fix: hindsight-embed profiles are not loaded correctly

feat: slim docker distro (#314 )
2026-02-06 17:13:39 +01:00 · 2026-02-06 16:52:08 +01:00 · 2026-02-06 15:00:24 +01:00 · 2026-02-06 13:56:59 +01:00 · 2026-02-06 10:50:45 +01:00 · 2026-02-06 10:49:13 +01:00
295 changed files with 29155 additions and 6020 deletions
@@ -139,7 +139,7 @@ jobs:
        path: hindsight-clients/typescript/*.tgz
        retention-days: 1

-  release-moltbot-integration:
+  release-openclaw-integration:
    runs-on: ubuntu-latest
    environment: npm

@@ -153,15 +153,15 @@ jobs:
        registry-url: 'https://registry.npmjs.org'

    - name: Install dependencies
-      working-directory: ./hindsight-integrations/moltbot
+      working-directory: ./hindsight-integrations/openclaw
      run: npm ci

    - name: Build
-      working-directory: ./hindsight-integrations/moltbot
+      working-directory: ./hindsight-integrations/openclaw
      run: npm run build

    - name: Publish to npm
-      working-directory: ./hindsight-integrations/moltbot
+      working-directory: ./hindsight-integrations/openclaw
      run: |
        set +e
        OUTPUT=$(npm publish --access public 2>&1)
@@ -178,14 +178,63 @@ jobs:
        NODE_AUTH_TOKEN: ${{ secrets.NPM_TOKEN }}

    - name: Pack for GitHub release
-      working-directory: ./hindsight-integrations/moltbot
+      working-directory: ./hindsight-integrations/openclaw
      run: npm pack

    - name: Upload artifacts
      uses: actions/upload-artifact@v4
      with:
-        name: moltbot-integration
-        path: hindsight-integrations/moltbot/*.tgz
+        name: openclaw-integration
+        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:
@@ -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-moltbot-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

@@ -438,11 +510,17 @@ jobs:
        name: typescript-client
        path: ./artifacts/typescript-client

-    - name: Download Moltbot Integration
+    - name: Download OpenClaw Integration
      uses: actions/download-artifact@v4
      with:
-        name: moltbot-integration
-        path: ./artifacts/moltbot-integration
+        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
@@ -485,8 +563,10 @@ jobs:
        cp artifacts/python-packages/hindsight-embed/dist/* release-assets/ || true
        # TypeScript client
        cp artifacts/typescript-client/*.tgz release-assets/ || true
-        # Moltbot Integration
-        cp artifacts/moltbot-integration/*.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
@@ -82,7 +51,7 @@ jobs:
    - name: Build TypeScript client
      run: npm run build --workspace=hindsight-clients/typescript

-  build-moltbot-integration:
+  build-openclaw-integration:
    runs-on: ubuntu-latest

    steps:
@@ -94,15 +63,38 @@ jobs:
        node-version: '22'

    - name: Install dependencies
-      working-directory: ./hindsight-integrations/moltbot
+      working-directory: ./hindsight-integrations/openclaw
      run: npm ci

    - name: Run tests
-      working-directory: ./hindsight-integrations/moltbot
+      working-directory: ./hindsight-integrations/openclaw
      run: npm test

    - name: Build
-      working-directory: ./hindsight-integrations/moltbot
+      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:
@@ -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 +324,30 @@ 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
+        cache-from: type=gha,scope=${{ matrix.name }}
+        cache-to: type=gha,mode=max,scope=${{ matrix.name }}

-    # 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 +770,9 @@ 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 +803,62 @@ 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 +1031,18 @@ 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 +1113,52 @@ 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
@@ -93,6 +93,34 @@ uv run ty check hindsight_api  # Type check
 3. Run tests to ensure nothing breaks
 4. Submit a PR with a clear description of changes

+## Release Process
+
+The project uses `scripts/release.sh` for creating releases. This script automates the entire release workflow:
+
+1. Bumps version in all components (API, clients, CLI, control plane, Helm)
+2. **Regenerates OpenAPI spec and client SDKs** (Python, TypeScript, Rust)
+3. Updates documentation versioning
+4. Creates a commit and git tag
+5. Pushes to GitHub (triggers CI/CD to publish packages)
+
+### Usage
+
+```bash
+./scripts/release.sh <version>
+```
+
+**Example:**
+```bash
+./scripts/release.sh 0.5.0
+```
+
+### Important for Developers
+
+- During development, version bumps in `__init__.py` do NOT require client regeneration
+- Clients are only regenerated during releases
+- Do not manually run `./scripts/generate-clients.sh` unless testing generation changes
+- Client version comments will reflect the API version from the latest release
+
 ## Reporting Issues

 Open an issue on GitHub with:
@@ -59,7 +59,7 @@ docker run --rm -it --pull always -p 8888:8888 -p 9999:9999 \

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

 Install client:
@@ -0,0 +1,135 @@
+# Docker Testing
+
+Scripts for testing Hindsight Docker images locally and in CI.
+
+## Scripts
+
+### `test-image.sh`
+
+General-purpose Docker image test script. Starts a container and verifies it becomes healthy.
+
+**Usage:**
+```bash
+./docker/test-image.sh <image> [target]
+```
+
+**Arguments:**
+- `image` - Docker image to test (e.g., `hindsight:test`, `ghcr.io/vectorize-io/hindsight:latest`)
+- `target` - Optional: `cp-only` for control plane, `api-only` for API, or `standalone` (default)
+
+**Environment Variables:**
+- `GROQ_API_KEY` - Required for API/standalone images
+- `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 (for slim images)
+- `HINDSIGHT_API_EMBEDDINGS_OPENAI_API_KEY` - OpenAI API key for embeddings
+- `HINDSIGHT_API_RERANKER_PROVIDER` - Reranker provider (for slim images)
+- `HINDSIGHT_API_COHERE_API_KEY` - Cohere API key for reranking
+- `SMOKE_TEST_TIMEOUT` - Timeout in seconds (default: 120)
+
+**Examples:**
+
+Test a full image (with local ML models):
+```bash
+export GROQ_API_KEY=gsk_xxx
+./docker/test-image.sh hindsight:test
+```
+
+Test a slim image (with external providers):
+```bash
+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
+```
+
+### `test-slim-local.sh`
+
+Convenience wrapper for testing slim images locally. Automatically configures external providers.
+
+**Usage:**
+```bash
+# Set API keys
+export GROQ_API_KEY=gsk_xxx
+export OPENAI_API_KEY=sk-xxx
+export COHERE_API_KEY=xxx
+
+# Run test
+./docker/test-slim-local.sh [image]
+```
+
+**Or inline:**
+```bash
+GROQ_API_KEY=gsk_xxx \
+OPENAI_API_KEY=sk-xxx \
+COHERE_API_KEY=xxx \
+./docker/test-slim-local.sh hindsight-slim:test
+```
+
+This script:
+- ✅ Validates API keys are set
+- ✅ Configures OpenAI embeddings automatically
+- ✅ Configures Cohere reranking automatically
+- ✅ Calls `test-image.sh` with the right configuration
+
+## Building and Testing Locally
+
+### Build a slim image
+
+```bash
+docker build \
+  --build-arg INCLUDE_LOCAL_MODELS=false \
+  --build-arg PRELOAD_ML_MODELS=false \
+  --target standalone \
+  -t hindsight-slim:test \
+  -f docker/standalone/Dockerfile \
+  .
+```
+
+### Test the slim image
+
+```bash
+# With API keys
+export GROQ_API_KEY=gsk_xxx
+export OPENAI_API_KEY=sk-xxx
+export COHERE_API_KEY=xxx
+
+# Run test
+./docker/test-slim-local.sh hindsight-slim:test
+```
+
+## Expected Output
+
+**Successful test:**
+```
+Starting smoke test for: hindsight-slim:test
+  Target: standalone
+  Health endpoint: http://localhost:8888/health
+  Timeout: 120s
+
+Starting container...
+Waiting for health endpoint at http://localhost:8888/health...
+  Still waiting... (10s)
+  Still waiting... (20s)
+
+Container is healthy after 25s
+
+=== Health Response ===
+{
+    "status": "healthy",
+    "database": "connected"
+}
+
+Smoke test PASSED
+```
+
+## CI Integration
+
+These scripts are used in CI to validate Docker images on every PR:
+
+- `.github/workflows/test.yml` - Runs `test-image.sh` for slim variants with OpenAI/Cohere
+- `.github/workflows/release.yml` - Can optionally run smoke tests during release
+
+See the workflows for the exact configuration.
@@ -169,16 +169,34 @@ ENV PATH="/app/api/.venv/bin:${PATH}"

 # 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
 ARG PRELOAD_ML_MODELS
 ARG INCLUDE_LOCAL_MODELS
+ENV HF_HUB_DOWNLOAD_TIMEOUT=600
 RUN if [ "$PRELOAD_ML_MODELS" = "true" ] && [ "$INCLUDE_LOCAL_MODELS" = "true" ]; then \
-    /app/api/.venv/bin/python -c "\
+    MAX_RETRIES=3; \
+    RETRY_DELAY=10; \
+    for i in $(seq 1 $MAX_RETRIES); do \
+      echo "Attempt $i/$MAX_RETRIES: Downloading ML models..."; \
+      /app/api/.venv/bin/python -c "\
+import os; os.environ['HF_HUB_DOWNLOAD_TIMEOUT'] = '600'; \
 from sentence_transformers import SentenceTransformer, CrossEncoder; \
 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('Models cached successfully')"; \
+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..."; \
+        sleep $RETRY_DELAY; \
+        RETRY_DELAY=$((RETRY_DELAY * 2)); \
+      fi; \
+    done; \
+    if [ $i -eq $MAX_RETRIES ] && ! /app/api/.venv/bin/python -c "from sentence_transformers import SentenceTransformer; SentenceTransformer('BAAI/bge-small-en-v1.5')" 2>/dev/null; then \
+      echo "ERROR: Failed to download models after $MAX_RETRIES attempts"; \
+      exit 1; \
+    fi; \
    elif [ "$INCLUDE_LOCAL_MODELS" != "true" ]; then echo "Skipping ML model preload (local-models not included)"; \
    else echo "Skipping ML model preload"; fi

@@ -190,6 +208,10 @@ 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"]

@@ -277,16 +299,34 @@ ENV PATH="/app/api/.venv/bin:${PATH}"

 # 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
 ARG PRELOAD_ML_MODELS
 ARG INCLUDE_LOCAL_MODELS
+ENV HF_HUB_DOWNLOAD_TIMEOUT=600
 RUN if [ "$PRELOAD_ML_MODELS" = "true" ] && [ "$INCLUDE_LOCAL_MODELS" = "true" ]; then \
-    /app/api/.venv/bin/python -c "\
+    MAX_RETRIES=3; \
+    RETRY_DELAY=10; \
+    for i in $(seq 1 $MAX_RETRIES); do \
+      echo "Attempt $i/$MAX_RETRIES: Downloading ML models..."; \
+      /app/api/.venv/bin/python -c "\
+import os; os.environ['HF_HUB_DOWNLOAD_TIMEOUT'] = '600'; \
 from sentence_transformers import SentenceTransformer, CrossEncoder; \
 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('Models cached successfully')"; \
+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..."; \
+        sleep $RETRY_DELAY; \
+        RETRY_DELAY=$((RETRY_DELAY * 2)); \
+      fi; \
+    done; \
+    if [ $i -eq $MAX_RETRIES ] && ! /app/api/.venv/bin/python -c "from sentence_transformers import SentenceTransformer; SentenceTransformer('BAAI/bge-small-en-v1.5')" 2>/dev/null; then \
+      echo "ERROR: Failed to download models after $MAX_RETRIES attempts"; \
+      exit 1; \
+    fi; \
    elif [ "$INCLUDE_LOCAL_MODELS" != "true" ]; then echo "Skipping ML model preload (local-models not included)"; \
    else echo "Skipping ML model preload"; fi

@@ -300,6 +340,10 @@ 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: 0.4.2
-appVersion: "0.4.2"
+version: 0.4.9
+appVersion: "0.4.9"
 keywords:
  - ai
  - memory
@@ -46,4 +46,4 @@ __all__ = [
    "RemoteTEICrossEncoder",
    "LLMConfig",
 ]
-__version__ = "0.4.2"
+__version__ = "0.4.9"
@@ -11,6 +11,7 @@ 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.
@@ -23,8 +24,21 @@ depends_on: str | Sequence[str] | None = None
 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(
@@ -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)
+    """)
@@ -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
@@ -35,7 +36,7 @@ from pydantic import BaseModel, ConfigDict, Field, field_validator

 from hindsight_api import MemoryEngine
 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 +46,8 @@ 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."""
@@ -520,7 +523,9 @@ 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
@@ -585,7 +590,7 @@ 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 +618,9 @@ 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.",
@@ -863,6 +870,7 @@ class ListDocumentsResponse(BaseModel):
                        "updated_at": "2024-01-15T10:30:00Z",
                        "text_length": 5420,
                        "memory_unit_count": 15,
+                        "tags": ["user_a", "session_123"],
                    }
                ],
                "total": 50,
@@ -1110,7 +1118,9 @@ class MentalModelResponse(BaseModel):
    bank_id: str
    name: str
    source_query: str
-    content: str
+    content: str = Field(
+        description="The mental model content as well-formatted markdown (auto-generated from reflect endpoint)"
+    )
    tags: list[str] = Field(default_factory=list)
    max_tokens: int = Field(default=2048)
    trigger: MentalModelTrigger = Field(default_factory=MentalModelTrigger)
@@ -1134,6 +1144,7 @@ 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"],
@@ -1143,6 +1154,9 @@ 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")
@@ -1153,7 +1167,8 @@ class CreateMentalModelRequest(BaseModel):
 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):
@@ -1398,14 +1413,19 @@ def create_app(

