "reflect" in tools
+
+    # Mental model tools should also be present (they're bank-scoped)
+    assert "list_mental_models" in tools
+    assert "get_mental_model" in tools
+    assert "create_mental_model" in tools
+    assert "update_mental_model" in tools
+    assert "delete_mental_model" in tools
+    assert "refresh_mental_model" in tools
+
+    # Should NOT have bank management tools
+    assert "list_banks" not in tools
+    assert "create_bank" not in tools
+
+
+def test_multi_bank_mode_tools_have_bank_id_param(mock_memory):
+    """Test that multi-bank mode tools include bank_id parameter."""
+    import inspect
+
+    from hindsight_api.api.mcp import create_mcp_server
+
+    mcp_server = create_mcp_server(mock_memory, multi_bank=True)
+    tools = mcp_server._tool_manager._tools
+
+    # All bank-scoped tools should have bank_id parameter in multi-bank mode
+    bank_scoped_tools = [
+        "retain",
+        "recall",
+        "reflect",
+        "list_mental_models",
+        "get_mental_model",
+        "create_mental_model",
+        "update_mental_model",
+        "delete_mental_model",
+        "refresh_mental_model",
+    ]
+    for tool_name in bank_scoped_tools:
+        tool = tools[tool_name]
+        sig = inspect.signature(tool.fn)
+        assert "bank_id" in sig.parameters, f"{tool_name} should have bank_id param in multi-bank mode"
+
+
+def test_single_bank_mode_tools_no_bank_id_param(mock_memory):
+    """Test that single-bank mode tools do NOT include bank_id parameter."""
+    import inspect
+
+    from hindsight_api.api.mcp import create_mcp_server
+
+    mcp_server = create_mcp_server(mock_memory, multi_bank=False)
+    tools = mcp_server._tool_manager._tools
+
+    # No bank-scoped tool should have bank_id parameter in single-bank mode
+    bank_scoped_tools = [
+        "retain",
+        "recall",
+        "reflect",
+        "list_mental_models",
+        "get_mental_model",
+        "create_mental_model",
+        "update_mental_model",
+        "delete_mental_model",
+        "refresh_mental_model",
+    ]
+    for tool_name in bank_scoped_tools:
+        tool = tools[tool_name]
+        sig = inspect.signature(tool.fn)
+        assert "bank_id" not in sig.parameters, f"{tool_name} should NOT have bank_id param in single-bank mode"
+
+
+@pytest.mark.asyncio
+async def test_middleware_handles_both_endpoints(mock_memory):
+    """Test that MCPMiddleware routes to correct server based on URL path."""
+    from hindsight_api.api.mcp import MCPMiddleware
+
+    # Create middleware (single instance)
+    middleware = MCPMiddleware(None, mock_memory)
+
+    # Verify both server instances exist
+    assert middleware.multi_bank_app is not None
+    assert middleware.single_bank_app is not None
+
+    # Verify they expose different tools
+    multi_bank_tools = middleware.multi_bank_server._tool_manager._tools
+    single_bank_tools = middleware.single_bank_server._tool_manager._tools
+
+    # Multi-bank should have all tools
+    assert "retain" in multi_bank_tools
+    assert "recall" in multi_bank_tools
+    assert "list_banks" in multi_bank_tools
+    assert "create_bank" in multi_bank_tools
+    assert "list_mental_models" in multi_bank_tools
+    assert "create_mental_model" in multi_bank_tools
+
+    # Single-bank should only have scoped tools
+    assert "retain" in single_bank_tools
+    assert "recall" in single_bank_tools
+    assert "list_mental_models" in single_bank_tools
+    assert "create_mental_model" in single_bank_tools
+    assert "list_banks" not in single_bank_tools
+    assert "create_bank" not in single_bank_tools
+
+
+@pytest.mark.asyncio
+async def test_routing_logic_from_url_path():
+    """Test that routing correctly selects server based on URL structure.
+
+    Simulates the path parsing logic from MCPMiddleware.__call__ after the
+    prefix has been stripped. Any first path segment is treated as a bank_id.
