Annotation-export curation loop, person merge service, and August sync

AGENTS.md

@@ -54,6 +54,7 @@ Chronicle includes an **interactive setup wizard** for easy configuration. The w
 - Memory configuration (agentic Markdown vault — Chronicle's single memory provider)
 - Network configuration and HTTPS setup
 - Optional services (speaker recognition, Parakeet ASR)
+- Immich photo library integration (photo discovery + person photos in the vault)
 
 ### Quick Start
 ```bash
@@ -128,7 +129,11 @@ All test operations are managed through a simple Makefile interface:
 cd tests
 
 # Full test workflow (recommended)
-make test              # Start containers + run all tests
+make test              # Start containers + run all tests (profile: mock, no credentials)
+
+# Same suite against real backing services (see tests/profiles.yml)
+make test PROFILE=deepgram-openai          # real Deepgram STT + real OpenAI LLM
+make test PROFILE=deepgram-openai-speaker  # ...plus the real speaker service
 
 # Or step by step
 make start             # Start test containers (with health checks)
@@ -588,6 +593,7 @@ tailscale ip -4
 
 ### Testing Strategy
 - **Makefile-Based**: All test operations through simple `make` commands (`make test`, `make start`, `make stop`)
+- **One suite, N service profiles**: tests are never selected by whether an API key is present. `tests/profiles.yml` declares which backing services are real for a run; stubs replay recorded real responses from `tests/cassettes/`, so the same assertions hold with or without credentials. Do not add a tag or a skip to work around a missing key — record a cassette (`make record-cassettes`) or fix the stub.
 - **Log Preservation**: Container logs always saved before cleanup (never lose debugging info)
 - **End-to-End Integration**: Robot Framework validates complete audio processing pipeline
 - **Environment Flexibility**: Tests work with both local .env files and CI environment variables
@@ -634,6 +640,7 @@ For detailed technical documentation, see:
 - **[@docs/podman.md](docs/podman.md)**: Running with Podman instead of Docker (engine selection, rootless/GPU setup)
 - **[@docs/screenpipe.md](docs/screenpipe.md)**: ScreenPipe capture-node architecture, services, desktop controls, and troubleshooting
 - **[@docs/audio-pipeline-architecture.md](docs/audio-pipeline-architecture.md)**: Audio pipeline design
+- **[@docs/backend/compose-stack.md](docs/backend/compose-stack.md)**: Backend compose services, shared mounts, and profiles
 - **[@docs/backend/auth.md](docs/backend/auth.md)**: Authentication architecture
 - **[@docs/backend/memories.md](docs/backend/memories.md)**: Memory system documentation
 - **[@docs/backend/plugin-development-guide.md](docs/backend/plugin-development-guide.md)**: Plugin development guide

backends/advanced/Dockerfile

@@ -22,7 +22,7 @@ COPY pyproject.toml uv.lock ./
 
 # Build a project .venv from the lockfile (includes git deps like Graphiti)
 RUN --mount=type=cache,target=/root/.cache/uv \
-    uv sync --frozen --no-dev --extra deepgram --extra galileo --extra benchmark --no-install-project
+    uv sync --frozen --no-dev --extra galileo --extra benchmark --no-install-project
 
 
 # ============================================
@@ -138,7 +138,7 @@ WORKDIR /app
 COPY --from=ghcr.io/astral-sh/uv:latest /uv /uvx /bin/
 COPY pyproject.toml uv.lock ./
 RUN --mount=type=cache,target=/root/.cache/uv \
-    uv sync --frozen --extra deepgram --extra galileo --extra benchmark --group test --no-install-project && \
+    uv sync --frozen --extra galileo --extra benchmark --group test --no-install-project && \
     rm /bin/uv /bin/uvx
 
 ENV VIRTUAL_ENV=/app/.venv

backends/advanced/Dockerfile.k8s

@@ -23,12 +23,12 @@ COPY uv.lock .
 RUN mkdir -p src/advanced_omi_backend
 COPY src/advanced_omi_backend/__init__.py src/advanced_omi_backend/
 
-# Install dependencies using uv with deepgram extra
+# Install dependencies using uv
 # Use cache mount for BuildKit, fallback for legacy builds
 # RUN --mount=type=cache,target=/root/.cache/uv \
-#     uv sync --extra deepgram
+#     uv sync
 # Fallback for legacy Docker builds (CI compatibility)
-RUN uv sync --extra deepgram
+RUN uv sync
 
 # Copy all application code
 COPY . .

backends/advanced/Docs/data-audit.md

@@ -135,6 +135,36 @@ records by `clip_id` + `conversation_id` + source times. Helpers live in
 the Data Audit toolbar (exports selected rows; modal also lists/downloads/
 deletes server-side exports).
 
+### Contents preview (verify before shipping)
+
+`POST /export/preview` (same body minus `sensitivity_policy`) is a synchronous
+dry-run: it returns the exact clips the settings would produce — boundaries,
+durations, and the sliced transcript each manifest record would carry —
+without writing any audio. Preview and job share one code path
+(`utils/export_planning.plan_conversation_clips`), so the preview cannot drift
+from the export. Unanalyzed conversations are reported skipped
+(`not analyzed`) rather than analyzed inline; the export job still runs VAD
+for them.
+
+The modal's **Dataset contents** panel renders this plan live (auto-refreshing
+as params or screen withholdings change): per-clip play (gapless player over
+the clip's range), the transcript slice (a `no transcript` badge marks clips
+an annotator would receive silent), and an include checkbox per clip. Unticked
+clips are passed to `/export` as `dropped_ranges`
+(`{conversation_id: [[start, end], …]}`) and carved out exactly like privacy
+ranges but accounted separately (`params.curated`, per-conversation and total
+`dropped_seconds`): one bucket means "too sensitive to share", the other "not
+worth annotating".
+
+### Export history
+
+The listing joins the on-disk `export.json` metadata to mark conversations a
+previous export actually shipped (skipped ones don't count): each row carries
+`last_export` (chipped `exported` in the table), and the `exported=never|exported`
+filter scopes a curation session to un-shipped audio. Reading history from the
+export directories means deleting an export naturally un-marks its
+conversations — no second source of truth.
+
 ## Privacy screen (shareability gate)
 
 Before sharing audio + transcripts with an outside annotator, the export can

backends/advanced/docker-compose-test.yml

@@ -166,10 +166,12 @@ services:
     build:
       context: ../..
       dockerfile: tests/Dockerfile.mock-streaming-stt
+    # 8879 on purpose: the old 9999 is a popular default port for other
+    # tools and collided with unrelated services on a dev host.
     ports:
-      - "9999:9999"
+      - "8879:8879"
     healthcheck:
-      test: ["CMD", "python", "-c", "import socket; s=socket.socket(); s.connect(('localhost',9999)); s.close()"]
+      test: ["CMD", "python", "-c", "import socket; s=socket.socket(); s.connect(('localhost',8879)); s.close()"]
       interval: 10s
       timeout: 5s
       retries: 3