        # Start worker poller if enabled (standalone mode)
        if config.worker_enabled and memory._pool is not None:
+            from ..config import DEFAULT_DATABASE_SCHEMA
+
            worker_id = config.worker_id or socket.gethostname()
+            # Convert default schema to None for SQL compatibility (no schema prefix)
+            schema = None if config.database_schema == DEFAULT_DATABASE_SCHEMA else config.database_schema
            poller = WorkerPoller(
                pool=memory._pool,
                worker_id=worker_id,
                executor=memory.execute_task,
                poll_interval_ms=config.worker_poll_interval_ms,
                max_retries=config.worker_max_retries,
-                tenant_extension=getattr(memory, "_tenant_extension", None),
+                schema=schema,
+                tenant_extension=memory._tenant_extension,
                max_slots=config.worker_max_slots,
                consolidation_max_slots=config.worker_consolidation_max_slots,
            )
@@ -1713,6 +1733,15 @@ 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)
@@ -1827,6 +1856,15 @@ 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

@@ -1884,17 +1922,17 @@ 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(
@@ -2285,6 +2323,23 @@ def _register_routes(app: FastAPI):
    ):
        """Get a mental model by ID."""
        try:
+            # Pre-operation validation hook
+            validator = app.state.memory._operation_validator
+            if validator:
+                from hindsight_api.extensions.operation_validator import MentalModelGetContext
+
+                ctx = MentalModelGetContext(
+                    bank_id=bank_id,
+                    mental_model_id=mental_model_id,
+                    request_context=request_context,
+                )
+                validation = await validator.validate_mental_model_get(ctx)
+                if not validation.allowed:
+                    raise OperationValidationError(
+                        validation.reason or "Operation not allowed",
+                        status_code=validation.status_code,
+                    )
+
            mental_model = await app.state.memory.get_mental_model(
                bank_id=bank_id,
                mental_model_id=mental_model_id,
@@ -2292,9 +2347,31 @@ def _register_routes(app: FastAPI):
            )
            if mental_model is None:
                raise HTTPException(status_code=404, detail=f"Mental model '{mental_model_id}' not found")
+
+            # Post-operation hook
+            if validator:
+                from hindsight_api.extensions.operation_validator import MentalModelGetResult
+
+                content = mental_model.get("content", "")
+                output_tokens = len(content) // 4 if content else 0
+
+                result_ctx = MentalModelGetResult(
+                    bank_id=bank_id,
+                    mental_model_id=mental_model_id,
+                    request_context=request_context,
+                    output_tokens=output_tokens,
+                    success=True,
+                )
+                try:
+                    await validator.on_mental_model_get_complete(result_ctx)
+                except Exception as hook_err:
+                    logger.warning(f"Post-mental-model-get hook error (non-fatal): {hook_err}")
+
            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

@@ -2319,12 +2396,30 @@ def _register_routes(app: FastAPI):
    ):
        """Create a mental model (async - returns operation_id)."""
        try:
+            # Pre-operation validation hook
+            validator = app.state.memory._operation_validator
+            if validator:
+                from hindsight_api.extensions.operation_validator import MentalModelRefreshContext
+
+                ctx = MentalModelRefreshContext(
+                    bank_id=bank_id,
+                    mental_model_id=None,  # Not yet created
+                    request_context=request_context,
+                )
+                validation = await validator.validate_mental_model_refresh(ctx)
+                if not validation.allowed:
+                    raise OperationValidationError(
+                        validation.reason or "Operation not allowed",
+                        status_code=validation.status_code,
+                    )
+
            # 1. Create the mental model with placeholder content
            mental_model = await app.state.memory.create_mental_model(
                bank_id=bank_id,
                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,
@@ -2336,11 +2431,13 @@ 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

@@ -2363,6 +2460,23 @@ def _register_routes(app: FastAPI):
    ):
        """Refresh a mental model by re-running its source query (async)."""
        try:
+            # Pre-operation validation hook
+            validator = app.state.memory._operation_validator
+            if validator:
+                from hindsight_api.extensions.operation_validator import MentalModelRefreshContext
+
+                ctx = MentalModelRefreshContext(
+                    bank_id=bank_id,
+                    mental_model_id=mental_model_id,
+                    request_context=request_context,
+                )
+                validation = await validator.validate_mental_model_refresh(ctx)
+                if not validation.allowed:
+                    raise OperationValidationError(
+                        validation.reason or "Operation not allowed",
+                        status_code=validation.status_code,
+                    )
+
            result = await app.state.memory.submit_async_refresh_mental_model(
                bank_id=bank_id,
                mental_model_id=mental_model_id,
@@ -2373,6 +2487,8 @@ 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

@@ -83,9 +83,12 @@ 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('LLM:')} {color(f'{llm_provider} / {llm_model}', 0.6)}")
@@ -154,7 +154,21 @@ 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
@@ -303,6 +317,11 @@ 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."""
@@ -431,14 +450,18 @@ class HindsightConfig:
    @classmethod
    def from_env(cls) -> "HindsightConfig":
        """Create configuration from environment variables."""
+        # 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)
+
        return cls(
            # Database
            database_url=os.getenv(ENV_DATABASE_URL, DEFAULT_DATABASE_URL),
            database_schema=os.getenv(ENV_DATABASE_SCHEMA, DEFAULT_DATABASE_SCHEMA),
            # 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 +476,12 @@ 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 +498,12 @@ 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 +522,12 @@ 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)
@@ -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
@@ -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
@@ -385,7 +430,8 @@ async def _process_memory(
    memory_id = memory["id"]
    fact_tags = memory.get("tags") or []

-    # Find related observations using the full recall system (NO tag filtering)
+    # 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,
@@ -393,6 +439,7 @@ async def _process_memory(
        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)
@@ -666,17 +713,20 @@ 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
    """
@@ -685,14 +735,19 @@ async def _find_related_observations(
    from ...config import get_config

    config = get_config()
+
+    # SECURITY: Use all_strict matching if tags provided to prevent cross-scope consolidation
+    tags_match = "all_strict" if tags else "any"
+
    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
-        # NO tags parameter - intentionally get ALL observations
    )

    # If no observations returned, return empty list
@@ -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

@@ -162,11 +163,28 @@ 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},
+                )
+            finally:
+                # Restore original logging level
+                transformers_logger.setLevel(original_level)

        # Initialize shared executor (limited workers naturally limits concurrency)
        if LocalSTCrossEncoder._executor is None:
@@ -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
@@ -157,11 +158,28 @@ 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},
+                )
+            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})")
@@ -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
@@ -303,8 +303,10 @@ class MemoryEngine(MemoryEngineInterface):
        db_url = db_url or config.database_url
        memory_llm_provider = memory_llm_provider or config.llm_provider
        memory_llm_api_key = memory_llm_api_key or config.llm_api_key
-        # Ollama and mock don't require an API key
-        if not memory_llm_api_key and memory_llm_provider not in ("ollama", "mock"):
+        # Ollama, openai-codex, claude-code, and mock don't require an API key
+        # openai-codex uses OAuth tokens from ~/.codex/auth.json
+        # claude-code uses OAuth tokens from macOS Keychain
+        if not memory_llm_api_key and memory_llm_provider not in ("ollama", "openai-codex", "claude-code", "mock"):
            raise ValueError("LLM API key is required. Set HINDSIGHT_API_LLM_API_KEY environment variable.")
        memory_llm_model = memory_llm_model or config.llm_model
        memory_llm_base_url = memory_llm_base_url or config.get_llm_base_url() or None
@@ -457,7 +459,11 @@ class MemoryEngine(MemoryEngineInterface):
        # Store operation validator extension (optional)
        self._operation_validator = operation_validator

-        # Store tenant extension (optional)
+        # Store tenant extension (always set, use default if none provided)
+        if tenant_extension is None:
+            from ..extensions.builtin.tenant import DefaultTenantExtension
+
+            tenant_extension = DefaultTenantExtension(config={})
        self._tenant_extension = tenant_extension

    async def _validate_operation(self, validation_coro) -> None:
@@ -495,22 +501,18 @@ class MemoryEngine(MemoryEngineInterface):
        Raises:
            AuthenticationError: If authentication fails or request_context is missing when required.
        """
-        if self._tenant_extension is None:
-            _current_schema.set("public")
-            return "public"
-
        from hindsight_api.extensions import AuthenticationError

        if request_context is None:
-            raise AuthenticationError("RequestContext is required when tenant extension is configured")
+            raise AuthenticationError("RequestContext is required")

        # For internal/background operations (e.g., worker tasks), skip extension authentication.
        # The task was already authenticated at submission time, and execute_task sets _current_schema
-        # from the task's _schema field. For public schema tasks, _current_schema keeps its default "public".
+        # from the task's _schema field.
        if request_context.internal:
            return _current_schema.get()

-        # Let AuthenticationError propagate - HTTP layer will convert to 401
+        # Authenticate through tenant extension (always set, may be default no-auth extension)
        tenant_context = await self._tenant_extension.authenticate(request_context)

        _current_schema.set(tenant_context.schema_name)
@@ -536,10 +538,15 @@ class MemoryEngine(MemoryEngineInterface):
            f"[BATCH_RETAIN_TASK] Starting background batch retain for bank_id={bank_id}, {len(contents)} items"
        )

-        # Use internal request context for background tasks (skips tenant auth when schema is pre-set)
+        # Restore tenant_id/api_key_id from task payload so downstream operations
+        # (e.g., consolidation and mental model refreshes) can attribute usage.
        from hindsight_api.models import RequestContext

-        internal_context = RequestContext(internal=True)
+        internal_context = RequestContext(
+            internal=True,
+            tenant_id=task_dict.get("_tenant_id"),
+            api_key_id=task_dict.get("_api_key_id"),
+        )
        await self.retain_batch_async(bank_id=bank_id, contents=contents, request_context=internal_context)

        logger.info(f"[BATCH_RETAIN_TASK] Completed background batch retain for bank_id={bank_id}")
@@ -565,7 +572,13 @@ class MemoryEngine(MemoryEngineInterface):

        from .consolidation import run_consolidation_job

-        internal_context = RequestContext(internal=True)
+        # Restore tenant_id/api_key_id from task payload so downstream operations
+        # (e.g., mental model refreshes) can attribute usage to the correct org.
+        internal_context = RequestContext(
+            internal=True,
+            tenant_id=task_dict.get("_tenant_id"),
+            api_key_id=task_dict.get("_api_key_id"),
+        )
        result = await run_consolidation_job(
            memory_engine=self,
            bank_id=bank_id,
@@ -597,7 +610,13 @@ class MemoryEngine(MemoryEngineInterface):

        from hindsight_api.models import RequestContext

-        internal_context = RequestContext(internal=True)
+        # Restore tenant_id/api_key_id from task payload so extensions can
+        # attribute the mental_model_refresh operation to the correct org.
+        internal_context = RequestContext(
+            internal=True,
+            tenant_id=task_dict.get("_tenant_id"),
+            api_key_id=task_dict.get("_api_key_id"),
+        )

        # Get the current mental model to get source_query
        mental_model = await self.get_mental_model(bank_id, mental_model_id, request_context=internal_context)
@@ -606,11 +625,19 @@ class MemoryEngine(MemoryEngineInterface):

        source_query = mental_model["source_query"]

+        # SECURITY: If the mental model has tags, pass them to reflect with "all_strict" matching
+        # to ensure it can only access other mental models/memories with the SAME tags.
+        # This prevents cross-tenant/cross-user information leakage by excluding untagged content.
+        tags = mental_model.get("tags")
+        tags_match = "all_strict" if tags else "any"
+
        # Run reflect to generate new content, excluding the mental model being refreshed
        reflect_result = await self.reflect_async(
            bank_id=bank_id,
            query=source_query,
            request_context=internal_context,
+            tags=tags,
+            tags_match=tags_match,
            exclude_mental_model_ids=[mental_model_id],
        )

@@ -641,6 +668,42 @@ class MemoryEngine(MemoryEngineInterface):
            request_context=internal_context,
        )

+        # Call post-operation hook if validator is configured
+        if self._operation_validator:
+            from hindsight_api.extensions.operation_validator import MentalModelRefreshResult
+
+            # Count facts and mental models from based_on
+            facts_used = 0
+            mental_models_used = 0
+            if reflect_result.based_on:
+                for fact_type, facts in reflect_result.based_on.items():
+                    if facts:
+                        if fact_type == "mental_models":
+                            mental_models_used += len(facts)
+                        else:
+                            facts_used += len(facts)
+
+            # Estimate tokens
+            query_tokens = len(source_query) // 4 if source_query else 0
+            output_tokens = len(generated_content) // 4 if generated_content else 0
+            context_tokens = 0  # refresh doesn't use additional context
+
+            result_ctx = MentalModelRefreshResult(
+                bank_id=bank_id,
+                mental_model_id=mental_model_id,
+                request_context=internal_context,
+                query_tokens=query_tokens,
+                output_tokens=output_tokens,
+                context_tokens=context_tokens,
+                facts_used=facts_used,
+                mental_models_used=mental_models_used,
+                success=True,
+            )
+            try:
+                await self._operation_validator.on_mental_model_refresh_complete(result_ctx)
+            except Exception as hook_err:
+                logger.warning(f"Post-mental-model-refresh hook error (non-fatal): {hook_err}")
+
        logger.info(f"[REFRESH_MENTAL_MODEL_TASK] Completed for bank_id={bank_id}, mental_model_id={mental_model_id}")

    async def execute_task(self, task_dict: dict[str, Any]):
@@ -884,30 +947,34 @@ class MemoryEngine(MemoryEngineInterface):

            if not self.db_url:
                raise ValueError("Database URL is required for migrations")
+
+            # Migrate all schemas from the tenant extension
+            # The tenant extension is the single source of truth for which schemas exist
            logger.info("Running database migrations...")
-            # Use configured database schema for migrations (defaults to "public")
-            run_migrations(self.db_url, schema=get_config().database_schema)
+            try:
+                tenants = await self._tenant_extension.list_tenants()
+                if tenants:
+                    logger.info(f"Running migrations on {len(tenants)} schema(s)...")
+                    for tenant in tenants:
+                        schema = tenant.schema
+                        if schema:
+                            try:
+                                run_migrations(self.db_url, schema=schema)
+                            except Exception as e:
+                                logger.warning(f"Failed to migrate schema {schema}: {e}")
+                    logger.info("Schema migrations completed")