+    """
+    from hindsight_api.api.mcp import MCPMiddleware
+
+    # Mock memory
+    mock_memory = MagicMock()
+
+    # Create middleware
+    middleware = MCPMiddleware(None, mock_memory)
+
+    # Simulate different URL patterns and verify routing
+    # Path is what remains after stripping the /mcp prefix
+    test_cases = [
+        # (path_after_prefix_strip, expected_bank_id_from_path, expected_bank_id, description)
+        ("/alice/messages", True, "alice", "Bank ID in path with endpoint"),
+        ("/my-agent-123/", True, "my-agent-123", "Bank ID in path with trailing slash"),
+        ("/sse/", True, "sse", "Bank named 'sse' routes to single-bank"),
+        ("/messages/", True, "messages", "Bank named 'messages' routes to single-bank"),
+        ("/", False, None, "Root path, no bank ID"),
+    ]
+
+    for path, expected_bank_from_path, expected_bank_id, description in test_cases:
+        bank_id = None
+        bank_id_from_path = False
+
+        if path.startswith("/") and len(path) > 1:
+            parts = path[1:].split("/", 1)
+            if parts[0]:
+                bank_id = parts[0]
+                bank_id_from_path = True
+
+        assert bank_id_from_path == expected_bank_from_path, f"Failed for: {description} (path={path})"
+        assert bank_id == expected_bank_id, f"Failed bank_id for: {description} (path={path}, got={bank_id})"
@@ -1,10 +1,17 @@
 """Tests for the shared MCP tools module."""

 from datetime import datetime, timezone
+from unittest.mock import AsyncMock, MagicMock

 import pytest

-from hindsight_api.mcp_tools import build_content_dict, parse_timestamp
+from hindsight_api.mcp_tools import (
+    MCPToolsConfig,
+    _validate_mental_model_inputs,
+    build_content_dict,
+    parse_timestamp,
+    register_mcp_tools,
+)


 class TestParseTimestamp:
@@ -61,3 +68,579 @@ class TestBuildContentDict:
        result, error = build_content_dict("test content", "test_context", None)
        assert error is None
        assert "event_date" not in result
+
+
+# =========================================================================
+# Mental Model MCP Tool Tests
+# =========================================================================
+
+
+@pytest.fixture
+def mock_memory():
+    """Create a mock MemoryEngine with mental model methods."""
+    memory = MagicMock()
+    memory.list_mental_models = AsyncMock(
+        return_value=[
+            {"id": "mm-1", "name": "Coding Prefs", "source_query": "coding preferences?", "content": "Prefers Python"},
+            {"id": "mm-2", "name": "Goals", "source_query": "current goals?", "content": "Ship v2"},
+        ]
+    )
+    memory.get_mental_model = AsyncMock(
+        return_value={
+            "id": "mm-1",
+            "name": "Coding Prefs",
+            "source_query": "coding preferences?",
+            "content": "Prefers Python",
+        }
+    )
+    memory.create_mental_model = AsyncMock(return_value={"id": "mm-new"})
+    memory.submit_async_refresh_mental_model = AsyncMock(return_value={"operation_id": "op-123"})
+    memory.update_mental_model = AsyncMock(
+        return_value={
+            "id": "mm-1",
+            "name": "Updated Name",
+            "source_query": "new query?",
+            "content": "Updated",
+        }
+    )
+    memory.delete_mental_model = AsyncMock(return_value=True)
+    return memory
+
+
+@pytest.fixture
+def mcp_server_with_mental_models(mock_memory):
+    """Create a FastMCP server with mental model tools registered (multi-bank mode)."""
+    from fastmcp import FastMCP
+
+    mcp = FastMCP("test", stateless_http=True)
+    config = MCPToolsConfig(
+        bank_id_resolver=lambda: "test-bank",
+        include_bank_id_param=True,
+        tools={
+            "list_mental_models",
+            "get_mental_model",
+            "create_mental_model",
+            "update_mental_model",
+            "delete_mental_model",
+            "refresh_mental_model",
+        },
+    )
+    register_mcp_tools(mcp, mock_memory, config)
+    return mcp
+
+
+@pytest.fixture
+def mcp_server_single_bank(mock_memory):
+    """Create a FastMCP server with mental model tools registered (single-bank mode)."""
+    from fastmcp import FastMCP
+
+    mcp = FastMCP("test")
+    config = MCPToolsConfig(
+        bank_id_resolver=lambda: "fixed-bank",
+        include_bank_id_param=False,
+        tools={
+            "list_mental_models",
+            "get_mental_model",
+            "create_mental_model",
+            "update_mental_model",
+            "delete_mental_model",
+            "refresh_mental_model",
+        },
+    )
+    register_mcp_tools(mcp, mock_memory, config)
+    return mcp
+
+
+class TestMentalModelToolRegistration:
+    """Test that mental model tools are registered correctly."""
