feat: workflow intelligence sidecar (v1, apagado por defecto)

CHANGELOG.md

@@ -1,11 +1,30 @@
-<!-- doc-head: public MemoryMaster release history through v4.8.4 -->
-Covers user-visible changes, migrations, governance boundaries, and operational fixes for each release.
-Key terms: graph observations, compiled profile, governed capture, recall, sensitivity, lineage.
+<!-- doc-head: public history through v4.8.4 plus unreleased Workflow Intelligence -->
+Covers user-visible changes, migrations, governance boundaries, workflow analytics, and operational fixes.
+Key terms: workflow intelligence, rule observations, graph observations, governed capture, recall, lineage.
 Read when upgrading MemoryMaster, preparing release notes, or checking migration and rollback impact.
 <!-- /doc-head -->
 
 # Changelog
 
+## [Unreleased]
+
+- Added on-demand **Workflow Intelligence** for Claude Code, Codex, and
+  Wezbridge histories. It uses a rebuildable WAL-mode sidecar, bounded redacted
+  excerpts, deterministic trajectory signals, self-contained local reports,
+  inert candidates, and explicit human review. LLM classification is opt-in and
+  cannot establish success or recurrence.
+- Added migration `0024_rule_observation_lineage` with SQLite/Postgres parity.
+  New governed skill candidates require three distinct human root sessions;
+  user/global patterns also require two projects. Same-session retries and
+  subagent/automation activity do not inflate independent support.
+- Added a fail-soft, content-free completion receipt hook. It ships
+  unregistered and defaults to `off`; shadow setup is explicit, advisory setup
+  fails closed until the 14-day/100-receipt/provider/precision gate passes, and
+  no blocking or automatic policy promotion exists.
+- This entry is implementation evidence only. No package release, hook
+  installation, scheduler activation, deployment, or external acceptance is
+  implied.
+
 ## [4.8.4] - 2026-08-19
 
 - **A correction aimed at another agent's claim is recorded instead of

DOCS-MAP.md

@@ -1,24 +1,26 @@
-<!-- doc-head: canonical map through v4.8.4; Dreaming V1 stays shadow by operator ruling -->
+<!-- doc-head: canonical map through the unreleased Workflow Intelligence implementation -->
 # DOCS-MAP - memorymaster
 # Covers: trust verdicts and replacements for every canonical documentation surface.
-# Key terms: CURRENT, SUPERSEDED, compiled user profile, roadmap, ADR.
+# Key terms: CURRENT, SUPERSEDED, workflow intelligence, compiled user profile, roadmap, ADR.
 # Read when: locating authoritative project documentation before reading doc bodies.
-# Updated: 2026-08-21 for version truthing; GLM retired, Dreaming shadow-only.
+# Updated: 2026-08-30 for Workflow Intelligence, rule lineage, and sidecar ADR.
 # Rule: generated profiles are disposable projections; MemoryMaster remains authority.
 <!-- /doc-head -->
 
