B2: ship Hermes auto-capture pipeline (#146)

Co-authored-by: Sami Rusani <sr@samirusani>

Author: samrusani

ARCHITECTURE.md

@@ -32,9 +32,11 @@ This keeps Alice continuity semantics centralized while adding Hermes-native aut
 Implemented in B1:
 - `alice_prefetch_context`
 
-Planned for B2+ (not shipped in B1):
+Implemented in B2:
 - `alice_capture_candidates`
 - `alice_commit_captures`
+
+Planned for B3+:
 - `alice_session_flush`
 - `alice_review_queue`
 - `alice_review_apply`
@@ -46,17 +48,17 @@ Planned for B2+ (not shipped in B1):
 3. Alice returns compact continuity payload (summary, relevant memories, open loops, recent decisions/changes, optional resume brief, trust posture).
 4. Provider injects payload into Hermes context before model response.
 
-### Flow 2: Post-response Candidate Extraction (planned B2+)
+### Flow 2: Post-response Candidate Extraction (implemented in B2)
 1. Hermes emits assistant response.
 2. Provider sends user/assistant turn pair to `alice_capture_candidates`.
 3. Alice returns typed candidates with confidence, trust class, proposed action, evidence snippet, and target object type.
 
-### Flow 3: Post-response Commit (planned B2+)
+### Flow 3: Post-response Commit (implemented in B2)
 1. In `assist`/`auto`, provider submits candidates to `alice_commit_captures`.
 2. Policy gates auto-save vs review queue using type allowlist + confidence threshold.
 3. Writes are idempotent across repeated sync attempts.
 
-### Flow 4: Session-End Flush (planned B2+)
+### Flow 4: Session-End Flush (planned B3+)
 1. On session end, provider calls `alice_session_flush`.
 2. Alice performs dedupe merge, contradiction checks, open-loop normalization, summary refresh, and review queue updates.
 

BUILD_REPORT.md

@@ -1,73 +1,83 @@
 # BUILD_REPORT
 
 ## sprint objective
-Implement Bridge Sprint 1 (`B1`) Hermes provider contract foundation by extending the shipped provider with bridge-phase config normalization, deterministic lifecycle hooks (`prefetch`, `queue_prefetch`, `sync_turn`, `on_session_end`), provider status/readiness validation, and the new automation-oriented MCP prefetch surface `alice_prefetch_context`.
+Implement Bridge Sprint 2 (`B2`) auto-capture pipeline on top of the shipped Hermes provider and B1 contract foundation: candidate extraction, commit policy, mode support (`manual`, `assist`, `auto`), review-queue persistence for non-auto-saved items, and deterministic idempotent/no-op behavior.
 