+
+    def test_tools_registered_multi_bank(self, mcp_server_with_mental_models):
+        tools = mcp_server_with_mental_models._tool_manager._tools
+        expected = {
+            "list_mental_models",
+            "get_mental_model",
+            "create_mental_model",
+            "update_mental_model",
+            "delete_mental_model",
+            "refresh_mental_model",
+        }
+        assert expected == set(tools.keys())
+
+    def test_tools_registered_single_bank(self, mcp_server_single_bank):
+        tools = mcp_server_single_bank._tool_manager._tools
+        expected = {
+            "list_mental_models",
+            "get_mental_model",
+            "create_mental_model",
+            "update_mental_model",
+            "delete_mental_model",
+            "refresh_mental_model",
+        }
+        assert expected == set(tools.keys())
+
+    @pytest.mark.asyncio
+    async def test_list_mental_models_propagates_request_context(self, mock_memory):
+        from fastmcp import FastMCP
+
+        mcp = FastMCP("test", stateless_http=True)
+        config = MCPToolsConfig(
+            bank_id_resolver=lambda: "test-bank",
+            api_key_resolver=lambda: "test-api-key",
+            include_bank_id_param=True,
+            tools={"list_mental_models"},
+        )
+        register_mcp_tools(mcp, mock_memory, config)
+        await _tools(mcp)["list_mental_models"].fn()
+        request_context = mock_memory.list_mental_models.call_args.kwargs["request_context"]
+        assert request_context.api_key == "test-api-key"
+
+    @pytest.mark.asyncio
+    async def test_create_mental_model_propagates_request_context(self, mock_memory):
+        from fastmcp import FastMCP
+
+        mcp = FastMCP("test", stateless_http=True)
+        config = MCPToolsConfig(
+            bank_id_resolver=lambda: "test-bank",
+            api_key_resolver=lambda: "test-api-key",
+            include_bank_id_param=True,
+            tools={"create_mental_model"},
+        )
+        register_mcp_tools(mcp, mock_memory, config)
+        await _tools(mcp)["create_mental_model"].fn(name="Test", source_query="query")
+        request_context = mock_memory.create_mental_model.call_args.kwargs["request_context"]
+        assert request_context.api_key == "test-api-key"
+
+    def test_mental_model_tools_in_default_set(self):
+        """Mental model tools should be in the default tools set when config.tools is None."""
+        from fastmcp import FastMCP
+
+        memory = MagicMock()
+        # Mock all engine methods that tools reference
+        memory.retain_batch_async = AsyncMock()
+        memory.submit_async_retain = AsyncMock(return_value={"operation_id": "op"})
+        memory.recall_async = AsyncMock(return_value=MagicMock(results=[]))
+        memory.reflect_async = AsyncMock()
+        memory.list_banks = AsyncMock(return_value=[])
+        memory.get_bank_profile = AsyncMock(return_value={})
+        memory.update_bank = AsyncMock()
+        memory.list_mental_models = AsyncMock(return_value=[])
+        memory.get_mental_model = AsyncMock()
+        memory.create_mental_model = AsyncMock()
+        memory.submit_async_refresh_mental_model = AsyncMock()
+        memory.update_mental_model = AsyncMock()
+        memory.delete_mental_model = AsyncMock()
+
+        mcp = FastMCP("test", stateless_http=True)
+        config = MCPToolsConfig(
+            bank_id_resolver=lambda: "bank",
+            include_bank_id_param=True,
+            tools=None,  # Default - all tools
+        )
+        register_mcp_tools(mcp, memory, config)
+        tools = mcp._tool_manager._tools
+        assert "list_mental_models" in tools
+        assert "create_mental_model" in tools
+        assert "refresh_mental_model" in tools
+
+
+@pytest.fixture
+def no_bank_mcp_server(mock_memory):
+    """Create a multi-bank MCP server where bank_id_resolver returns None."""