 | File | Verdict | Last change | Reason |
 |---|---|---|---|
 | CHANGELOG.md | CURRENT | 2026-08-19 | Public release history through v4.8.4, matching `pyproject.toml`. Earlier heads said v4.7.5; the entries were there, the verdict line was stale. |
-| README.md | CURRENT | 2026-08-13 | Primary product overview and setup for governed claims, PPR-7 observations, compiled profile, and release-first scheduled review. Its "Gemini plus GLM" wording predates the 2026-08-21 GLM retirement (#226) and needs a pass. |
-| ROADMAP.md | CURRENT | 2026-08-21 | Sole authoritative roadmap; P5 merged and locally deployed, isolated PPR-7 active, and Dreaming V1 recorded as the main active workstream held in shadow by operator ruling. |
+| README.md | CURRENT | 2026-08-30 | Primary product overview for governed claims plus the optional non-authoritative Workflow Intelligence sidecar; hook activation remains separate. |
+| ROADMAP.md | CURRENT | 2026-08-30 | Sole authoritative roadmap; records Workflow Intelligence as locally implemented but unreleased/uninstalled and preserves separate shadow/advisory decisions. |
+| docs/workflow-intelligence.md | CURRENT | 2026-08-30 | Operator and architecture contract for local trajectory analytics, deterministic evidence, inert candidates, and the off/shadow/advisory receipt gate. |
+| docs/adr/0016-workflow-analytics-uses-disposable-sidecar.md | CURRENT | 2026-08-30 | Accepted boundary: authoritative memory remains one database while rebuildable workflow analytics use a non-authoritative local sidecar. |
 | .planning/GRAPH-OBSERVATIONS-V1.md | CURRENT | 2026-08-12 | Bounded PPR-7 implementation ledger for deterministic supported components, governed synthesis, lifecycle, opt-in recall, and verification. |
 | .planning/COMPILED-USER-PROFILE-V1.md | CURRENT | 2026-08-12 | Implementation contract for the zero-curation, evidence-bound user profile projection and weekly GLM map/reduce job. |
 | .planning/PAPER-RADAR-REVIEW-2026-08-08.md | CURRENT | 2026-08-08 | Primary-paper ledger covers 57-paper triage, 18 deep reviews, exact MemoryMaster gaps, and ordered PPR-1 through PPR-6 decisions subordinate to ROADMAP.md. |
 | .planning/PAPER-RESEARCH-IMPLEMENTATION-2026-08-08.md | CURRENT | 2026-08-08 | Executable status ledger for PPR-1 through PPR-6; records acceptance criteria and evidence without competing with ROADMAP.md. |
 | .planning/HERMES-SCOPE-SKILLS-INTEGRATION-2026-08-07.md | CURRENT | 2026-08-12 | Executable Tencent-derived ledger; repair5 passed and PR #189 is open without merge authority. |
 | .planning/audits/2026-08-07-hermes-scope-skills/audit-delta.md | CURRENT | 2026-08-12 | Bounded delta preserves failed runs and records repair5 pass plus PR #189 creation. |
-| docs/governed-skills.md | CURRENT | 2026-08-08 | Operator guide for proposal/promotion, progressive confirmed-skill recall, isolation, and staging-only export. |
+| docs/governed-skills.md | CURRENT | 2026-08-30 | Operator guide for three-human-root recurrence, human promotion, confirmed-skill recall, isolation, and staging-only export. |
 | .planning/AUTORESEARCH-PROGRAM-2026-08-03.md | CURRENT | 2026-08-04 | Completed six-phase execution overlay records retrieval, graph, capture, full-QA, OAuth quality, public release, and the sole remaining longitudinal gate. |
 | .planning/audits/2026-08-04-autoresearch-convergence/audit-delta.md | CURRENT | 2026-08-04 | Bounded convergence delta records SQLite-only tests, 40-case quality evidence, public v4.6.0 release evidence, and the remaining seven-day observation. |
 | COMPETITOR_ANALYSIS.md | CURRENT | 2026-07-27 | Current prior-art analysis corrects Cognee provenance, temporal, tenant, graph, and document capabilities without adopting it as a dependency. |