 ## completed work
-- Extended (not replaced) the shipped Hermes Alice provider with bridge contract behavior:
-  - normalized canonical bridge config keys for lifecycle operation
-  - preserved legacy config-key compatibility
-  - added provider status/readiness reporting without live network dependency
-  - added deterministic session-end flush behavior
-  - added duplicate-write suppression for repeated `sync_turn` and `on_memory_write` callback execution
-  - fixed capture-queue worker start race and bounded dedupe behavior to avoid suppressing valid later repeats
-- Added MCP tool `alice_prefetch_context` with deterministic prefetch text assembled from existing continuity resumption semantics.
-- Updated smoke script to emit provider bridge status/lifecycle readiness evidence.
-- Updated install and integration docs for bridge contract keys, lifecycle mapping, compatibility keys, and MCP tool surface.
-- Added/updated unit and integration coverage for:
-  - provider config compatibility and invalid-config status
-  - deterministic lifecycle dedupe/flush behavior
-  - MCP tool surface stability including `alice_prefetch_context`
-- Aligned control-doc truth checks and active control docs to the current bridge-phase baseline.
-- Clarified architecture docs so B2+ capture/review surfaces are marked as planned (not shipped in B1).
-
-Legacy compatibility keys preserved:
-- `prefetch_limit`
-- `max_recent_changes`
-- `max_open_loops`
-- `include_non_promotable_facts`
-- `auto_capture`
-- `mirror_memory_writes`
+- Added B2 capture pipeline core in Alice continuity:
+  - implemented `alice_capture_candidates` extraction from user/assistant turn pairs
+  - implemented `alice_commit_captures` commit policy over extracted candidates
+  - implemented candidate classes: `decision`, `commitment`, `waiting_for`, `blocker`, `preference`, `correction`, `note`, `no_op`
+  - candidate payloads now include confidence, trust class, evidence snippet, and proposed action
+- Implemented commit policy operating modes:
+  - `manual`: routes non-`no_op` candidates to review persistence
+  - `assist`: auto-saves only explicit high-confidence allowlist candidates
+  - `auto`: auto-saves allowlist candidates at the auto-mode confidence gate
+- Implemented policy allowlist and review routing evidence:
+  - auto-save allowlist categories: `correction`, `preference`, `decision`, `commitment`, `waiting_for`, `blocker`
+  - review-routed categories by type policy: `note`
+  - additionally, low-confidence or policy-disallowed candidates route to review under mode gates
+- Added idempotent commit behavior using commit fingerprint + candidate fingerprint lookup to prevent duplicate writes on repeated sync attempts.
+- Added no-op protection so no-op turns (`no_op`) produce no memory writes.
+- Wired new HTTP surfaces:
+  - `POST /v0/continuity/captures/candidates`
+  - `POST /v0/continuity/captures/commit`
+- Wired new MCP surfaces:
+  - `alice_capture_candidates`
+  - `alice_commit_captures`
+  - preserved existing `alice_capture` and other shipped tools for fallback/manual workflows
+- Wired Hermes provider B2 flow in `sync_turn`:
+  - `assist`/`auto` modes now run candidate extraction then commit
+  - `manual` mode suppresses automatic `sync_turn` capture
+  - fallback to legacy `/v0/continuity/captures` path when B2 endpoints are unavailable
+  - preserved dedupe queue and session-end flush behavior
+- Updated sprint-scoped integration docs and smoke script for B2 mode/pipeline truth.
+- Updated control-doc truth checker markers from B1-active to B2-active so required verification reflects active sprint state.
 
 ## incomplete work
-- No code deliverable remains incomplete in B1 scope.
+- None in B2 packet scope.
 
 ## files changed
 - `ARCHITECTURE.md`
 - `BUILD_REPORT.md`
 - `PRODUCT_BRIEF.md`
 - `README.md`
+- `REVIEW_REPORT.md`
 - `ROADMAP.md`
-- `RULES.md`
-- `docs/integrations/hermes-memory-provider/plugins/memory/alice/__init__.py`
+- `apps/api/src/alicebot_api/continuity_capture.py`
+- `apps/api/src/alicebot_api/contracts.py`
+- `apps/api/src/alicebot_api/main.py`
 - `apps/api/src/alicebot_api/mcp_tools.py`
+- `apps/api/src/alicebot_api/store.py`
+- `docs/integrations/hermes-memory-provider/plugins/memory/alice/__init__.py`
 - `docs/integrations/hermes-memory-provider.md`
 - `docs/integrations/mcp.md`
-- `scripts/check_control_doc_truth.py`
-- `scripts/install_hermes_alice_memory_provider.py`
 - `scripts/run_hermes_memory_provider_smoke.py`
-- `tests/unit/test_hermes_memory_provider.py`
+- `scripts/check_control_doc_truth.py`
+- `tests/unit/test_continuity_capture.py`
+- `tests/integration/test_continuity_capture_api.py`
 - `tests/unit/test_mcp.py`
-- `tests/integration/test_mcp_server.py`
+- `tests/unit/test_hermes_memory_provider.py`
 