+    from fastmcp import FastMCP
+
+    mcp = FastMCP("test", stateless_http=True)
+    config = MCPToolsConfig(
+        bank_id_resolver=lambda: None,
+        include_bank_id_param=True,
+        tools={
+            "list_mental_models",
+            "get_mental_model",
+            "create_mental_model",
+            "update_mental_model",
+            "delete_mental_model",
+            "refresh_mental_model",
+        },
+    )
+    register_mcp_tools(mcp, mock_memory, config)
+    return mcp
+
+
+def _tools(mcp_server):
+    """Helper to get tools dict from MCP server."""
+    return mcp_server._tool_manager._tools
+
+
+@pytest.mark.asyncio
+class TestListMentalModels:
+    async def test_list_multi_bank(self, mcp_server_with_mental_models, mock_memory):
+        result = await _tools(mcp_server_with_mental_models)["list_mental_models"].fn()
+        assert '"mm-1"' in result
+        assert '"mm-2"' in result
+        mock_memory.list_mental_models.assert_called_once()
+        assert mock_memory.list_mental_models.call_args.kwargs["bank_id"] == "test-bank"
+
+    async def test_list_with_bank_id_override(self, mcp_server_with_mental_models, mock_memory):
+        """Explicit bank_id should override the resolver."""
+        await _tools(mcp_server_with_mental_models)["list_mental_models"].fn(bank_id="other-bank")
+        assert mock_memory.list_mental_models.call_args.kwargs["bank_id"] == "other-bank"
+
+    async def test_list_with_tags(self, mcp_server_with_mental_models, mock_memory):
+        await _tools(mcp_server_with_mental_models)["list_mental_models"].fn(tags=["work"])
+        assert mock_memory.list_mental_models.call_args.kwargs["tags"] == ["work"]
+
+    async def test_list_single_bank(self, mcp_server_single_bank, mock_memory):
+        result = await _tools(mcp_server_single_bank)["list_mental_models"].fn()
+        assert isinstance(result, dict)
+        assert len(result["items"]) == 2
+        assert mock_memory.list_mental_models.call_args.kwargs["bank_id"] == "fixed-bank"
+
+    async def test_list_no_bank_returns_error(self, no_bank_mcp_server):
+        result = await _tools(no_bank_mcp_server)["list_mental_models"].fn()
+        assert "error" in result
+
+    async def test_list_engine_error_multi_bank(self, mcp_server_with_mental_models, mock_memory):
+        mock_memory.list_mental_models.side_effect = RuntimeError("DB connection lost")
+        result = await _tools(mcp_server_with_mental_models)["list_mental_models"].fn()
+        assert "error" in result
+        assert "DB connection lost" in result
+
+    async def test_list_engine_error_single_bank(self, mcp_server_single_bank, mock_memory):
+        mock_memory.list_mental_models.side_effect = RuntimeError("DB connection lost")
+        result = await _tools(mcp_server_single_bank)["list_mental_models"].fn()
+        assert isinstance(result, dict)
+        assert "error" in result
+
+
+@pytest.mark.asyncio
+class TestGetMentalModel:
+    async def test_get_multi_bank(self, mcp_server_with_mental_models, mock_memory):
+        result = await _tools(mcp_server_with_mental_models)["get_mental_model"].fn(mental_model_id="mm-1")
+        assert '"mm-1"' in result
+        assert mock_memory.get_mental_model.call_args.kwargs["mental_model_id"] == "mm-1"
+
+    async def test_get_with_bank_id_override(self, mcp_server_with_mental_models, mock_memory):
+        await _tools(mcp_server_with_mental_models)["get_mental_model"].fn(mental_model_id="mm-1", bank_id="other-bank")
+        assert mock_memory.get_mental_model.call_args.kwargs["bank_id"] == "other-bank"
+
+    async def test_get_not_found_multi_bank(self, mcp_server_with_mental_models, mock_memory):
+        mock_memory.get_mental_model.return_value = None
+        result = await _tools(mcp_server_with_mental_models)["get_mental_model"].fn(mental_model_id="missing")
+        assert "not found" in result
+
+    async def test_get_not_found_single_bank(self, mcp_server_single_bank, mock_memory):
+        mock_memory.get_mental_model.return_value = None
+        result = await _tools(mcp_server_single_bank)["get_mental_model"].fn(mental_model_id="missing")
+        assert isinstance(result, dict)
+        assert "not found" in result["error"]
+
+    async def test_get_single_bank(self, mcp_server_single_bank, mock_memory):