@@ -57,7 +59,7 @@
 | .planning/audits/2026-07-14-phase4-budget-delta/production-release-verification.md | CURRENT | 2026-07-14 | This is a stable historical verification record for the released v4.5.0 version, documenting immutable facts like commit hashes and CI status.  |
 | docs/handbook.md | CURRENT | 2026-07-14 | Strong cross-references from root files and active maintenance indicators (v4.2 changelog, changelog mentions) confirm this is the canonical operational guide.  |
 | .planning/codebase/CONVENTIONS.md | CURRENT | 2026-06-09 | The document explicitly states it was regenerated on 2026-06-09 from the current tree (v3.28.0), making it a living reference for project standards.  |
-| docs/adr/0001-local-first-single-database.md | CURRENT | 2026-07-14 | The document records a stable architectural decision (Local-First SQLite) that was reaffirmed recently (2026-07-14) and defines the core persistence model for the project.  |
+| docs/adr/0001-local-first-single-database.md | CURRENT | 2026-08-30 | Governing data stays in one logical store; ADR-0016's disposable analytics sidecar is explicitly excluded from recall and correctness paths. |
 | docs/archive/security-audit/threat-model.md | CURRENT | 2026-06-20 | The file is a specific, dated security audit report and threat model that remains an accurate snapshot of the system's architecture and potential vulnerabilities as of the last commit.  |
 | docs/atlas-inbox-v1-technical-spec.md | CURRENT | 2026-05-07 | The document serves as the authoritative implementation plan for the V1 Atlas Inbox feature, mapping PRD requirements to existing MemoryMaster primitives and defining the data model for WhatsApp ingestion and action proposals.  |
 | docs/cross-project-patterns.md | CURRENT | 2026-07-14 | The document is behavior-derived, cites specific implementation lines and lifecycle rules, and describes R1.3 features like Qdrant quarantine and scope allowlists that appear to be the active architecture.  |
@@ -78,7 +80,7 @@
 | docs/archive/security-audit-2026-05-11.md | CURRENT | 2026-06-20 | The content is a dated but highly specific and technical security audit report that remains accurate as a historical record and reference for vulnerabilities unless explicitly fixed.  |
 | .planning/REMEDIATION-EXECUTION-V3-BUDGET.md | CURRENT | 2026-07-14 | The document contains a completed execution record from 2026-07-12, finalizing Phase 1 and explicitly stopping; it serves as the controlling authority for the V3 budget mode and audit delta.  |
 | .planning/codebase/INTEGRATIONS.md | CURRENT | 2026-06-09 | The document was regenerated recently (2026-06-09) from the current tree (v3.28.0), accurately reflecting the existing architecture, file paths, and integration logic.  |
-| docs/adr/README.md | CURRENT | 2026-05-12 | This index is the canonical navigation hub for 14 numbered architectural decision records and dated notes, and is actively referenced by the project root.  |
+| docs/adr/README.md | CURRENT | 2026-08-30 | Canonical navigation for 16 numbered decisions, including the disposable workflow-sidecar boundary. |
 | docs/adr/0012-readme-stays-concise-handbook-holds-operator-depth.md | CURRENT | 2026-05-11 | This is an accepted architectural decision record (ADR) establishing a permanent documentation split strategy that remains relevant for project organization.  |
 | .planning/P2-CENSUS.md | CURRENT | 2026-06-15 | Recent (June 2026) snapshot census with actionable verdicts and specific code insights for the P2 restructure.  |
 | docs/adr/0006-sensitivity-filter-boundary.md | CURRENT | 2026-05-11 | The document establishes a stable architectural boundary between raw source storage and sanitized claims storage that remains applicable to the project's design.  |