-            # Migrate all existing tenant schemas (if multi-tenant)
-            if self._tenant_extension is not None:
-                try:
-                    tenants = await self._tenant_extension.list_tenants()
-                    if tenants:
-                        logger.info(f"Running migrations on {len(tenants)} tenant schemas...")
-                        for tenant in tenants:
-                            schema = tenant.schema
-                            if schema and schema != "public":
-                                try:
-                                    run_migrations(self.db_url, schema=schema)
-                                except Exception as e:
-                                    logger.warning(f"Failed to migrate tenant schema {schema}: {e}")
-                        logger.info("Tenant schema migrations completed")
-                except Exception as e:
-                    logger.warning(f"Failed to run tenant schema migrations: {e}")
-
-            # Ensure embedding column dimension matches the model's dimension
-            # This is done after migrations and after embeddings.initialize()
-            ensure_embedding_dimension(self.db_url, self.embeddings.dimension, schema=get_config().database_schema)
+                    # Ensure embedding column dimension matches the model's dimension
+                    # This is done after migrations and after embeddings.initialize()
+                    for tenant in tenants:
+                        schema = tenant.schema
+                        if schema:
+                            try:
+                                ensure_embedding_dimension(self.db_url, self.embeddings.dimension, schema=schema)
+                            except Exception as e:
+                                logger.warning(f"Failed to ensure embedding dimension for schema {schema}: {e}")
+            except Exception as e:
+                logger.warning(f"Failed to run schema migrations: {e}")

        logger.info(f"Connecting to PostgreSQL at {self.db_url}")

@@ -3245,7 +3312,8 @@ class MemoryEngine(MemoryEngineInterface):
                    created_at,
                    updated_at,
                    LENGTH(original_text) as text_length,
-                    retain_params
+                    retain_params,
+                    tags
                FROM {fq_table("documents")}
                {where_clause}
                ORDER BY created_at DESC
@@ -3301,6 +3369,7 @@ class MemoryEngine(MemoryEngineInterface):
                        "text_length": row["text_length"] or 0,
                        "memory_unit_count": unit_count,
                        "retain_params": row["retain_params"] if row["retain_params"] else None,
+                        "tags": row["tags"] if row["tags"] else [],
                    }
                )

@@ -3604,12 +3673,14 @@ class MemoryEngine(MemoryEngineInterface):

        # Load directives from the dedicated directives table
        # Directives are hard rules that must be followed in all responses
+        # Use isolation_mode=True to prevent tag-scoped directives from leaking into untagged operations
        directives_raw = await self.list_directives(
            bank_id=bank_id,
            tags=tags,
            tags_match=tags_match,
            active_only=True,
            request_context=request_context,
+            isolation_mode=True,
        )
        # Convert directive format to the expected format for reflect agent
        # The agent expects: name, description (optional), observations (list of {title, content})
@@ -3678,7 +3749,16 @@ class MemoryEngine(MemoryEngineInterface):
        # Extract memories from recall tool outputs - only include memories the agent actually used
        # agent_result.used_memory_ids contains validated IDs from the done action
        used_memory_ids_set = set(agent_result.used_memory_ids) if agent_result.used_memory_ids else set()
-        based_on: dict[str, list[MemoryFact]] = {"world": [], "experience": [], "opinion": [], "observation": []}
+        # based_on stores facts, mental models, and directives
+        # Note: directives list stores raw directive dicts (not MemoryFact), which will be converted to Directive objects
+        based_on: dict[str, list[MemoryFact] | list[dict[str, Any]]] = {
+            "world": [],
+            "experience": [],
+            "opinion": [],
+            "observation": [],
+            "mental-models": [],
+            "directives": [],
+        }
        seen_memory_ids: set[str] = set()
        for tc in agent_result.tool_trace:
            if tc.tool == "recall" and "memories" in tc.output:
@@ -3780,38 +3860,15 @@ class MemoryEngine(MemoryEngineInterface):
                        )
                # List all models lookup - don't add to based_on (too verbose, just a listing)

-        # Add directives to based_on["mental-models"] (they are mental models with subtype='directive')
-        for directive in directives:
-            # Extract summary from observations
-            summary_parts: list[str] = []
-            for obs in directive.get("observations", []):
-                # Support both Pydantic Observation objects and dicts
-                if hasattr(obs, "content"):
-                    content = obs.content
-                    title = obs.title
-                else:
-                    content = obs.get("content", "")
-                    title = obs.get("title", "")
-                if title and content:
-                    summary_parts.append(f"{title}: {content}")
-                elif content:
-                    summary_parts.append(content)
-
-            # Fallback to description if no observations
-            if not summary_parts and directive.get("description"):
-                summary_parts.append(directive["description"])
-
-            directive_name = directive.get("name", "")
-            directive_summary = "; ".join(summary_parts) if summary_parts else ""
-            based_on["mental-models"].append(
-                MemoryFact(
-                    id=directive.get("id", ""),
-                    text=f"{directive_name}: {directive_summary}",
-                    fact_type="mental-models",
-                    context="directive (directive)",
-                    occurred_start=None,
-                    occurred_end=None,
-                )
+        # Add directives to based_on["directives"]
+        # Store raw directive dicts (with id, name, content) for http.py to convert to ReflectDirective
+        for directive_raw in directives_raw:
+            based_on["directives"].append(
+                {
+                    "id": directive_raw["id"],
+                    "name": directive_raw["name"],
+                    "content": directive_raw["content"],
+                }
            )

        # Build directives_applied from agent result
@@ -4660,11 +4717,19 @@ class MemoryEngine(MemoryEngineInterface):
        if not mental_model:
            return None

+        # SECURITY: If the mental model has tags, pass them to reflect with "all_strict" matching
+        # to ensure it can only access other mental models/memories with the SAME tags.
+        # This prevents cross-tenant/cross-user information leakage by excluding untagged content.
+        tags = mental_model.get("tags")
+        tags_match = "all_strict" if tags else "any"
+
        # Run reflect with the source query, excluding the mental model being refreshed
        reflect_result = await self.reflect_async(
            bank_id=bank_id,
            query=mental_model["source_query"],
            request_context=request_context,
+            tags=tags,
+            tags_match=tags_match,
            exclude_mental_model_ids=[mental_model_id],
        )

@@ -4677,6 +4742,7 @@ class MemoryEngine(MemoryEngineInterface):
                        "id": str(fact.id),
                        "text": fact.text,
                        "type": fact_type,
+                        "context": fact.context,  # Include context to distinguish directives from mental models in UI
                    }
                    for fact in facts
                ]
@@ -4865,6 +4931,7 @@ class MemoryEngine(MemoryEngineInterface):
        limit: int = 100,
        offset: int = 0,
        request_context: "RequestContext",
+        isolation_mode: bool = False,
    ) -> list[dict[str, Any]]:
        """List directives for a bank.

@@ -4876,6 +4943,9 @@ class MemoryEngine(MemoryEngineInterface):
            limit: Maximum number of results
            offset: Offset for pagination
            request_context: Request context for authentication
+            isolation_mode: When True and tags=None, only return directives with no tags.
+                This prevents tag-scoped directives from leaking into untagged operations.
+                Default False (normal API behavior - returns all directives when tags=None)

        Returns:
            List of directive dicts
@@ -4885,6 +4955,8 @@ class MemoryEngine(MemoryEngineInterface):

        async with acquire_with_retry(pool) as conn:
            # Build filters
+            from .search.tags import build_tags_where_clause
+
            filters = ["bank_id = $1"]
            params: list[Any] = [bank_id]
            param_idx = 2
@@ -4892,15 +4964,23 @@ class MemoryEngine(MemoryEngineInterface):
            if active_only:
                filters.append("is_active = TRUE")

+            # Apply tags filter:
+            # - If tags provided: use standard filtering (with strict modes support)
+            # - If tags=None and isolation_mode=True: only include directives with NO tags
+            #   (prevents tag-scoped directives from leaking into untagged reflect/refresh)
+            # - If tags=None and isolation_mode=False: no filtering (normal API behavior)
            if tags:
-                if tags_match == "all":
-                    filters.append(f"tags @> ${param_idx}::varchar[]")
-                elif tags_match == "exact":
-                    filters.append(f"tags = ${param_idx}::varchar[]")
-                else:  # any
-                    filters.append(f"tags && ${param_idx}::varchar[]")
-                params.append(tags)
-                param_idx += 1
+                tags_clause, tags_params, param_idx = build_tags_where_clause(
+                    tags=tags, param_offset=param_idx, table_alias="", match=tags_match
+                )
+                if tags_clause:
+                    # Remove leading "AND " from clause since we're building filters list
+                    filters.append(tags_clause.replace("AND ", "", 1))
+                    params.extend(tags_params)
+            elif isolation_mode:
+                # Isolation mode: only include directives with empty/null tags
+                # This ensures tag-scoped directives don't apply to untagged operations
+                filters.append("(tags IS NULL OR tags = '{}')")

            params.extend([limit, offset])

@@ -5416,6 +5496,13 @@ class MemoryEngine(MemoryEngineInterface):
        task_payload: dict[str, Any] = {"contents": contents}
        if document_tags:
            task_payload["document_tags"] = document_tags
+        # Pass tenant_id and api_key_id through task payload so the worker
+        # can propagate request context to downstream operations (e.g.,
+        # consolidation and mental model refreshes triggered after retain).
+        if request_context.tenant_id:
+            task_payload["_tenant_id"] = request_context.tenant_id
+        if request_context.api_key_id:
+            task_payload["_api_key_id"] = request_context.api_key_id

        result = await self._submit_async_operation(
            bank_id=bank_id,
@@ -5448,11 +5535,21 @@ class MemoryEngine(MemoryEngineInterface):
            Dict with operation_id
        """
        await self._authenticate_tenant(request_context)
+
+        # Pass tenant_id and api_key_id through task payload so the worker
+        # can provide request context to extension hooks (e.g., usage metering
+        # for mental model refreshes triggered by consolidation).
+        task_payload: dict[str, Any] = {}
+        if request_context.tenant_id:
+            task_payload["_tenant_id"] = request_context.tenant_id
+        if request_context.api_key_id:
+            task_payload["_api_key_id"] = request_context.api_key_id
+
        return await self._submit_async_operation(
            bank_id=bank_id,
            operation_type="consolidation",
            task_type="consolidation",
-            task_payload={},
+            task_payload=task_payload,
            dedupe_by_bank=True,
        )

@@ -5482,13 +5579,21 @@ class MemoryEngine(MemoryEngineInterface):
        if not mental_model:
            raise ValueError(f"Mental model {mental_model_id} not found in bank {bank_id}")

+        # Pass tenant_id and api_key_id through task payload so the worker
+        # can provide request context to extension hooks.
+        task_payload: dict[str, Any] = {
+            "mental_model_id": mental_model_id,
+        }
+        if request_context.tenant_id:
+            task_payload["_tenant_id"] = request_context.tenant_id
+        if request_context.api_key_id:
+            task_payload["_api_key_id"] = request_context.api_key_id
+
        return await self._submit_async_operation(
            bank_id=bank_id,
            operation_type="refresh_mental_model",
            task_type="refresh_mental_model",
-            task_payload={
-                "mental_model_id": mental_model_id,
-            },
+            task_payload=task_payload,
            result_metadata={"mental_model_id": mental_model_id, "name": mental_model["name"]},
            dedupe_by_bank=False,
        )
@@ -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 +1,434 @@
+"""
+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="test",
+                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,
+                )
+
+                # 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()
+                metrics.record_llm_call(
+                    provider=self.provider,
+                    model=self.model,
+                    scope=scope,
+                    duration=time.time() - start_time,
+                    input_tokens=input_tokens,
+                    output_tokens=output_tokens,
+                    success=True,
+                )
+
+                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 +1,493 @@
+"""
+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="test",
+                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,
+                )
+
+                # 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 +1,578 @@
+"""
+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,
+            )
+            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,
+                )
+
+                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,
+            )
+
+            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 +1,502 @@
+"""
+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,
+            )
+            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,
+                )
+
+                # 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,
+                )
+
+                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 +1,234 @@
+"""
+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
+
+    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}")
+
+        # 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)
+
+        if self._mock_response is not None:
+            if isinstance(self._mock_response, LLMToolCallResult):
+                return self._mock_response
+            # Allow setting just tool calls as a list
+            if isinstance(self._mock_response, list):
+                return 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",
+                )
+
+        return LLMToolCallResult(content="mock response", finish_reason="stop")
+
+    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 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."""
+        self._mock_calls = []
@@ -0,0 +1,745 @@
+"""
+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,
+            )
+            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,
+                )
+
+                # 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,
+                )
+
+                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()
@@ -871,21 +871,21 @@ 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 +904,18 @@ 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",
@@ -57,21 +57,25 @@ def _infer_temporal_date(fact_text: str, event_date: datetime) -> str | None:
    return None


-def _sanitize_text(text: str) -> str:
+def _sanitize_text(text: str | None) -> str | None:
    """
-    Sanitize text by removing invalid Unicode surrogate characters.
+    Sanitize text by removing characters that break downstream systems.