+        result = await _tools(mcp_server_single_bank)["get_mental_model"].fn(mental_model_id="mm-1")
+        assert isinstance(result, dict)
+        assert result["id"] == "mm-1"
+
+    async def test_get_no_bank_returns_error(self, no_bank_mcp_server):
+        result = await _tools(no_bank_mcp_server)["get_mental_model"].fn(mental_model_id="mm-1")
+        assert "error" in result
+
+    async def test_get_engine_error(self, mcp_server_with_mental_models, mock_memory):
+        mock_memory.get_mental_model.side_effect = RuntimeError("DB error")
+        result = await _tools(mcp_server_with_mental_models)["get_mental_model"].fn(mental_model_id="mm-1")
+        assert "error" in result
+
+
+@pytest.mark.asyncio
+class TestCreateMentalModel:
+    async def test_create_multi_bank(self, mcp_server_with_mental_models, mock_memory):
+        result = await _tools(mcp_server_with_mental_models)["create_mental_model"].fn(
+            name="Test Model",
+            source_query="What are the user's preferences?",
+        )
+        assert '"mm-new"' in result
+        assert '"op-123"' in result
+        mock_memory.create_mental_model.assert_called_once()
+        call_kwargs = mock_memory.create_mental_model.call_args.kwargs
+        assert call_kwargs["name"] == "Test Model"
+        assert call_kwargs["source_query"] == "What are the user's preferences?"
+        assert call_kwargs["content"] == "Generating content..."
+        # Verify async refresh was scheduled
+        mock_memory.submit_async_refresh_mental_model.assert_called_once()
+        assert mock_memory.submit_async_refresh_mental_model.call_args.kwargs["mental_model_id"] == "mm-new"
+
+    async def test_create_with_custom_id(self, mcp_server_with_mental_models, mock_memory):
+        await _tools(mcp_server_with_mental_models)["create_mental_model"].fn(
+            name="Test", source_query="query", mental_model_id="custom-id"
+        )
+        assert mock_memory.create_mental_model.call_args.kwargs["mental_model_id"] == "custom-id"
+
+    async def test_create_with_tags_and_max_tokens(self, mcp_server_with_mental_models, mock_memory):
+        await _tools(mcp_server_with_mental_models)["create_mental_model"].fn(
+            name="Test", source_query="query", tags=["work", "coding"], max_tokens=4096
+        )
+        call_kwargs = mock_memory.create_mental_model.call_args.kwargs
+        assert call_kwargs["tags"] == ["work", "coding"]
+        assert call_kwargs["max_tokens"] == 4096
+
+    async def test_create_with_bank_id_override(self, mcp_server_with_mental_models, mock_memory):
+        await _tools(mcp_server_with_mental_models)["create_mental_model"].fn(
+            name="Test", source_query="query", bank_id="other-bank"
+        )
+        assert mock_memory.create_mental_model.call_args.kwargs["bank_id"] == "other-bank"
+        assert mock_memory.submit_async_refresh_mental_model.call_args.kwargs["bank_id"] == "other-bank"
+
+    async def test_create_single_bank(self, mcp_server_single_bank, mock_memory):
+        result = await _tools(mcp_server_single_bank)["create_mental_model"].fn(name="Test", source_query="query")
+        assert isinstance(result, dict)
+        assert result["mental_model_id"] == "mm-new"
+        assert result["operation_id"] == "op-123"
+
+    async def test_create_no_bank_returns_error(self, no_bank_mcp_server):
+        result = await _tools(no_bank_mcp_server)["create_mental_model"].fn(name="Test", source_query="query")
+        assert "error" in result
+
+    async def test_create_value_error_multi_bank(self, mcp_server_with_mental_models, mock_memory):
+        """ValueError from engine (e.g. invalid ID format) should return error, not crash."""
+        mock_memory.create_mental_model.side_effect = ValueError("ID must be alphanumeric lowercase")
+        result = await _tools(mcp_server_with_mental_models)["create_mental_model"].fn(
+            name="Test", source_query="query", mental_model_id="INVALID!!"
+        )
+        assert "alphanumeric" in result
+
+    async def test_create_value_error_single_bank(self, mcp_server_single_bank, mock_memory):
+        mock_memory.create_mental_model.side_effect = ValueError("ID must be alphanumeric lowercase")
+        result = await _tools(mcp_server_single_bank)["create_mental_model"].fn(
+            name="Test", source_query="query", mental_model_id="INVALID!!"