README.md

@@ -1,6 +1,6 @@
-<!-- doc-head: MemoryMaster 4.8 governed memory, observations, profiles, and operations -->
-<!-- Covers: installation, governed capture and recall, graph observations, compiled profile, scheduling, and safety. -->
-<!-- Key terms: claims, citations, steward, graph observations, compiled profile, Gemini, Antigravity CLI, SQLite, MCP. -->
+<!-- doc-head: MemoryMaster 4.8 governed memory, observations, profiles, workflow analytics, and operations -->
+<!-- Covers: installation, governed recall, graph observations, compiled profile, Workflow Intelligence, scheduling, and safety. -->
+<!-- Key terms: claims, citations, graph observations, workflow sidecar, compiled profile, Gemini, SQLite, MCP. -->
 <!-- Read when: evaluating, installing, operating, or upgrading MemoryMaster. -->
 <!-- Default: private local SQLite; observations and generated views never bypass claim governance. -->
 <!-- /doc-head -->
@@ -39,6 +39,7 @@ evidence -> candidate claim -> steward -> confirmed claim -> governed recall
 | Governed skills | Stores reviewed reusable procedures separately from ordinary facts | Recall opt-in |
 | Hermes integration | Provides exact session/project scoping, local HTTP/stdio compatibility and replay-safe outbox behavior | Optional integration |
 | Operational review | Performs a read-only six-hour integrity, queue, profile, intake and retrieval review | Optional Windows task |
+| Workflow Intelligence | Mines Claude/Codex trajectories into local redacted analytics and inert improvement candidates | On demand; separate sidecar, no automatic promotion |
 
 MemoryMaster 4.8 uses the configured **Gemini extraction + Gemini consolidation**
 path for this installation; consolidation moved off GLM on 2026-08-20 when that
@@ -111,6 +112,24 @@ queued = improve(scope="project:atlas")
 
 See [Public v1](docs/public-v1.md) for receipts and full parameter contracts.
 