-    Surrogate characters (U+D800 to U+DFFF) are used in UTF-16 encoding
-    but cannot be encoded in UTF-8. They can appear in Python strings
-    from improperly decoded data (e.g., from JavaScript or broken files).
+    Removes:
+    - Null bytes (\\x00): Invalid in PostgreSQL UTF-8 encoding
+    - Unicode surrogates (U+D800-U+DFFF): Invalid in UTF-8, break LLM APIs

-    This function removes unpaired surrogates to prevent UnicodeEncodeError
-    when the text is sent to the LLM API.
+    Surrogate characters are used in UTF-16 encoding but cannot be encoded
+    in UTF-8. They can appear in Python strings from improperly decoded data
+    (e.g., from JavaScript or broken files). Null bytes commonly appear in
+    OCR output, PDF extraction, or copy-paste from binary sources.
    """
+    if text is None:
+        return None
    if not text:
        return text
-    # Remove surrogate characters (U+D800 to U+DFFF) using regex
-    # These are invalid in UTF-8 and cause encoding errors
+    # Remove null bytes and surrogate characters
+    text = text.replace("\x00", "")
    return re.sub(r"[\ud800-\udfff]", "", text)


@@ -538,7 +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(
@@ -637,6 +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)
@@ -759,6 +769,10 @@ async def _extract_facts_from_chunk(

    # Build user message with metadata and chunk content in a clear format
    # Format event_date with day of week for better temporal reasoning
+    # Handle both datetime objects and ISO string formats (from deserialized async tasks)
+    from .orchestrator import parse_datetime_flexible
+
+    event_date = parse_datetime_flexible(event_date)
    event_date_formatted = event_date.strftime("%A, %B %d, %Y")  # e.g., "Monday, June 10, 2024"
    user_message = f"""Extract facts from the following text chunk.

@@ -1339,31 +1353,26 @@ 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.
    """
-    # Group facts by content_index
-    current_content_idx = 0
-    content_fact_start = 0
+    from .orchestrator import parse_datetime_flexible

    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
+        # Use absolute position across all facts to ensure uniqueness across different contents
+        offset = timedelta(seconds=i * SECONDS_PER_FACT)

-        # Calculate position within this content
-        fact_position = i - content_fact_start
-        offset = timedelta(seconds=fact_position * SECONDS_PER_FACT)
-
-        # Apply offset to all temporal fields
+        # Apply offset to all temporal fields (handle both datetime objects and ISO strings)
        if fact.occurred_start:
-            fact.occurred_start = fact.occurred_start + offset
+            fact.occurred_start = parse_datetime_flexible(fact.occurred_start) + offset
        if fact.occurred_end:
-            fact.occurred_end = fact.occurred_end + offset
+            fact.occurred_end = parse_datetime_flexible(fact.occurred_end) + offset
        if fact.mentioned_at:
-            fact.mentioned_at = fact.mentioned_at + offset
+            fact.mentioned_at = parse_datetime_flexible(fact.mentioned_at) + offset
@@ -8,6 +8,7 @@ import json
 import logging

 from ..memory_engine import fq_table
+from .fact_extraction import _sanitize_text
 from .types import ProcessedFact

 logger = logging.getLogger(__name__)
@@ -47,7 +48,7 @@ async def insert_facts_batch(
    tags_list = []

    for fact in facts:
-        fact_texts.append(fact.fact_text)
+        fact_texts.append(_sanitize_text(fact.fact_text))
        # Convert embedding to string for asyncpg vector type
        embeddings.append(str(fact.embedding))
        # event_date: Use occurred_start if available, otherwise use mentioned_at
@@ -56,7 +57,7 @@ async def insert_facts_batch(
        occurred_starts.append(fact.occurred_start)
        occurred_ends.append(fact.occurred_end)
        mentioned_ats.append(fact.mentioned_at)
-        contexts.append(fact.context)
+        contexts.append(_sanitize_text(fact.context))
        fact_types.append(fact.fact_type)
        # confidence_score is only for opinion facts
        confidence_scores.append(1.0 if fact.fact_type == "opinion" else None)
@@ -157,7 +158,8 @@ async def handle_document_tracking(
    """
    import hashlib

-    # Calculate content hash
+    # Sanitize and calculate content hash
+    combined_content = _sanitize_text(combined_content) or ""
    content_hash = hashlib.sha256(combined_content.encode()).hexdigest()

    # Always delete old document first if it exists (cascades to units and links)
@@ -8,6 +8,7 @@ import logging
 import time
 import uuid
 from datetime import UTC, datetime
+from typing import Any

 from ..db_utils import acquire_with_retry
 from . import bank_utils
@@ -18,6 +19,39 @@ def utcnow():
    return datetime.now(UTC)


+def parse_datetime_flexible(value: Any) -> datetime:
+    """
+    Parse a datetime value that could be either a datetime object or an ISO string.
+
+    This handles datetime values from both direct Python calls and deserialized JSON
+    (where datetime objects are serialized as ISO strings).
+
+    Args:
+        value: Either a datetime object or an ISO format string
+
+    Returns:
+        datetime object (timezone-aware)
+
+    Raises:
+        TypeError: If value is neither datetime nor string
+        ValueError: If string is not a valid ISO datetime
+    """
+    if isinstance(value, datetime):
+        # Ensure timezone-aware
+        if value.tzinfo is None:
+            return value.replace(tzinfo=UTC)
+        return value
+    elif isinstance(value, str):
+        # Parse ISO format string (handles both 'Z' and '+00:00' timezone formats)
+        dt = datetime.fromisoformat(value.replace("Z", "+00:00"))
+        # Ensure timezone-aware
+        if dt.tzinfo is None:
+            return dt.replace(tzinfo=UTC)
+        return dt
+    else:
+        raise TypeError(f"Expected datetime or string, got {type(value).__name__}")
+
+
 from ..response_models import TokenUsage
 from . import (
    chunk_storage,
@@ -89,10 +123,18 @@ async def retain_batch(
        # Merge item-level tags with document-level tags
        item_tags = item.get("tags", []) or []
        merged_tags = list(set(item_tags + (document_tags or [])))
+
+        # Handle event_date: parse flexibly (handles both datetime objects and ISO strings)
+        event_date_value = item.get("event_date")
+        if event_date_value:
+            event_date_value = parse_datetime_flexible(event_date_value)
+        else:
+            event_date_value = utcnow()
+
        content = RetainContent(
            content=item["content"],
            context=item.get("context", ""),
-            event_date=item.get("event_date") or utcnow(),
+            event_date=event_date_value,
            metadata=item.get("metadata", {}),
            entities=item.get("entities", []),
            tags=merged_tags,
@@ -116,6 +158,13 @@ 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]
@@ -130,7 +179,7 @@ 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
@@ -144,6 +193,13 @@ 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]
@@ -158,7 +214,7 @@ 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
@@ -210,6 +266,13 @@ 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"):
@@ -224,7 +287,7 @@ 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
@@ -252,6 +315,13 @@ 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:
@@ -274,7 +344,7 @@ async def retain_batch(
                                combined_content,
                                is_first_batch,
                                retain_params,
-                                document_tags,
+                                merged_tags,
                            )
                            document_ids_added.append(actual_doc_id)

@@ -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(
@@ -182,7 +182,16 @@ class BrokerTaskBackend(TaskBackend):
        operation_id = task_dict.get("operation_id")
        task_type = task_dict.get("type", "unknown")
        bank_id = task_dict.get("bank_id")
-        payload_json = json.dumps(task_dict)
+
+        # Custom encoder to handle datetime objects
+        from datetime import datetime
+
+        def datetime_encoder(obj):
+            if isinstance(obj, datetime):
+                return obj.isoformat()
+            raise TypeError(f"Object of type {type(obj).__name__} is not JSON serializable")
+
+        payload_json = json.dumps(task_dict, default=datetime_encoder)

        schema = self._schema_getter() if self._schema_getter else self._schema
        table = fq_table("async_operations", schema)
@@ -24,6 +24,11 @@ from hindsight_api.extensions.operation_validator import (
    # Consolidation operation
    ConsolidateContext,
    ConsolidateResult,
+    # Mental Model operations
+    MentalModelGetContext,
+    MentalModelGetResult,
+    MentalModelRefreshContext,
+    MentalModelRefreshResult,
    # Core operations
    OperationValidationError,
    OperationValidatorExtension,
@@ -65,6 +70,11 @@ __all__ = [
    # Operation Validator - Consolidation
    "ConsolidateContext",
    "ConsolidateResult",
+    # Operation Validator - Mental Model
+    "MentalModelGetContext",
+    "MentalModelGetResult",
+    "MentalModelRefreshContext",
+    "MentalModelRefreshResult",
    # Tenant/Auth
    "ApiKeyTenantExtension",
    "AuthenticationError",
@@ -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.
@@ -196,6 +196,57 @@ 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 +453,81 @@ 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
@@ -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 .daemon import (
    DEFAULT_DAEMON_PORT,
    DEFAULT_IDLE_TIMEOUT,
-    DaemonLock,
    IdleTimeoutMiddleware,
    daemonize,
 )
@@ -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()
@@ -362,6 +346,7 @@ 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
@@ -165,6 +165,81 @@ 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:
@@ -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()
@@ -200,15 +207,20 @@ def main():
        if tenant_extension:
            print("Tenant extension loaded - schemas will be discovered dynamically on each poll")
        else:
-            print("No tenant extension configured, using public schema only")
+            print(f"No tenant extension configured, using schema: {config.database_schema}")

        # Create a single poller that handles all schemas dynamically
+        # Convert default schema to None for SQL compatibility (no schema prefix)
+        from hindsight_api.config import DEFAULT_DATABASE_SCHEMA
+
+        schema = None if config.database_schema == DEFAULT_DATABASE_SCHEMA else config.database_schema
        poller = WorkerPoller(
            pool=memory._pool,
            worker_id=args.worker_id,
            executor=memory.execute_task,
            poll_interval_ms=args.poll_interval,
            max_retries=args.max_retries,
+            schema=schema,
            tenant_extension=tenant_extension,
            max_slots=config.worker_max_slots,
            consolidation_max_slots=config.worker_consolidation_max_slots,
@@ -217,15 +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(
@@ -244,7 +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
@@ -99,13 +106,12 @@ class WorkerPoller:
        self._in_flight_by_type: dict[str, int] = {}

    async def _get_schemas(self) -> list[str | None]:
-        """Get list of schemas to poll. Returns [None] for public schema."""
-        if self._tenant_extension is not None:
-            tenants = await self._tenant_extension.list_tenants()
-            # Convert "public" to None for SQL compatibility, keep others as-is
-            return [t.schema if t.schema != "public" else None for t in tenants]
-        # Single schema mode
-        return [self._schema]
+        """Get list of schemas to poll. Returns [None] for default schema (no prefix)."""
+        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]

    async def _get_available_slots(self) -> tuple[int, int]:
        """
@@ -194,7 +200,9 @@ class WorkerPoller:
        try:
            return await self._claim_batch_for_schema_inner(schema, limit, consolidation_limit)
        except Exception as e:
-            logger.warning(f"Worker {self._worker_id} failed to claim tasks for schema {schema or 'public'}: {e}")
+            # Format schema for logging: custom schemas in quotes, None as-is
+            schema_display = f'"{schema}"' if schema else str(schema)
+            logger.warning(f"Worker {self._worker_id} failed to claim tasks for schema {schema_display}: {e}")
            return []

    async def _claim_batch_for_schema_inner(
@@ -418,7 +426,9 @@ class WorkerPoller:
                count = int(result.split()[-1]) if result else 0
                total_count += count
            except Exception as e:
-                logger.warning(f"Worker {self._worker_id} failed to recover tasks for schema {schema or 'public'}: {e}")
+                # Format schema for logging: custom schemas in quotes, None as-is
+                schema_display = f'"{schema}"' if schema else str(schema)
+                logger.warning(f"Worker {self._worker_id} failed to recover tasks for schema {schema_display}: {e}")

        if total_count > 0:
            logger.info(f"Worker {self._worker_id} recovered {total_count} stale tasks from previous run")
@@ -457,7 +467,8 @@ class WorkerPoller:
                            consolidation_count += 1

                    types_str = ", ".join(f"{k}:{v}" for k, v in task_types.items())
-                    schemas_str = ", ".join(s or "public" for s in schemas_seen)
+                    # Display None as "default" in logs
+                    schemas_str = ", ".join(s if s else "default" for s in schemas_seen)
                    logger.info(
                        f"Worker {self._worker_id} claimed {len(tasks)} tasks "
                        f"({consolidation_count} consolidation): {types_str} (schemas: {schemas_str})"
@@ -591,7 +602,8 @@ class WorkerPoller:
                    other_workers.append(f"{wid}:{cnt}")
            others_str = ", ".join(other_workers) if other_workers else "none"

-            schemas_str = ", ".join(s or "public" for s in schemas)
+            # Display None as "default" in logs
+            schemas_str = ", ".join(s if s else "default" for s in schemas)
            logger.info(
                f"[WORKER_STATS] worker={self._worker_id} "
                f"slots={in_flight}/{self._max_slots} (consolidation={consolidation_count}/{self._consolidation_max_slots}) | "
@@ -4,7 +4,7 @@ build-backend = "hatchling.build"