 ## tests run
 1. `python3 scripts/check_control_doc_truth.py`
    - Result: PASS
-   - Output: `Control-doc truth check: PASS`
+   - Output summary: verified README, ROADMAP, sprint packet, RULES, current state, archive planning marker
 2. `./.venv/bin/python -m pytest tests/unit tests/integration -q`
    - Result: PASS
-   - Output: `1174 passed in 186.28s (0:03:06)`
+   - Output: `1188 passed in 191.85s (0:03:11)`
 3. `./.venv/bin/python scripts/run_hermes_memory_provider_smoke.py`
    - Result: PASS
    - Output summary:
      - single external provider enforcement validated
-     - provider registered with expected tool set
-     - `structural.bridge_status.ready=true`
-     - `structural.bridge_status.errors=[]`
-     - lifecycle hooks status present for `prefetch`, `queue_prefetch`, `sync_turn`, `on_session_end`
+     - provider registered with expected tool schemas
+     - `bridge_contract_version` reported as `bridge_b2`
+     - bridge status `ready=true`, `errors=[]`
+     - config includes `bridge_mode=assist`
+     - lifecycle hooks report `prefetch`, `queue_prefetch`, `sync_turn`, `on_session_end`, and `bridge_mode`
 
 ## blockers/issues
 - None.
 
 ## recommended next step
-Proceed with reviewer re-check against B1 acceptance criteria using this updated evidence set.
+Run B2 review against acceptance criteria with focus on policy calibration (confidence thresholds) and confirm desired `auto`-mode aggressiveness before promoting B3 review actions.

PRODUCT_BRIEF.md

@@ -70,7 +70,7 @@ Review-required:
 - low-confidence extractions
 
 ## Active Sprint Status
-Bridge Sprint 1 (`B1`) is now the active execution sprint. It is limited to provider-contract foundation work on top of the shipped Hermes provider surface.
+Bridge Sprint 2 (`B2`) is now the active execution sprint. It is limited to the auto-capture pipeline on top of the shipped Hermes provider surface and the `B1` contract foundation.
 
 ## Known Gaps To Resolve Before Build
 - Candidate scoring rubric and confidence calibration method are not specified.

README.md

@@ -39,7 +39,8 @@ Phase 11 is complete and shipped:
 - `P11-S6` Model Packs Tier 2 + Launch Clarity Assets is shipped
 - `P11-R1` Provider Runtime Hardening is shipped
 - A bridge phase is now active: Hermes Auto-Capture
-- `B1` Hermes Provider Contract Foundation is the active sprint
+- `B1` Hermes Provider Contract Foundation is shipped
+- `B2` Auto-Capture Pipeline is the active sprint
 - Historical planning and control docs: [docs/archive/planning/2026-04-08-context-compaction/README.md](docs/archive/planning/2026-04-08-context-compaction/README.md)
 