+## Workflow Intelligence
+
+The optional `workflow` command family audits retained Claude Code, Codex, and
+Wezbridge trajectories in a rebuildable local sidecar. It does not write claims,
+modify agent instructions, or treat an LLM judgment as proof of success.
+
+```powershell
+memorymaster workflow scan --deep human
+memorymaster workflow report
+memorymaster workflow candidates
+```
+
+LLM classification is a separate opt-in command. The provider-neutral
+completion receipt hook ships unregistered and defaults to `off`; advisory mode
+cannot be configured until its 14-day shadow evidence gate passes. See
+[Workflow Intelligence](docs/workflow-intelligence.md) for the schema, commands,
+redaction boundary, recurrence rules, and rollout gate.
+
 ## Graph observations (PPR-7)
 
 > **Off by default.** Enable with `MEMORYMASTER_GRAPH_OBSERVATIONS=1`.
@@ -216,8 +235,10 @@ producer
   -> optional observations / skills / compiled profile
 ```
 
-The default product is one private local SQLite database in WAL mode plus a
-stdio MCP server. No new database, cloud service or vector server is required.
+The authoritative product is one private local SQLite database in WAL mode plus
+a stdio MCP server. Optional Workflow Intelligence uses a disposable analytics
+sidecar that cannot participate in recall or lifecycle authority. No cloud
+service or vector server is required.
 PostgreSQL team operation remains explicitly deferred; Qdrant remains an
 optional semantic accelerator rather than a source of truth.
 

ROADMAP.md

@@ -1,7 +1,7 @@
-<!-- doc-head: sole roadmap; PPR-7 integrated, compiled profile in development -->
+<!-- doc-head: sole roadmap; Workflow Intelligence implemented locally and held off -->
 # MemoryMaster roadmap
-# Covers: post-v4.6 sequence, governed observations, compiled profile, and deferrals.
-# Key terms: Hermes, governed skills, graph observations, user profile, sustainability.
+# Covers: post-v4.6 sequence, workflow analytics, governed observations, compiled profile, and deferrals.
+# Key terms: Workflow Intelligence, governed skills, graph observations, user profile, sustainability.
 # Read when: choosing release scope, accepting a feature, or checking deferrals.
 # Authority: sole roadmap; planning ledgers implement it and never replace it.
 # Safety: SQLite remains authority; generated user.md is disposable and feature-off by default.
@@ -73,6 +73,12 @@ its governed-claims authority:
   loops, hardening live extraction, and closing the first-soak gaps. Activation
   is a separate operator decision that has not been made, so no rollout date
   belongs here.
+- Workflow Intelligence v1 is implemented on an isolated feature branch: local
+  Claude/Codex/Wezbridge census, deterministic trajectory analysis, bounded
+  opt-in classification, local reports, inert candidates, independent-session
+  rule governance, and an off-by-default receipt hook. Implementation and tests
+  do not authorize installing the hook, starting its 14-day shadow window,
+  releasing 4.9.0, deploying, or promoting any candidate.
 - PPR-7 governed graph observations are implemented and locally verified on an
   isolated feature branch. Implementation is additive, candidate-first, and
   opt-in; its separate PR/CI, feature-off deployment, and rollout evidence are
@@ -91,6 +97,13 @@ its governed-claims authority:
 
 ## Next
 
+- Review and integrate Workflow Intelligence without activating it. After a
+  separately approved shadow installation, collect at least 14 days and 100
+  eligible receipts (20 per Claude and Codex), manually review warnings, require
+  at least 90% precision and zero read-only false positives, then request a
+  separate advisory-mode decision. Never auto-promote analytics into rules or
+  skills.
+
 - Complete the bounded session-scope, native Hermes MemoryProvider, and
   governed-skill proposal program defined by
   `.planning/HERMES-SCOPE-SKILLS-INTEGRATION-2026-08-07.md`; Windows SQLite

benchmarks/bench_transcript_confidence.py

@@ -0,0 +1,66 @@
+"""Micro-benchmark del costo por llamada de `_transcript_confidence`.
+
+Aisla el mecanismo que rompio el SLO de `cycle_p95` en el PR #248: la funcion
+se llama DENTRO de un bucle de minado y, desde que el registro de linaje se
+puso antes del gate de bootstrap, abre y cierra una conexion SQLite en cada
+iteracion. Antes del cambio, con bootstrap deshabilitado —el default— la
+funcion retornaba sin tocar la base.
+
+Mide el camino real, no un proxy: mismo servicio, misma base en disco, mismas
+kwargs de linaje. Correr antes y despues del arreglo.
+
+Uso: python benchmarks/bench_transcript_confidence.py [--calls 200]
+"""
+from __future__ import annotations
+
+import argparse
+import sys
+import tempfile
+import time
+from pathlib import Path
+
+sys.path.insert(0, str(Path(__file__).resolve().parents[1]))
+
+from memorymaster.core.service import MemoryService  # noqa: E402
+from memorymaster.knowledge import rule_miner  # noqa: E402
+
+LINAJE = dict(
+    scope="project:bench",
+    provider="google",
+    source_ref="transcript:bench",
+    evidence_hash="e" * 64,
+    session_kind="human",
+)
+
+
+def main(argv: list[str]) -> int:
+    ap = argparse.ArgumentParser()
+    ap.add_argument("--calls", type=int, default=200)
+    args = ap.parse_args(argv[1:])
+
+    tmp = Path(tempfile.mkdtemp())
+    svc = MemoryService(tmp / "bench.db", workspace_root=tmp)
+    svc.init_db()
+    rule = {"trigger": "cuando falle el deploy", "action": "revisar el log"}
+
+    # calentar: la primera llamada paga imports y creacion de tabla
+    rule_miner._transcript_confidence(svc, rule, root_session_id="warm", **LINAJE)
+
+    inicio = time.perf_counter()
+    for n in range(args.calls):
+        # raiz distinta por llamada: el caso real de un lote de correcciones,
+        # y evita que el dedup por (provider, root) haga el trabajo trivial
+        rule_miner._transcript_confidence(
+            svc, rule, root_session_id=f"root-{n}", **LINAJE
+        )
+    total = time.perf_counter() - inicio
+
+    print(f"llamadas      : {args.calls}")
+    print(f"total         : {total:.3f}s")
+    print(f"por llamada   : {total / args.calls * 1000:.3f} ms")
+    print(f"llamadas/seg  : {args.calls / total:.1f}")
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main(sys.argv))

docs/adr/0001-local-first-single-database.md

@@ -1,4 +1,10 @@
+<!-- doc-head: authoritative MemoryMaster data remains one local-first database -->
 # 0001 Local-First Single Database
+# Covers: the authority boundary for claims, evidence, lifecycle, and recall data.
+# Key terms: SQLite WAL, authoritative store, Postgres parity, Qdrant.
+# Read when: changing governed persistence or interpreting ADR-0016's analytics sidecar.
+# Decision: all correctness-critical MemoryMaster tables share one logical store.
+<!-- /doc-head -->
 
 Date: 2026-05-05; reaffirmed 2026-07-14
 
@@ -40,3 +46,9 @@ disabled and documented, but do not block an accurately scoped local release.
 Consumers must not create a split layout such as one database for claims and another for Atlas Inbox. A split layout creates disconnected state and silently breaks the claim/source relationship.
 
 Schema and concurrency changes must preserve WAL behavior and the all-tables-in-one-store assumption.
+
+ADR-0016 does not weaken this rule. It permits one disposable operational
+sidecar for rebuildable coding-agent trajectory analytics. That sidecar cannot
+participate in claims, recall, lifecycle, or cross-database correctness paths;
+the minimal hashed recurrence lineage that affects skill governance remains in
+the authoritative database.