+        )
+        assert isinstance(result, dict)
+        assert "alphanumeric" in result["error"]
+
+    async def test_create_engine_error(self, mcp_server_with_mental_models, mock_memory):
+        mock_memory.create_mental_model.side_effect = RuntimeError("DB error")
+        result = await _tools(mcp_server_with_mental_models)["create_mental_model"].fn(
+            name="Test", source_query="query"
+        )
+        assert "error" in result
+
+
+@pytest.mark.asyncio
+class TestUpdateMentalModel:
+    async def test_update_multi_bank(self, mcp_server_with_mental_models, mock_memory):
+        result = await _tools(mcp_server_with_mental_models)["update_mental_model"].fn(
+            mental_model_id="mm-1", name="Updated Name"
+        )
+        assert '"Updated Name"' in result
+        call_kwargs = mock_memory.update_mental_model.call_args.kwargs
+        assert call_kwargs["name"] == "Updated Name"
+        assert call_kwargs["source_query"] is None  # Not updated
+
+    async def test_update_multiple_fields(self, mcp_server_with_mental_models, mock_memory):
+        await _tools(mcp_server_with_mental_models)["update_mental_model"].fn(
+            mental_model_id="mm-1", name="New Name", source_query="new query?", tags=["updated"], max_tokens=4096
+        )
+        call_kwargs = mock_memory.update_mental_model.call_args.kwargs
+        assert call_kwargs["name"] == "New Name"
+        assert call_kwargs["source_query"] == "new query?"
+        assert call_kwargs["tags"] == ["updated"]
+        assert call_kwargs["max_tokens"] == 4096
+
+    async def test_update_with_bank_id_override(self, mcp_server_with_mental_models, mock_memory):
+        await _tools(mcp_server_with_mental_models)["update_mental_model"].fn(
+            mental_model_id="mm-1", name="X", bank_id="other-bank"
+        )
+        assert mock_memory.update_mental_model.call_args.kwargs["bank_id"] == "other-bank"
+
+    async def test_update_not_found_multi_bank(self, mcp_server_with_mental_models, mock_memory):
+        mock_memory.update_mental_model.return_value = None
+        result = await _tools(mcp_server_with_mental_models)["update_mental_model"].fn(
+            mental_model_id="missing", name="X"
+        )
+        assert "not found" in result
+
+    async def test_update_single_bank(self, mcp_server_single_bank, mock_memory):
+        result = await _tools(mcp_server_single_bank)["update_mental_model"].fn(mental_model_id="mm-1", name="Updated")
+        assert isinstance(result, dict)
+        assert mock_memory.update_mental_model.call_args.kwargs["bank_id"] == "fixed-bank"
+
+    async def test_update_not_found_single_bank(self, mcp_server_single_bank, mock_memory):
+        mock_memory.update_mental_model.return_value = None
+        result = await _tools(mcp_server_single_bank)["update_mental_model"].fn(mental_model_id="missing", name="X")
+        assert isinstance(result, dict)
+        assert "not found" in result["error"]
+
+    async def test_update_no_bank_returns_error(self, no_bank_mcp_server):
+        result = await _tools(no_bank_mcp_server)["update_mental_model"].fn(mental_model_id="mm-1", name="X")
+        assert "error" in result
+
+    async def test_update_engine_error(self, mcp_server_with_mental_models, mock_memory):
+        mock_memory.update_mental_model.side_effect = RuntimeError("DB error")
+        result = await _tools(mcp_server_with_mental_models)["update_mental_model"].fn(mental_model_id="mm-1", name="X")
+        assert "error" in result
+
+
+@pytest.mark.asyncio
+class TestDeleteMentalModel:
+    async def test_delete_multi_bank(self, mcp_server_with_mental_models, mock_memory):
+        result = await _tools(mcp_server_with_mental_models)["delete_mental_model"].fn(mental_model_id="mm-1")
+        assert '"deleted"' in result
+        assert mock_memory.delete_mental_model.call_args.kwargs["mental_model_id"] == "mm-1"
+
+    async def test_delete_with_bank_id_override(self, mcp_server_with_mental_models, mock_memory):
+        await _tools(mcp_server_with_mental_models)["delete_mental_model"].fn(