 [project]
 name = "hindsight-api"
-version = "0.4.2"
+version = "0.4.9"
 description = "Hindsight: Agent Memory That Works Like Human Memory"
 readme = "README.md"
 requires-python = ">=3.11"
@@ -25,7 +25,7 @@ dependencies = [
    "psycopg2-binary>=2.9.11",
    "tiktoken>=0.12.0",
    "httpx>=0.27.0",
-    "fastmcp>=2.14.0",  # CVE-2025-66416
+    "fastmcp>=2.14.0", # CVE-2025-66416
    "pg0-embedded>=0.11.0",
    "python-dateutil>=2.8.0",
    "opentelemetry-api>=1.20.0",
@@ -41,16 +41,17 @@ 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 +1,109 @@
+"""
+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.
@@ -1098,3 +1098,134 @@ async def test_version_endpoint_returns_correct_version(api_client):
    assert isinstance(features["worker"], bool)

    print(f"Version endpoint returned: api_version={result['api_version']}, features={features}")
+
+
+@pytest.mark.asyncio
+async def test_retain_with_timestamp_async(api_client, test_bank_id):
+    """Test that async retain accepts timestamp field and serializes correctly."""
+    response = await api_client.post(
+        f"/v1/default/banks/{test_bank_id}/memories",
+        json={
+            "items": [
+                {
+                    "content": "Test memory with timestamp",
+                    "context": "test",
+                    "timestamp": "2026-01-30T11:45:00Z"
+                }
+            ],
+            "async": True
+        }
+    )
+    
+    assert response.status_code == 200, f"Expected 200, got {response.status_code}: {response.text}"
+    data = response.json()
+    assert data["success"] is True
+    assert data["async"] is True
+    assert "operation_id" in data
+    
+
+@pytest.mark.asyncio
+async def test_retain_with_timestamp_sync(api_client, test_bank_id):
+    """Test that sync retain accepts timestamp field."""
+    response = await api_client.post(
+        f"/v1/default/banks/{test_bank_id}/memories",
+        json={
+            "items": [
+                {
+                    "content": "Test memory with timestamp sync",
+                    "context": "test", 
+                    "timestamp": "2026-01-30T11:45:00Z"
+                }
+            ],
+            "async": False
+        }
+    )
+    
+    assert response.status_code == 200, f"Expected 200, got {response.status_code}: {response.text}"
+    data = response.json()
+    assert data["success"] is True
+    assert data["async"] is False
+    
+
+@pytest.mark.asyncio
+async def test_retain_with_multiple_timestamps(api_client, test_bank_id):
+    """Test that multiple items with different timestamp formats work."""
+    response = await api_client.post(
+        f"/v1/default/banks/{test_bank_id}/memories",
+        json={
+            "items": [
+                {
+                    "content": "Event 1",
+                    "timestamp": "2026-01-30T11:45:00Z"  # With Z
+                },
+                {
+                    "content": "Event 2", 
+                    "timestamp": "2026-01-30T12:00:00+00:00"  # With timezone
+                },
+                {
+                    "content": "Event 3"  # No timestamp
+                }
+            ],
+            "async": True
+        }
+    )
+    
+    assert response.status_code == 200, f"Expected 200, got {response.status_code}: {response.text}"
+    data = response.json()
+    assert data["success"] is True
+    assert data["items_count"] == 3
+
+
+@pytest.mark.asyncio
+async def test_retain_with_timestamp_async_complete_processing(api_client, test_bank_id):
+    """Test that async retain with timestamp completes full processing including fact extraction."""
+    # Submit async retain with timestamp
+    response = await api_client.post(
+        f"/v1/default/banks/{test_bank_id}/memories",
+        json={
+            "items": [
+                {
+                    "content": "The quarterly meeting was held on January 30th 2026",
+                    "context": "meetings",
+                    "timestamp": "2026-01-30T11:45:00Z"
+                }
+            ],
+            "async": True
+        }
+    )
+    
+    assert response.status_code == 200, f"Expected 200, got {response.status_code}: {response.text}"
+    data = response.json()
+    assert data["success"] is True
+    assert data["async"] is True
+    operation_id = data["operation_id"]
+    
+    # Wait for async processing to complete (poll operation status)
+    max_wait_seconds = 30
+    poll_interval = 0.5
+    elapsed = 0
+    operation_completed = False
+    
+    while elapsed < max_wait_seconds:
+        response = await api_client.get(f"/v1/default/banks/{test_bank_id}/operations/{operation_id}")
+        if response.status_code == 200:
+            op_status = response.json()
+            if op_status.get("status") == "completed":
+                operation_completed = True
+                break
+            elif op_status.get("status") == "failed":
+                raise AssertionError(f"Operation failed: {op_status.get('error_message')}")
+        
+        await asyncio.sleep(poll_interval)
+        elapsed += poll_interval
+    
+    assert operation_completed, f"Async operation did not complete within {max_wait_seconds} seconds"
+    
+    # Verify memories were actually stored
+    response = await api_client.get(
+        f"/v1/default/banks/{test_bank_id}/memories/list",
+        params={"limit": 10}
+    )
+    assert response.status_code == 200
+    items = response.json()["items"]
+    assert len(items) > 0, "Should have stored memories after async processing"
@@ -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,
@@ -0,0 +1,206 @@
+"""Unit tests for mental model operation validator hooks.
+
+Tests that the operation validator hooks are called correctly for
+mental model GET and refresh operations.
+"""
+
+import pytest
+
+from hindsight_api.extensions.operation_validator import (
+    MentalModelGetContext,
+    MentalModelGetResult,
+    MentalModelRefreshResult,
+    OperationValidatorExtension,
+    ValidationResult,
+)
+
+
+class TestMentalModelGetContextDataclass:
+    """Tests for MentalModelGetContext dataclass."""
+
+    def test_create_context(self):
+        """Test creating a MentalModelGetContext."""
+        from unittest.mock import MagicMock
+
+        request_context = MagicMock()
+        ctx = MentalModelGetContext(
+            bank_id="bank-1",
+            mental_model_id="mm-1",
+            request_context=request_context,
+        )
+
+        assert ctx.bank_id == "bank-1"
+        assert ctx.mental_model_id == "mm-1"
+        assert ctx.request_context is request_context
+
+
+class TestMentalModelGetResultDataclass:
+    """Tests for MentalModelGetResult dataclass."""
+
+    def test_create_result_success(self):
+        """Test creating a successful MentalModelGetResult."""
+        from unittest.mock import MagicMock
+
+        request_context = MagicMock()
+        result = MentalModelGetResult(
+            bank_id="bank-1",
+            mental_model_id="mm-1",
+            request_context=request_context,
+            output_tokens=250,
+        )
+
+        assert result.bank_id == "bank-1"
+        assert result.mental_model_id == "mm-1"
+        assert result.output_tokens == 250
+        assert result.success is True
+        assert result.error is None
+
+    def test_create_result_failure(self):
+        """Test creating a failed MentalModelGetResult."""
+        from unittest.mock import MagicMock
+
+        result = MentalModelGetResult(
+            bank_id="bank-1",
+            mental_model_id="mm-1",
+            request_context=MagicMock(),
+            output_tokens=0,
+            success=False,
+            error="Not found",
+        )
+
+        assert result.success is False
+        assert result.error == "Not found"
+
+
+class TestMentalModelRefreshResultDataclass:
+    """Tests for MentalModelRefreshResult dataclass."""
+
+    def test_create_result_with_all_fields(self):
+        """Test creating a MentalModelRefreshResult with all fields."""
+        from unittest.mock import MagicMock
+
+        result = MentalModelRefreshResult(
+            bank_id="bank-1",
+            mental_model_id="mm-1",
+            request_context=MagicMock(),
+            query_tokens=50,
+            output_tokens=500,
+            context_tokens=0,
+            facts_used=10,
+            mental_models_used=2,
+        )
+
+        assert result.query_tokens == 50
+        assert result.output_tokens == 500
+        assert result.context_tokens == 0
+        assert result.facts_used == 10
+        assert result.mental_models_used == 2
+        assert result.success is True
+        assert result.error is None
+
+    def test_create_result_failure(self):
+        """Test creating a failed MentalModelRefreshResult."""
+        from unittest.mock import MagicMock
+
+        result = MentalModelRefreshResult(
+            bank_id="bank-1",
+            mental_model_id="mm-1",
+            request_context=MagicMock(),
+            query_tokens=50,
+            output_tokens=0,
+            context_tokens=0,
+            facts_used=0,
+            mental_models_used=0,
+            success=False,
+            error="Reflect failed",
+        )
+
+        assert result.success is False
+        assert result.error == "Reflect failed"
+
+
+class TestDefaultHookBehavior:
+    """Tests for default (no-op) behavior of mental model hooks on OperationValidatorExtension."""
+
+    @pytest.fixture
+    def validator(self):
+        """Create a concrete subclass for testing default behavior."""
+        from unittest.mock import MagicMock
+
+        # Create a concrete subclass that implements the abstract methods
+        class TestValidator(OperationValidatorExtension):
+            async def validate_retain(self, ctx):
+                return ValidationResult.accept()
+
+            async def validate_recall(self, ctx):
+                return ValidationResult.accept()
+
+            async def validate_reflect(self, ctx):
+                return ValidationResult.accept()
+
+        return TestValidator(config={})
+
+    @pytest.mark.asyncio
+    async def test_validate_mental_model_get_default_accepts(self, validator):
+        """Test that default validate_mental_model_get accepts."""
+        from unittest.mock import MagicMock
+
+        ctx = MentalModelGetContext(
+            bank_id="bank-1",
+            mental_model_id="mm-1",
+            request_context=MagicMock(),
+        )
+
+        result = await validator.validate_mental_model_get(ctx)
+
+        assert result.allowed is True
+
+    @pytest.mark.asyncio
+    async def test_on_mental_model_get_complete_default_noop(self, validator):
+        """Test that default on_mental_model_get_complete is a no-op."""
+        from unittest.mock import MagicMock
+
+        result = MentalModelGetResult(
+            bank_id="bank-1",
+            mental_model_id="mm-1",
+            request_context=MagicMock(),
+            output_tokens=100,
+        )
+
+        # Should not raise
+        await validator.on_mental_model_get_complete(result)
+
+    @pytest.mark.asyncio
+    async def test_on_mental_model_refresh_complete_default_noop(self, validator):
+        """Test that default on_mental_model_refresh_complete is a no-op."""
+        from unittest.mock import MagicMock
+
+        result = MentalModelRefreshResult(
+            bank_id="bank-1",
+            mental_model_id="mm-1",
+            request_context=MagicMock(),
+            query_tokens=50,
+            output_tokens=500,
+            context_tokens=0,
+            facts_used=5,
+            mental_models_used=1,
+        )
+
+        # Should not raise
+        await validator.on_mental_model_refresh_complete(result)
+
+
+class TestExportsAvailable:
+    """Test that mental model hooks are properly exported."""
+
+    def test_imports_from_extensions_package(self):
+        """Test that all mental model types can be imported from hindsight_api.extensions."""
+        from hindsight_api.extensions import (
+            MentalModelGetContext,
+            MentalModelGetResult,
+            MentalModelRefreshResult,
+        )
+
+        assert MentalModelGetContext is not None
+        assert MentalModelGetResult is not None
+        assert MentalModelRefreshResult is not None
@@ -312,6 +312,49 @@ class TestDirectiveTags:
        # Cleanup
        await memory.delete_bank(bank_id, request_context=request_context)

+    async def test_list_all_directives_without_filter(self, memory: MemoryEngine, request_context):
+        """Test that listing directives without tags returns ALL directives (both tagged and untagged)."""
+        bank_id = f"test-directive-list-all-{uuid.uuid4().hex[:8]}"
+
+        # Ensure bank exists
+        await memory.get_bank_profile(bank_id, request_context=request_context)
+
+        # Create untagged directive
+        await memory.create_directive(
+            bank_id=bank_id,
+            name="Untagged Directive",
+            content="This has no tags",
+            request_context=request_context,
+        )
+
+        # Create tagged directive
+        await memory.create_directive(
+            bank_id=bank_id,
+            name="Tagged Directive",
+            content="This has tags",
+            tags=["project-x"],
+            request_context=request_context,
+        )
+
+        # List ALL directives (no tag filter, isolation_mode defaults to False)
+        all_directives = await memory.list_directives(
+            bank_id=bank_id,
+            request_context=request_context,
+        )
+
+        # Should return BOTH tagged and untagged directives
+        assert len(all_directives) == 2
+        directive_names = {d["name"] for d in all_directives}
+        assert "Untagged Directive" in directive_names
+        assert "Tagged Directive" in directive_names
+
+        # Verify the tagged directive has its tags
+        tagged = next(d for d in all_directives if d["name"] == "Tagged Directive")
+        assert tagged["tags"] == ["project-x"]
+
+        # Cleanup
+        await memory.delete_bank(bank_id, request_context=request_context)
+

 class TestReflect:
    """Test reflect endpoint."""
@@ -399,6 +442,161 @@ class TestDirectivesInReflect:
        # Cleanup
        await memory.delete_bank(bank_id, request_context=request_context)