 ## Why Alice exists

REVIEW_REPORT.md

@@ -4,40 +4,50 @@
 PASS
 
 ## criteria met
-- The shipped Hermes provider was extended, not replaced (`docs/integrations/hermes-memory-provider/plugins/memory/alice/__init__.py`).
-- Bridge config normalization and readiness/status reporting are present, including legacy compatibility reporting.
-- Lifecycle hooks `prefetch`, `queue_prefetch`, `sync_turn`, and `on_session_end` are implemented with deterministic behavior, and duplicate callback suppression is covered.
-- New MCP tool `alice_prefetch_context` is implemented, wired, and documented (`apps/api/src/alicebot_api/mcp_tools.py`, `docs/integrations/mcp.md`).
-- Existing Hermes tools (`alice_recall`, `alice_resumption_brief`, `alice_open_loops`) remain intact.
-- B1 required verification commands all pass:
-  - `python3 scripts/check_control_doc_truth.py` -> PASS
-  - `./.venv/bin/python -m pytest tests/unit tests/integration -q` -> `1174 passed in 186.28s (0:03:06)`
-  - `./.venv/bin/python scripts/run_hermes_memory_provider_smoke.py` -> PASS
-- No local identifiers were found in changed code/docs after fixes.
+- `alice_capture_candidates` is implemented and wired through API and MCP.
+- `alice_commit_captures` is implemented and wired through API and MCP.
+- Commit operating modes are implemented and exercised: `manual`, `assist`, `auto`.
+- Candidate classes required by B2 are present: `decision`, `commitment`, `waiting_for`, `blocker`, `preference`, `correction`, `note`, `no_op`.
+- Candidate outputs include confidence, trust class, evidence snippet, and proposed action.
+- Auto-save allowlist categories are explicit and implemented: `correction`, `preference`, `decision`, `commitment`, `waiting_for`, `blocker`.
+- Review-routed category by type policy is explicit and implemented: `note`.
+- Low-confidence/policy-gated candidates route to review persistence rather than auto-save.
+- No-op turns produce no memory writes.
+- Repeated sync attempts are idempotent via `(sync_fingerprint, candidate_id)` duplicate guard.
+- Hermes provider uses the B2 pipeline in `assist`/`auto`, preserves `manual` behavior, and keeps legacy fallback.
+- B2 docs/code do not claim `alice_review_queue` or `alice_review_apply` as shipped.
+- Previously flagged gaps are fixed:
+  - explicit `auto` mode tests added
+  - strict boolean validation for candidate `explicit` added
+  - `ARCHITECTURE.md` bridge status updated to reflect B2 implementation
+- Local identifier hygiene check on changed files passed (no local machine paths/usernames leaked in changed code/docs/reports).
 
 ## criteria missed
 - None.
 
 ## quality issues
-- No blocking quality issues found after fixes.
-- Minor residual note: callback dedupe now uses a short bounded window; if callback replay behavior changes materially in Hermes, this window may need retuning.
+- No blocking quality issues found in B2 scope after fixes.
 
 ## regression risks
-- Low.
-- Primary risk area remains concurrent lifecycle callback timing; current unit coverage and full suite pass reduce risk materially.
+- Moderate-low risk in policy calibration (confidence thresholds), but implementation behavior is deterministic and covered by unit/integration tests.
+- Idempotency/no-op regressions are specifically covered.
 
 ## docs issues
-- Previously identified doc/report inconsistencies were corrected:
-  - architecture now marks B2+ surfaces as planned
-  - build report file list and command results are aligned with current state
-- Local identifier hygiene: PASS.
+- None blocking.
+- Bridge status documentation is now aligned with B2 implementation state.
 
 ## should anything be added to RULES.md?
-- Optional improvement: add a permanent rule that sprint reports must match `git diff --name-only` for the sprint-owned file list.
+- No required change.
 
 ## should anything update ARCHITECTURE.md?
-- No further update required for B1 acceptance after current “Implemented in B1” vs “Planned for B2+” separation.
+- Completed in this fix pass: B2 surfaces and runtime flow status markers were updated from planned to implemented where applicable.
 
 ## recommended next action
-1. Approve B1 for merge review.
-2. Carry the new capture-queue/dedupe tests forward as required regression coverage for future bridge sprints.
+1. Approve B2 for merge.
+2. Start B3 review-action scope (`alice_review_queue`, `alice_review_apply`) using B2 persisted review items as baseline fixtures.
+
+## evidence summary
+- Required verification commands (re-run):
+  - `python3 scripts/check_control_doc_truth.py` -> PASS
+  - `./.venv/bin/python -m pytest tests/unit tests/integration -q` -> `1188 passed in 191.85s (0:03:11)`
+  - `./.venv/bin/python scripts/run_hermes_memory_provider_smoke.py` -> PASS

ROADMAP.md

@@ -8,7 +8,7 @@
 Phase 11 remains baseline truth and is not future scope.
 
 ## Active Planning Status
-- Bridge Sprint 1 (`B1`) is the active execution sprint.
+- Bridge Sprint 2 (`B2`) is the active execution sprint.
 - The remaining bridge-phase milestones are planned but not yet promoted.
 
 ## Bridge Phase: Hermes Auto-Capture (Planned)