+            mental_model_id="mm-1", bank_id="other-bank"
+        )
+        assert mock_memory.delete_mental_model.call_args.kwargs["bank_id"] == "other-bank"
+
+    async def test_delete_not_found_multi_bank(self, mcp_server_with_mental_models, mock_memory):
+        mock_memory.delete_mental_model.return_value = False
+        result = await _tools(mcp_server_with_mental_models)["delete_mental_model"].fn(mental_model_id="missing")
+        assert "not found" in result
+
+    async def test_delete_not_found_single_bank(self, mcp_server_single_bank, mock_memory):
+        mock_memory.delete_mental_model.return_value = False
+        result = await _tools(mcp_server_single_bank)["delete_mental_model"].fn(mental_model_id="missing")
+        assert isinstance(result, dict)
+        assert "not found" in result["error"]
+
+    async def test_delete_single_bank(self, mcp_server_single_bank, mock_memory):
+        result = await _tools(mcp_server_single_bank)["delete_mental_model"].fn(mental_model_id="mm-1")
+        assert isinstance(result, dict)
+        assert result["status"] == "deleted"
+
+    async def test_delete_no_bank_returns_error(self, no_bank_mcp_server):
+        result = await _tools(no_bank_mcp_server)["delete_mental_model"].fn(mental_model_id="mm-1")
+        assert "error" in result
+
+    async def test_delete_engine_error(self, mcp_server_with_mental_models, mock_memory):
+        mock_memory.delete_mental_model.side_effect = RuntimeError("DB error")
+        result = await _tools(mcp_server_with_mental_models)["delete_mental_model"].fn(mental_model_id="mm-1")
+        assert "error" in result
+
+
+@pytest.mark.asyncio
+class TestRefreshMentalModel:
+    async def test_refresh_multi_bank(self, mcp_server_with_mental_models, mock_memory):
+        result = await _tools(mcp_server_with_mental_models)["refresh_mental_model"].fn(mental_model_id="mm-1")
+        assert '"op-123"' in result
+        assert '"queued"' in result
+
+    async def test_refresh_with_bank_id_override(self, mcp_server_with_mental_models, mock_memory):
+        await _tools(mcp_server_with_mental_models)["refresh_mental_model"].fn(
+            mental_model_id="mm-1", bank_id="other-bank"
+        )
+        assert mock_memory.submit_async_refresh_mental_model.call_args.kwargs["bank_id"] == "other-bank"
+
+    async def test_refresh_not_found_multi_bank(self, mcp_server_with_mental_models, mock_memory):
+        mock_memory.submit_async_refresh_mental_model.side_effect = ValueError("Mental model 'missing' not found")
+        result = await _tools(mcp_server_with_mental_models)["refresh_mental_model"].fn(mental_model_id="missing")
+        assert "not found" in result
+
+    async def test_refresh_not_found_single_bank(self, mcp_server_single_bank, mock_memory):
+        mock_memory.submit_async_refresh_mental_model.side_effect = ValueError("not found")
+        result = await _tools(mcp_server_single_bank)["refresh_mental_model"].fn(mental_model_id="missing")
+        assert isinstance(result, dict)
+        assert "not found" in result["error"]
+
+    async def test_refresh_single_bank(self, mcp_server_single_bank, mock_memory):
+        result = await _tools(mcp_server_single_bank)["refresh_mental_model"].fn(mental_model_id="mm-1")
+        assert isinstance(result, dict)
+        assert result["operation_id"] == "op-123"
+
+    async def test_refresh_no_bank_returns_error(self, no_bank_mcp_server):
+        result = await _tools(no_bank_mcp_server)["refresh_mental_model"].fn(mental_model_id="mm-1")
+        assert "error" in result
+
+    async def test_refresh_engine_error(self, mcp_server_with_mental_models, mock_memory):
+        mock_memory.submit_async_refresh_mental_model.side_effect = RuntimeError("DB error")
+        result = await _tools(mcp_server_with_mental_models)["refresh_mental_model"].fn(mental_model_id="mm-1")
+        assert "error" in result
+
+
+class TestValidateMentalModelInputs:
+    """Tests for the _validate_mental_model_inputs helper."""