+    async def test_tagged_directive_not_applied_without_tags(self, memory: MemoryEngine, request_context):
+        """Test that directives with tags are NOT applied to untagged reflect operations."""
+        bank_id = f"test-directive-isolation-{uuid.uuid4().hex[:8]}"
+
+        # Ensure bank exists
+        await memory.get_bank_profile(bank_id, request_context=request_context)
+
+        # Add some untagged content
+        await memory.retain_batch_async(
+            bank_id=bank_id,
+            contents=[
+                {"content": "The sky is blue."},
+                {"content": "Water is wet."},
+            ],
+            request_context=request_context,
+        )
+
+        # Add some tagged content for the project-x context
+        await memory.retain_batch_async(
+            bank_id=bank_id,
+            contents=[
+                {"content": "The sky is blue according to project X standards.", "tags": ["project-x"]},
+                {"content": "Project X color guidelines specify sky is blue.", "tags": ["project-x"]},
+            ],
+            request_context=request_context,
+        )
+        await memory.wait_for_background_tasks()
+
+        # Create an untagged directive (should be applied)
+        await memory.create_directive(
+            bank_id=bank_id,
+            name="General Policy",
+            content="Always be polite and start responses with 'Hello!'",
+            request_context=request_context,
+        )
+
+        # Create a tagged directive (should NOT be applied to untagged reflect)
+        await memory.create_directive(
+            bank_id=bank_id,
+            name="Tagged Policy",
+            content="ALWAYS respond in ALL CAPS and end with 'PROJECT-X ONLY'",
+            tags=["project-x"],
+            request_context=request_context,
+        )
+
+        # Run reflect without tags - should only apply the untagged directive
+        result = await memory.reflect_async(
+            bank_id=bank_id,
+            query="What color is the sky?",
+            request_context=request_context,
+        )
+
+        response_lower = result.text.lower()
+
+        # Should follow the untagged directive (polite greeting)
+        assert "hello" in response_lower, f"Expected 'Hello' from untagged directive, but got: {result.text}"
+
+        # Should NOT follow the tagged directive (all caps and PROJECT-X)
+        # If it did follow, the entire response would be in caps
+        all_caps = result.text.replace(" ", "").replace("!", "").replace(".", "").isupper()
+        assert not all_caps, f"Tagged directive was incorrectly applied to untagged operation: {result.text}"
+        assert "project-x only" not in response_lower, f"Tagged directive was incorrectly applied: {result.text}"
+
+        # Now run reflect WITH the tag - should apply BOTH directives
+        result_tagged = await memory.reflect_async(
+            bank_id=bank_id,
+            query="What color is the sky?",
+            tags=["project-x"],
+            tags_match="all_strict",
+            request_context=request_context,
+        )
+
+        response_tagged_lower = result_tagged.text.lower()
+
+        # With strict matching and tags, should apply the tagged directive
+        assert "project-x only" in response_tagged_lower, f"Tagged directive should be applied with tags: {result_tagged.text}"
+
+        # Cleanup
+        await memory.delete_bank(bank_id, request_context=request_context)
+
+    async def test_reflect_based_on_structure(self, memory: MemoryEngine, request_context):
+        """Test that reflect returns correct based_on structure with directives and memories separated."""
+        bank_id = f"test-reflect-based-on-{uuid.uuid4().hex[:8]}"
+
+        # Ensure bank exists
+        await memory.get_bank_profile(bank_id, request_context=request_context)
+
+        # Add some memories
+        await memory.retain_batch_async(
+            bank_id=bank_id,
+            contents=[
+                {"content": "Alice works at Google as a software engineer."},
+                {"content": "Bob is a product manager at Microsoft."},
+                {"content": "The team meets every Monday at 9am."},
+            ],
+            request_context=request_context,
+        )
+        await memory.wait_for_background_tasks()
+
+        # Create a directive
+        directive = await memory.create_directive(
+            bank_id=bank_id,
+            name="Professional Tone",
+            content="Always maintain a professional and formal tone in responses.",
+            request_context=request_context,
+        )
+        directive_id = directive["id"]
+
+        # Run reflect which returns the core result
+        result = await memory.reflect_async(
+            bank_id=bank_id,
+            query="Who works at Google?",
+            request_context=request_context,
+        )
+
+        # Verify based_on structure exists
+        assert result.based_on is not None
+
+        # Verify directives key exists and contains our directive
+        assert "directives" in result.based_on
+        directives_list = result.based_on.get("directives", [])
+
+        # Verify directives are dicts with id, name, content (not MemoryFact objects)
+        assert len(directives_list) > 0, "Should have at least one directive"
+        directive_found = False
+        for d in directives_list:
+            assert isinstance(d, dict), f"Directive should be dict, got {type(d)}"
+            assert "id" in d, "Directive dict should have 'id'"
+            assert "name" in d, "Directive dict should have 'name'"
+            assert "content" in d, "Directive dict should have 'content'"
+            # Check if this is our directive
+            if d["id"] == directive_id:
+                directive_found = True
+                assert d["name"] == "Professional Tone"
+                assert "professional" in d["content"].lower()
+
+        assert directive_found, f"Our directive {directive_id} should be in based_on.directives"
+
+        # Verify memories (world/experience) are separate from directives
+        has_memories = "world" in result.based_on or "experience" in result.based_on
+        assert has_memories, "Should have world or experience memories"
+
+        # Verify that if mental-models key exists, it's separate from directives
+        if "mental-models" in result.based_on:
+            mental_models = result.based_on.get("mental-models", [])
+            # Verify mental models are MemoryFact objects, not dicts like directives
+            for mm in mental_models:
+                assert hasattr(mm, "fact_type"), "Mental model should be MemoryFact with fact_type"
+                assert mm.fact_type == "mental-models"
+                assert hasattr(mm, "context")
+                assert "mental model" in mm.context.lower()
+
+        # Cleanup
+        await memory.delete_bank(bank_id, request_context=request_context)
+

 class TestDirectivesPromptInjection:
    """Test that directives are properly injected into the system prompt."""
@@ -451,3 +649,206 @@ class TestDirectivesPromptInjection:
        directives_pos = prompt.find("## DIRECTIVES")
        critical_rules_pos = prompt.find("## CRITICAL RULES")
        assert directives_pos < critical_rules_pos
+
+
+class TestMentalModelRefreshTagSecurity:
+    """Test that mental model refresh respects tag-based security boundaries."""
+
+    async def test_refresh_with_tags_only_accesses_same_tagged_models(
+        self, memory: MemoryEngine, request_context
+    ):
+        """Test that refreshing a mental model with tags can only access other models with the same tags.
+
+        This is a security test to ensure that mental models with tags (e.g., user:alice)
+        cannot access mental models from other scopes (e.g., user:bob or no tags) during refresh.
+        """
+        bank_id = f"test-refresh-tags-{uuid.uuid4().hex[:8]}"
+
+        # Ensure bank exists
+        await memory.get_bank_profile(bank_id, request_context=request_context)
+
+        # Add some facts with different tags
+        await memory.retain_batch_async(
+            bank_id=bank_id,
+            contents=[
+                {"content": "Alice works on the frontend React project. Alice's favorite color is blue.", "tags": ["user:alice"]},
+                {"content": "Alice prefers working in the morning. Alice drinks coffee every day.", "tags": ["user:alice"]},
+                {"content": "Bob works on the backend API services. Bob's favorite language is Python.", "tags": ["user:bob"]},
+                {"content": "Bob prefers working at night. Bob drinks tea every day.", "tags": ["user:bob"]},
+                {"content": "The company has 100 employees and is growing fast.", "tags": []},  # No tags
+            ],
+            request_context=request_context,
+        )
+
+        # Wait for background processing
+        await memory.wait_for_background_tasks()
+
+        # Create mental model for user:alice with sensitive data
+        mm_alice = await memory.create_mental_model(
+            bank_id=bank_id,
+            name="Alice's Work Profile",
+            source_query="What does Alice work on?",
+            content="Alice is a frontend engineer specializing in React",
+            tags=["user:alice"],
+            request_context=request_context,
+        )
+
+        # Create mental model for user:bob with sensitive data
+        mm_bob = await memory.create_mental_model(
+            bank_id=bank_id,
+            name="Bob's Work Profile",
+            source_query="What does Bob work on?",
+            content="Bob is a backend engineer specializing in Python",
+            tags=["user:bob"],
+            request_context=request_context,
+        )
+
+        # Create mental model with no tags (should not be accessible from tagged models)
+        mm_untagged = await memory.create_mental_model(
+            bank_id=bank_id,
+            name="Company Info",
+            source_query="What is the company info?",
+            content="The company has 100 employees",
+            request_context=request_context,
+        )
+
+        # Create a mental model for user:alice that will be refreshed
+        mm_alice_refresh = await memory.create_mental_model(
+            bank_id=bank_id,
+            name="Alice's Summary",
+            source_query="What are all the facts about work and preferences?",  # Broad query that should match all facts
+            content="Initial content",
+            tags=["user:alice"],
+            request_context=request_context,
+        )
+
+        # Refresh Alice's mental model
+        refreshed = await memory.refresh_mental_model(
+            bank_id=bank_id,
+            mental_model_id=mm_alice_refresh["id"],
+            request_context=request_context,
+        )
+
+        # SECURITY CHECK: The refreshed content should ONLY include information from
+        # memories/models tagged with user:alice, NOT from user:bob or untagged
+        refreshed_content = refreshed["content"].lower()
+
+        # Should include Alice's content (either from facts or mental models)
+        assert "alice" in refreshed_content, \
+            "Refreshed model should access memories/models with matching tags (user:alice)"
+
+        # MUST NOT include Bob's content (security violation)
+        # Use word boundary matching to avoid false positives (e.g., "team" contains "tea")
+        import re
+        def contains_word(text: str, word: str) -> bool:
+            """Check if text contains word as a whole word (not substring)."""
+            return bool(re.search(rf'\b{re.escape(word)}\b', text, re.IGNORECASE))
+
+        assert not contains_word(refreshed_content, "bob") and \
+               not contains_word(refreshed_content, "python") and \
+               not contains_word(refreshed_content, "tea"), \
+            f"SECURITY VIOLATION: Refreshed model accessed memories/models with different tags (user:bob). Content: {refreshed['content']}"
+
+        # MUST NOT include untagged content (security violation)
+        assert "100 employees" not in refreshed_content and "growing fast" not in refreshed_content, \
+            f"SECURITY VIOLATION: Refreshed model accessed untagged memories/models. Content: {refreshed['content']}"
+
+        # Cleanup
+        await memory.delete_bank(bank_id, request_context=request_context)
+
+    async def test_consolidation_only_refreshes_matching_tagged_models(
+        self, memory: MemoryEngine, request_context
+    ):
+        """Test that consolidation only triggers refresh for mental models with matching tags.
+
+        This is a security test to ensure that when tagged memories are consolidated,
+        only mental models with overlapping tags get refreshed, not all mental models.
+        """
+        bank_id = f"test-consolidation-refresh-{uuid.uuid4().hex[:8]}"
+
+        # Ensure bank exists
+        await memory.get_bank_profile(bank_id, request_context=request_context)
+
+        # Create mental models with different tags, all with refresh_after_consolidation=true
+        mm_alice = await memory.create_mental_model(
+            bank_id=bank_id,
+            name="Alice's Model",
+            source_query="What about Alice?",
+            content="Initial Alice content",
+            tags=["user:alice"],
+            trigger={"refresh_after_consolidation": True},
+            request_context=request_context,
+        )
+
+        mm_bob = await memory.create_mental_model(
+            bank_id=bank_id,
+            name="Bob's Model",
+            source_query="What about Bob?",
+            content="Initial Bob content",
+            tags=["user:bob"],
+            trigger={"refresh_after_consolidation": True},
+            request_context=request_context,
+        )
+
+        mm_untagged = await memory.create_mental_model(
+            bank_id=bank_id,
+            name="Untagged Model",
+            source_query="What about general stuff?",
+            content="Initial untagged content",
+            trigger={"refresh_after_consolidation": True},
+            request_context=request_context,
+        )
+
+        # Record initial last_refreshed_at timestamps
+        alice_initial = mm_alice["last_refreshed_at"]
+        bob_initial = mm_bob["last_refreshed_at"]
+        untagged_initial = mm_untagged["last_refreshed_at"]
+
+        # Add memories with user:alice tags
+        await memory.retain_batch_async(
+            bank_id=bank_id,
+            contents=[
+                {"content": "Alice likes React", "tags": ["user:alice"]},
+                {"content": "Alice drinks coffee", "tags": ["user:alice"]},
+            ],
+            request_context=request_context,
+        )
+
+        # Trigger consolidation manually (this should only refresh Alice's mental model)
+        from hindsight_api.engine.consolidation.consolidator import run_consolidation_job
+
+        result = await run_consolidation_job(
+            memory_engine=memory,
+            bank_id=bank_id,
+            request_context=request_context,
+        )
+
+        # Wait for background refresh tasks to complete
+        await memory.wait_for_background_tasks()
+
+        # Check that mental models were refreshed appropriately
+        mm_alice_after = await memory.get_mental_model(
+            bank_id, mm_alice["id"], request_context=request_context
+        )
+        mm_bob_after = await memory.get_mental_model(
+            bank_id, mm_bob["id"], request_context=request_context
+        )
+        mm_untagged_after = await memory.get_mental_model(
+            bank_id, mm_untagged["id"], request_context=request_context
+        )
+
+        # SECURITY CHECK: Only Alice's mental model and untagged model should be refreshed
+        # Alice's model should be refreshed (tags match)
+        assert mm_alice_after["last_refreshed_at"] != alice_initial or mm_alice_after["content"] != mm_alice["content"], \
+            "Alice's mental model should be refreshed when user:alice memories are consolidated"
+
+        # Bob's model should NOT be refreshed (tags don't match)
+        assert mm_bob_after["last_refreshed_at"] == bob_initial, \
+            "SECURITY VIOLATION: Bob's mental model was refreshed even though user:bob memories were not consolidated"
+
+        # Untagged model should be refreshed (untagged models are always refreshed)
+        assert mm_untagged_after["last_refreshed_at"] != untagged_initial or mm_untagged_after["content"] != mm_untagged["content"], \
+            "Untagged mental model should be refreshed after any consolidation"
+
+        # Cleanup
+        await memory.delete_bank(bank_id, request_context=request_context)
@@ -0,0 +1,123 @@
+"""Test provider-specific default models in config."""
+
+import os
+
+import pytest
+
+
+def test_provider_default_models():
+    """Test that each provider has a default model and it's used when model is not explicitly set."""
+    from hindsight_api.config import PROVIDER_DEFAULT_MODELS, HindsightConfig, clear_config_cache
+
+    # Save original env vars
+    original_provider = os.environ.get("HINDSIGHT_API_LLM_PROVIDER")
+    original_model = os.environ.get("HINDSIGHT_API_LLM_MODEL")
+
+    try:
+        # Test each provider has a default
+        for provider, expected_model in PROVIDER_DEFAULT_MODELS.items():
+            clear_config_cache()
+            os.environ["HINDSIGHT_API_LLM_PROVIDER"] = provider
+            # Remove explicit model setting to test default
+            if "HINDSIGHT_API_LLM_MODEL" in os.environ:
+                del os.environ["HINDSIGHT_API_LLM_MODEL"]
+
+            config = HindsightConfig.from_env()
+            assert config.llm_provider == provider, f"Provider mismatch for {provider}"
+            assert config.llm_model == expected_model, f"Expected {expected_model} for {provider}, got {config.llm_model}"
+
+    finally:
+        # Restore original env vars
+        clear_config_cache()
+        if original_provider:
+            os.environ["HINDSIGHT_API_LLM_PROVIDER"] = original_provider
+        elif "HINDSIGHT_API_LLM_PROVIDER" in os.environ:
+            del os.environ["HINDSIGHT_API_LLM_PROVIDER"]
+
+        if original_model:
+            os.environ["HINDSIGHT_API_LLM_MODEL"] = original_model
+        elif "HINDSIGHT_API_LLM_MODEL" in os.environ:
+            del os.environ["HINDSIGHT_API_LLM_MODEL"]
+
+
+def test_explicit_model_overrides_provider_default():
+    """Test that explicit model setting overrides provider default."""
+    from hindsight_api.config import HindsightConfig, clear_config_cache
+
+    original_provider = os.environ.get("HINDSIGHT_API_LLM_PROVIDER")
+    original_model = os.environ.get("HINDSIGHT_API_LLM_MODEL")
+
+    try:
+        clear_config_cache()
+        os.environ["HINDSIGHT_API_LLM_PROVIDER"] = "anthropic"
+        os.environ["HINDSIGHT_API_LLM_MODEL"] = "claude-sonnet-4-5-20250929"
+
+        config = HindsightConfig.from_env()
+        assert config.llm_provider == "anthropic"
+        assert config.llm_model == "claude-sonnet-4-5-20250929", "Explicit model should override default"
+
+    finally:
+        clear_config_cache()
+        if original_provider:
+            os.environ["HINDSIGHT_API_LLM_PROVIDER"] = original_provider
+        elif "HINDSIGHT_API_LLM_PROVIDER" in os.environ:
+            del os.environ["HINDSIGHT_API_LLM_PROVIDER"]
+
+        if original_model:
+            os.environ["HINDSIGHT_API_LLM_MODEL"] = original_model
+        elif "HINDSIGHT_API_LLM_MODEL" in os.environ:
+            del os.environ["HINDSIGHT_API_LLM_MODEL"]
+
+
+def test_per_operation_provider_default_model():
+    """Test that per-operation providers use their own default models."""
+    from hindsight_api.config import HindsightConfig, clear_config_cache
+
+    original_provider = os.environ.get("HINDSIGHT_API_LLM_PROVIDER")
+    original_model = os.environ.get("HINDSIGHT_API_LLM_MODEL")
+    original_retain_provider = os.environ.get("HINDSIGHT_API_RETAIN_LLM_PROVIDER")
+    original_retain_model = os.environ.get("HINDSIGHT_API_RETAIN_LLM_MODEL")
+
+    try:
+        clear_config_cache()
+        os.environ["HINDSIGHT_API_LLM_PROVIDER"] = "openai"
+        # Remove explicit model to use provider default
+        if "HINDSIGHT_API_LLM_MODEL" in os.environ:
+            del os.environ["HINDSIGHT_API_LLM_MODEL"]
+
+        # Set retain-specific provider but not model
+        os.environ["HINDSIGHT_API_RETAIN_LLM_PROVIDER"] = "anthropic"
+        if "HINDSIGHT_API_RETAIN_LLM_MODEL" in os.environ:
+            del os.environ["HINDSIGHT_API_RETAIN_LLM_MODEL"]
+
+        config = HindsightConfig.from_env()
+
+        # Global LLM should use OpenAI default
+        assert config.llm_model == "o3-mini", f"Expected o3-mini, got {config.llm_model}"
+
+        # Retain should use Anthropic default
+        assert (
+            config.retain_llm_model == "claude-haiku-4-5-20251001"
+        ), f"Expected claude-haiku-4-5-20251001, got {config.retain_llm_model}"
+
+    finally:
+        clear_config_cache()
+        if original_provider:
+            os.environ["HINDSIGHT_API_LLM_PROVIDER"] = original_provider
+        elif "HINDSIGHT_API_LLM_PROVIDER" in os.environ:
+            del os.environ["HINDSIGHT_API_LLM_PROVIDER"]
+
+        if original_model:
+            os.environ["HINDSIGHT_API_LLM_MODEL"] = original_model
+        elif "HINDSIGHT_API_LLM_MODEL" in os.environ:
+            del os.environ["HINDSIGHT_API_LLM_MODEL"]
+
+        if original_retain_provider:
+            os.environ["HINDSIGHT_API_RETAIN_LLM_PROVIDER"] = original_retain_provider
+        elif "HINDSIGHT_API_RETAIN_LLM_PROVIDER" in os.environ:
+            del os.environ["HINDSIGHT_API_RETAIN_LLM_PROVIDER"]
+
+        if original_retain_model:
+            os.environ["HINDSIGHT_API_RETAIN_LLM_MODEL"] = original_retain_model
+        elif "HINDSIGHT_API_RETAIN_LLM_MODEL" in os.environ:
+            del os.environ["HINDSIGHT_API_RETAIN_LLM_MODEL"]
@@ -175,6 +175,45 @@ class TestMentalModelsCRUD:
        # Cleanup
        await memory.delete_bank(bank_id, request_context=request_context)

+    @pytest.mark.asyncio
+    async def test_create_mental_model_with_custom_id(self, memory: MemoryEngine, request_context):
+        """Test creating a mental model with a custom ID."""
+        bank_id = f"test-mental-model-custom-id-{uuid.uuid4().hex[:8]}"
+
+        # Create the bank first
+        await memory.get_bank_profile(bank_id=bank_id, request_context=request_context)
+
+        # Create a mental model with a custom ID
+        custom_id = "team-communication-preferences"
+        mental_model = await memory.create_mental_model(
+            bank_id=bank_id,
+            mental_model_id=custom_id,
+            name="Team Communication Preferences",
+            source_query="How does the team prefer to communicate?",
+            content="The team prefers async communication via Slack",
+            tags=["team", "communication"],
+            request_context=request_context,
+        )
+
+        # Verify the custom ID was used
+        assert mental_model["id"] == custom_id
+        assert mental_model["name"] == "Team Communication Preferences"
+        assert mental_model["tags"] == ["team", "communication"]
+
+        # Verify we can retrieve it with the custom ID
+        fetched = await memory.get_mental_model(
+            bank_id=bank_id,
+            mental_model_id=custom_id,
+            request_context=request_context,
+        )
+
+        assert fetched is not None
+        assert fetched["id"] == custom_id
+        assert fetched["name"] == "Team Communication Preferences"
+
+        # Cleanup
+        await memory.delete_bank(bank_id, request_context=request_context)
+

 class TestObservationsAPI:
    """Test observations API endpoints.
@@ -2193,3 +2193,65 @@ If the text contains both Italian and English content, extract ONLY the Italian

        # Clear cache again to restore original config
        clear_config_cache()
+
+
+@pytest.mark.asyncio
+async def test_retain_batch_with_per_item_tags_on_document(memory, request_context):
+    """
+    Test that per-item tags are correctly stored on documents.
+
+    This test verifies the fix for a bug where per-item tags in content dictionaries
+    were not being merged and passed to document tracking, causing tags to be lost
+    even though they were correctly sent through the API.
+
+    Without the fix, this test would fail because:
+    - Tags are correctly passed in the content dict
+    - Tags are correctly stored on memory_units (facts)
+    - BUT tags were NOT stored on the document record itself
+    """
+    bank_id = f"test_doc_tags_{datetime.now(timezone.utc).timestamp()}"
+    document_id = "app-state-testuser"
+
+    try:
+        # Retain content with per-item tags (simulating the TasteAI use case)
+        contents = [
+            {
+                "content": '{"username":"testuser","meals":[],"preferences":{"nickname":"testuser"}}',
+                "document_id": document_id,
+                "tags": ["user:testuser", "app-type:taste-ai"],
+            }
+        ]
+
+        result = await memory.retain_batch_async(
+            bank_id=bank_id,
+            contents=contents,
+            request_context=request_context,
+        )
+
+        assert len(result) > 0, "Should have retained content"
+        print(f"\n=== Retained content with tags ===")
+
+        # Retrieve the document
+        doc = await memory.get_document(
+            document_id=document_id,
+            bank_id=bank_id,
+            request_context=request_context,
+        )
+
+        assert doc is not None, "Document should exist"
+        assert "tags" in doc, "Document should have tags field"
+
+        # This is the critical assertion - tags should be stored on the document
+        doc_tags = doc["tags"] or []
+        print(f"Document tags: {doc_tags}")
+
+        assert "user:testuser" in doc_tags, \
+            f"Document should have 'user:testuser' tag, but got: {doc_tags}"
+        assert "app-type:taste-ai" in doc_tags, \
+            f"Document should have 'app-type:taste-ai' tag, but got: {doc_tags}"
+
+        print("✓ Per-item tags correctly stored on document")
+
+    finally:
+        await memory.delete_bank(bank_id, request_context=request_context)
+        print(f"\n=== Cleaned up bank: {bank_id} ===")
@@ -1020,6 +1020,81 @@ class TestDynamicTenantDiscovery:
        for task in claimed:
            assert task.schema is None

+    @pytest.mark.asyncio
+    async def test_poller_with_custom_schema(self, pool):
+        """Test that poller uses custom schema when schema parameter is provided."""
+        from hindsight_api.worker import WorkerPoller
+
+        # Create a custom schema for testing
+        test_schema = "test_custom_schema"
+
+        try:
+            # Create schema and copy table structure
+            await pool.execute(f'CREATE SCHEMA IF NOT EXISTS "{test_schema}"')
+            await pool.execute(
+                f"""
+                CREATE TABLE "{test_schema}".async_operations (
+                    LIKE public.async_operations INCLUDING ALL
+                )
+                """
+            )
+
+            # Create pending tasks in the custom schema
+            bank_id = f"test-worker-{uuid.uuid4().hex[:8]}"
+            task_ids = []
+            for i in range(3):
+                op_id = uuid.uuid4()
+                task_ids.append(str(op_id))
+                payload = json.dumps({"type": "test_task", "index": i, "bank_id": bank_id})
+                await pool.execute(
+                    f"""
+                    INSERT INTO "{test_schema}".async_operations (operation_id, bank_id, operation_type, status, task_payload)
+                    VALUES ($1, $2, 'test', 'pending', $3::jsonb)
+                    """,
+                    op_id,
+                    bank_id,
+                    payload,
+                )
+
+            # Create poller with custom schema
+            poller = WorkerPoller(
+                pool=pool,
+                worker_id="test-worker-custom-schema",
+                executor=lambda x: None,
+                schema=test_schema,
+            )
+
+            # Claim tasks
+            claimed = await poller.claim_batch()
+            assert len(claimed) == 3, f"Expected 3 tasks, got {len(claimed)}"
+
+            # All tasks should have schema=test_schema
+            claimed_ids = []
+            for task in claimed:
+                assert task.schema == test_schema, f"Expected schema '{test_schema}', got '{task.schema}'"
+                claimed_ids.append(task.operation_id)
+
+            # Verify claimed tasks match what we inserted
+            assert set(claimed_ids) == set(task_ids)
+
+            # Verify tasks are marked as processing in the custom schema
+            rows = await pool.fetch(
+                f"""
+                SELECT operation_id, status, worker_id
+                FROM "{test_schema}".async_operations
+                WHERE operation_id = ANY($1)
+                """,
+                [uuid.UUID(tid) for tid in task_ids],
+            )
+            assert len(rows) == 3
+            for row in rows:
+                assert row["status"] == "processing"
+                assert row["worker_id"] == "test-worker-custom-schema"
+
+        finally:
+            # Clean up: drop the custom schema
+            await pool.execute(f'DROP SCHEMA IF EXISTS "{test_schema}" CASCADE')
+

 async def test_worker_fire_and_forget_nonblocking(pool, clean_operations):
    """
@@ -1,6 +1,6 @@
 [package]
 name = "hindsight-cli"
-version = "0.4.2"
+version = "0.4.9"
 edition = "2021"
 authors = ["Hindsight Team"]
 description = "A beautiful CLI for Hindsight - semantic memory system"
@@ -98,6 +98,7 @@ pub fn create(
    bank_id: &str,
    name: &str,
    source_query: &str,
+    id: Option<&str>,
    verbose: bool,
    output_format: OutputFormat,
 ) -> Result<()> {
@@ -108,6 +109,7 @@ pub fn create(
    };

    let request = types::CreateMentalModelRequest {
+        id: id.map(|s| s.to_string()),
        name: name.to_string(),
        source_query: source_query.to_string(),
        max_tokens: 2048,
@@ -596,6 +596,10 @@ enum MentalModelCommands {

        /// Source query to generate the mental model from
        source_query: String,
+
+        /// Optional custom ID for the mental model (alphanumeric lowercase with hyphens)
+        #[arg(long)]
+        id: Option<String>,
    },

    /// Update a mental model
@@ -863,8 +867,8 @@ fn run() -> Result<()> {
            MentalModelCommands::Get { bank_id, mental_model_id } => {
                commands::mental_model::get(&client, &bank_id, &mental_model_id, verbose, output_format)
            }
-            MentalModelCommands::Create { bank_id, name, source_query } => {
-                commands::mental_model::create(&client, &bank_id, &name, &source_query, verbose, output_format)
+            MentalModelCommands::Create { bank_id, name, source_query, id } => {
+                commands::mental_model::create(&client, &bank_id, &name, &source_query, id.as_deref(), verbose, output_format)
            }
            MentalModelCommands::Update { bank_id, mental_model_id, name } => {
                commands::mental_model::update(&client, &bank_id, &mental_model_id, name, verbose, output_format)
@@ -7,7 +7,7 @@

    HTTP API for Hindsight

-    The version of the OpenAPI document: 0.4.2
+    The version of the OpenAPI document: 0.4.9
    Generated by OpenAPI Generator (https://openapi-generator.tech)