+
+    def test_valid_inputs(self):
+        assert _validate_mental_model_inputs(name="Test", source_query="query", max_tokens=2048) is None
+
+    def test_none_inputs(self):
+        assert _validate_mental_model_inputs() is None
+
+    def test_empty_name(self):
+        result = _validate_mental_model_inputs(name="")
+        assert result == "name cannot be empty"
+
+    def test_whitespace_name(self):
+        result = _validate_mental_model_inputs(name="   ")
+        assert result == "name cannot be empty"
+
+    def test_empty_source_query(self):
+        result = _validate_mental_model_inputs(source_query="")
+        assert result == "source_query cannot be empty"
+
+    def test_whitespace_source_query(self):
+        result = _validate_mental_model_inputs(source_query="  \t  ")
+        assert result == "source_query cannot be empty"
+
+    def test_max_tokens_too_low(self):
+        result = _validate_mental_model_inputs(max_tokens=0)
+        assert "max_tokens must be between 256 and 8192" in result
+
+    def test_max_tokens_too_high(self):
+        result = _validate_mental_model_inputs(max_tokens=10000)
+        assert "max_tokens must be between 256 and 8192" in result
+
+    def test_max_tokens_at_lower_bound(self):
+        assert _validate_mental_model_inputs(max_tokens=256) is None
+
+    def test_max_tokens_at_upper_bound(self):
+        assert _validate_mental_model_inputs(max_tokens=8192) is None
+
+
+@pytest.mark.asyncio
+class TestMentalModelInputValidation:
+    """Tests that validation is applied in create/update tools before engine calls."""
+
+    async def test_create_empty_name_returns_error_multi_bank(self, mcp_server_with_mental_models, mock_memory):
+        result = await _tools(mcp_server_with_mental_models)["create_mental_model"].fn(name="", source_query="query")
+        assert "name cannot be empty" in result
+        mock_memory.create_mental_model.assert_not_called()
+
+    async def test_create_empty_source_query_returns_error_multi_bank(self, mcp_server_with_mental_models, mock_memory):
+        result = await _tools(mcp_server_with_mental_models)["create_mental_model"].fn(name="Test", source_query="")
+        assert "source_query cannot be empty" in result
+        mock_memory.create_mental_model.assert_not_called()
+
+    async def test_create_max_tokens_too_low_multi_bank(self, mcp_server_with_mental_models, mock_memory):
+        result = await _tools(mcp_server_with_mental_models)["create_mental_model"].fn(
+            name="Test", source_query="query", max_tokens=0
+        )
+        assert "max_tokens must be between 256 and 8192" in result
+        mock_memory.create_mental_model.assert_not_called()
+
+    async def test_create_max_tokens_too_high_single_bank(self, mcp_server_single_bank, mock_memory):
+        result = await _tools(mcp_server_single_bank)["create_mental_model"].fn(
+            name="Test", source_query="query", max_tokens=10000
+        )
+        assert isinstance(result, dict)
+        assert "max_tokens must be between 256 and 8192" in result["error"]
+        mock_memory.create_mental_model.assert_not_called()
+
+    async def test_update_empty_name_returns_error_multi_bank(self, mcp_server_with_mental_models, mock_memory):
+        result = await _tools(mcp_server_with_mental_models)["update_mental_model"].fn(mental_model_id="mm-1", name="")
+        assert "name cannot be empty" in result
+        mock_memory.update_mental_model.assert_not_called()
+
+    async def test_update_empty_name_returns_error_single_bank(self, mcp_server_single_bank, mock_memory):
+        result = await _tools(mcp_server_single_bank)["update_mental_model"].fn(mental_model_id="mm-1", name="  ")
+        assert isinstance(result, dict)
+        assert "name cannot be empty" in result["error"]
+        mock_memory.update_mental_model.assert_not_called()
+
+    async def test_not_found_error_includes_bank_id_multi_bank(self, mcp_server_with_mental_models, mock_memory):
+        mock_memory.get_mental_model.return_value = None
+        result = await _tools(mcp_server_with_mental_models)["get_mental_model"].fn(mental_model_id="missing")
+        assert "test-bank" in result
+
+    async def test_not_found_error_includes_bank_id_single_bank(self, mcp_server_single_bank, mock_memory):
+        mock_memory.get_mental_model.return_value = None
+        result = await _tools(mcp_server_single_bank)["get_mental_model"].fn(mental_model_id="missing")
+        assert isinstance(result, dict)
+        assert "fixed-bank" in result["error"]
--- a/Show More
+++ b/Show More