    Do not edit the class manually.
@@ -5,7 +5,7 @@

    HTTP API for Hindsight

-    The version of the OpenAPI document: 0.4.2
+    The version of the OpenAPI document: 0.4.9
    Generated by OpenAPI Generator (https://openapi-generator.tech)

    Do not edit the class manually.
@@ -5,7 +5,7 @@

    HTTP API for Hindsight

-    The version of the OpenAPI document: 0.4.2
+    The version of the OpenAPI document: 0.4.9
    Generated by OpenAPI Generator (https://openapi-generator.tech)

    Do not edit the class manually.
@@ -5,7 +5,7 @@

    HTTP API for Hindsight

-    The version of the OpenAPI document: 0.4.2
+    The version of the OpenAPI document: 0.4.9
    Generated by OpenAPI Generator (https://openapi-generator.tech)

    Do not edit the class manually.
@@ -5,7 +5,7 @@

    HTTP API for Hindsight

-    The version of the OpenAPI document: 0.4.2
+    The version of the OpenAPI document: 0.4.9
    Generated by OpenAPI Generator (https://openapi-generator.tech)

    Do not edit the class manually.
@@ -5,7 +5,7 @@

    HTTP API for Hindsight

-    The version of the OpenAPI document: 0.4.2
+    The version of the OpenAPI document: 0.4.9
    Generated by OpenAPI Generator (https://openapi-generator.tech)

    Do not edit the class manually.
@@ -5,7 +5,7 @@

    HTTP API for Hindsight

-    The version of the OpenAPI document: 0.4.2
+    The version of the OpenAPI document: 0.4.9
    Generated by OpenAPI Generator (https://openapi-generator.tech)

    Do not edit the class manually.
@@ -5,7 +5,7 @@

    HTTP API for Hindsight

-    The version of the OpenAPI document: 0.4.2
+    The version of the OpenAPI document: 0.4.9
    Generated by OpenAPI Generator (https://openapi-generator.tech)

    Do not edit the class manually.
@@ -5,7 +5,7 @@

    HTTP API for Hindsight

-    The version of the OpenAPI document: 0.4.2
+    The version of the OpenAPI document: 0.4.9
    Generated by OpenAPI Generator (https://openapi-generator.tech)

    Do not edit the class manually.
@@ -5,7 +5,7 @@

    HTTP API for Hindsight

-    The version of the OpenAPI document: 0.4.2
+    The version of the OpenAPI document: 0.4.9
    Generated by OpenAPI Generator (https://openapi-generator.tech)

    Do not edit the class manually.
@@ -5,7 +5,7 @@

    HTTP API for Hindsight

-    The version of the OpenAPI document: 0.4.2
+    The version of the OpenAPI document: 0.4.9
    Generated by OpenAPI Generator (https://openapi-generator.tech)

    Do not edit the class manually.
@@ -489,7 +489,7 @@ class Configuration:
        return "Python SDK Debug Report:\n"\
               "OS: {env}\n"\
               "Python Version: {pyversion}\n"\
-               "Version of the API: 0.4.2\n"\
+               "Version of the API: 0.4.9\n"\
               "SDK Package Version: 0.0.7".\
               format(env=sys.platform, pyversion=sys.version)

@@ -5,7 +5,7 @@

    HTTP API for Hindsight

-    The version of the OpenAPI document: 0.4.2
+    The version of the OpenAPI document: 0.4.9
    Generated by OpenAPI Generator (https://openapi-generator.tech)

    Do not edit the class manually.
@@ -6,7 +6,7 @@

    HTTP API for Hindsight

-    The version of the OpenAPI document: 0.4.2
+    The version of the OpenAPI document: 0.4.9
    Generated by OpenAPI Generator (https://openapi-generator.tech)

    Do not edit the class manually.
@@ -5,7 +5,7 @@

    HTTP API for Hindsight

-    The version of the OpenAPI document: 0.4.2
+    The version of the OpenAPI document: 0.4.9
    Generated by OpenAPI Generator (https://openapi-generator.tech)

    Do not edit the class manually.
@@ -5,7 +5,7 @@

    HTTP API for Hindsight

-    The version of the OpenAPI document: 0.4.2
+    The version of the OpenAPI document: 0.4.9
    Generated by OpenAPI Generator (https://openapi-generator.tech)

    Do not edit the class manually.
@@ -5,7 +5,7 @@

    HTTP API for Hindsight

-    The version of the OpenAPI document: 0.4.2
+    The version of the OpenAPI document: 0.4.9
    Generated by OpenAPI Generator (https://openapi-generator.tech)

    Do not edit the class manually.
@@ -5,7 +5,7 @@

    HTTP API for Hindsight

-    The version of the OpenAPI document: 0.4.2
+    The version of the OpenAPI document: 0.4.9
    Generated by OpenAPI Generator (https://openapi-generator.tech)

    Do not edit the class manually.
@@ -5,7 +5,7 @@

    HTTP API for Hindsight

-    The version of the OpenAPI document: 0.4.2
+    The version of the OpenAPI document: 0.4.9
    Generated by OpenAPI Generator (https://openapi-generator.tech)

    Do not edit the class manually.
@@ -5,7 +5,7 @@

    HTTP API for Hindsight

-    The version of the OpenAPI document: 0.4.2
+    The version of the OpenAPI document: 0.4.9
    Generated by OpenAPI Generator (https://openapi-generator.tech)

    Do not edit the class manually.
@@ -5,7 +5,7 @@

    HTTP API for Hindsight

-    The version of the OpenAPI document: 0.4.2
+    The version of the OpenAPI document: 0.4.9
    Generated by OpenAPI Generator (https://openapi-generator.tech)

    Do not edit the class manually.
@@ -5,7 +5,7 @@

    HTTP API for Hindsight

-    The version of the OpenAPI document: 0.4.2
+    The version of the OpenAPI document: 0.4.9
    Generated by OpenAPI Generator (https://openapi-generator.tech)

    Do not edit the class manually.
@@ -5,7 +5,7 @@

    HTTP API for Hindsight

-    The version of the OpenAPI document: 0.4.2
+    The version of the OpenAPI document: 0.4.9
    Generated by OpenAPI Generator (https://openapi-generator.tech)

    Do not edit the class manually.
@@ -5,7 +5,7 @@

    HTTP API for Hindsight

-    The version of the OpenAPI document: 0.4.2
+    The version of the OpenAPI document: 0.4.9
    Generated by OpenAPI Generator (https://openapi-generator.tech)

    Do not edit the class manually.
@@ -5,7 +5,7 @@

    HTTP API for Hindsight

-    The version of the OpenAPI document: 0.4.2
+    The version of the OpenAPI document: 0.4.9
    Generated by OpenAPI Generator (https://openapi-generator.tech)

    Do not edit the class manually.
@@ -5,7 +5,7 @@

    HTTP API for Hindsight

-    The version of the OpenAPI document: 0.4.2
+    The version of the OpenAPI document: 0.4.9
    Generated by OpenAPI Generator (https://openapi-generator.tech)

    Do not edit the class manually.
@@ -5,7 +5,7 @@

    HTTP API for Hindsight

-    The version of the OpenAPI document: 0.4.2
+    The version of the OpenAPI document: 0.4.9
    Generated by OpenAPI Generator (https://openapi-generator.tech)

    Do not edit the class manually.
@@ -5,7 +5,7 @@

    HTTP API for Hindsight

-    The version of the OpenAPI document: 0.4.2
+    The version of the OpenAPI document: 0.4.9
    Generated by OpenAPI Generator (https://openapi-generator.tech)

    Do not edit the class manually.
@@ -5,7 +5,7 @@

    HTTP API for Hindsight

-    The version of the OpenAPI document: 0.4.2
+    The version of the OpenAPI document: 0.4.9
    Generated by OpenAPI Generator (https://openapi-generator.tech)

    Do not edit the class manually.
@@ -5,7 +5,7 @@

    HTTP API for Hindsight

-    The version of the OpenAPI document: 0.4.2
+    The version of the OpenAPI document: 0.4.9
    Generated by OpenAPI Generator (https://openapi-generator.tech)

    Do not edit the class manually.
@@ -28,12 +28,13 @@ class CreateMentalModelRequest(BaseModel):
    """
    Request model for creating a mental model.
    """ # noqa: E501
+    id: Optional[StrictStr] = None
    name: StrictStr = Field(description="Human-readable name for the mental model")
    source_query: StrictStr = Field(description="The query to run to generate content")
    tags: Optional[List[StrictStr]] = Field(default=None, description="Tags for scoped visibility")
    max_tokens: Optional[Annotated[int, Field(le=8192, strict=True, ge=256)]] = Field(default=2048, description="Maximum tokens for generated content")
    trigger: Optional[MentalModelTrigger] = Field(default=None, description="Trigger settings")
-    __properties: ClassVar[List[str]] = ["name", "source_query", "tags", "max_tokens", "trigger"]
+    __properties: ClassVar[List[str]] = ["id", "name", "source_query", "tags", "max_tokens", "trigger"]

    model_config = ConfigDict(
        populate_by_name=True,
@@ -77,6 +78,11 @@ class CreateMentalModelRequest(BaseModel):
        # override the default output from pydantic by calling `to_dict()` of trigger
        if self.trigger:
            _dict['trigger'] = self.trigger.to_dict()
+        # set to None if id (nullable) is None
+        # and model_fields_set contains the field
+        if self.id is None and "id" in self.model_fields_set:
+            _dict['id'] = None
+
        return _dict

    @classmethod
@@ -89,6 +95,7 @@ class CreateMentalModelRequest(BaseModel):
            return cls.model_validate(obj)

        _obj = cls.model_validate({
+            "id": obj.get("id"),
            "name": obj.get("name"),
            "source_query": obj.get("source_query"),
            "tags": obj.get("tags"),
@@ -5,7 +5,7 @@

    HTTP API for Hindsight

-    The version of the OpenAPI document: 0.4.2
+    The version of the OpenAPI document: 0.4.9
    Generated by OpenAPI Generator (https://openapi-generator.tech)

    Do not edit the class manually.
@@ -18,7 +18,7 @@ import re  # noqa: F401
 import json

 from pydantic import BaseModel, ConfigDict, Field, StrictStr
-from typing import Any, ClassVar, Dict, List
+from typing import Any, ClassVar, Dict, List, Optional
 from typing import Optional, Set
 from typing_extensions import Self

@@ -26,8 +26,9 @@ class CreateMentalModelResponse(BaseModel):
    """
    Response model for mental model creation.
    """ # noqa: E501
-    operation_id: StrictStr = Field(description="Operation ID to track progress")
-    __properties: ClassVar[List[str]] = ["operation_id"]
+    mental_model_id: Optional[StrictStr] = None
+    operation_id: StrictStr = Field(description="Operation ID to track refresh progress")
+    __properties: ClassVar[List[str]] = ["mental_model_id", "operation_id"]

    model_config = ConfigDict(
        populate_by_name=True,
@@ -68,6 +69,11 @@ class CreateMentalModelResponse(BaseModel):
            exclude=excluded_fields,
            exclude_none=True,
        )
+        # set to None if mental_model_id (nullable) is None
+        # and model_fields_set contains the field
+        if self.mental_model_id is None and "mental_model_id" in self.model_fields_set:
+            _dict['mental_model_id'] = None
+
        return _dict

    @classmethod
@@ -80,6 +86,7 @@ class CreateMentalModelResponse(BaseModel):
            return cls.model_validate(obj)

        _obj = cls.model_validate({
+            "mental_model_id": obj.get("mental_model_id"),
            "operation_id": obj.get("operation_id")
        })
        return _obj
@@ -5,7 +5,7 @@

    HTTP API for Hindsight

-    The version of the OpenAPI document: 0.4.2
+    The version of the OpenAPI document: 0.4.9
    Generated by OpenAPI Generator (https://openapi-generator.tech)

    Do not edit the class manually.
@@ -5,7 +5,7 @@

    HTTP API for Hindsight

-    The version of the OpenAPI document: 0.4.2
+    The version of the OpenAPI document: 0.4.9
    Generated by OpenAPI Generator (https://openapi-generator.tech)

    Do not edit the class manually.
@@ -5,7 +5,7 @@

    HTTP API for Hindsight

-    The version of the OpenAPI document: 0.4.2
+    The version of the OpenAPI document: 0.4.9
    Generated by OpenAPI Generator (https://openapi-generator.tech)

    Do not edit the class manually.
@@ -5,7 +5,7 @@

    HTTP API for Hindsight

-    The version of the OpenAPI document: 0.4.2
+    The version of the OpenAPI document: 0.4.9
    Generated by OpenAPI Generator (https://openapi-generator.tech)

    Do not edit the class manually.
@@ -5,7 +5,7 @@

    HTTP API for Hindsight

-    The version of the OpenAPI document: 0.4.2
+    The version of the OpenAPI document: 0.4.9
    Generated by OpenAPI Generator (https://openapi-generator.tech)

    Do not edit the class manually.
@@ -5,7 +5,7 @@

    HTTP API for Hindsight

-    The version of the OpenAPI document: 0.4.2
+    The version of the OpenAPI document: 0.4.9
    Generated by OpenAPI Generator (https://openapi-generator.tech)

    Do not edit the class manually.
@@ -5,7 +5,7 @@

    HTTP API for Hindsight

-    The version of the OpenAPI document: 0.4.2
+    The version of the OpenAPI document: 0.4.9
    Generated by OpenAPI Generator (https://openapi-generator.tech)

    Do not edit the class manually.
@@ -5,7 +5,7 @@

    HTTP API for Hindsight

-    The version of the OpenAPI document: 0.4.2
+    The version of the OpenAPI document: 0.4.9
    Generated by OpenAPI Generator (https://openapi-generator.tech)

    Do not edit the class manually.
--- a/Show More
+++ b/Show More