[Dream Cycle 2026-09-02] intelligence: LearningBridge.consolidate() reward-blind fix + capabilities,memory scan

[ruvnet]

1. Tonight's Rotation

DATE=2026-09-02
DEEP=intelligence
SCAN=capabilities,memory
SLOT=2 (DAYINT % 5)
COMMIT=4d0134e59b4fa5e8552cb7b98c6b9846f08b0c82

No bonus deep dive (DAYINT%25=2, %75=27 — neither trigger).

2. Ledger Check

Last 14 rows inspected. Prior gist (2026-09-01, security/turnWindowMs) scored 10/10 on the STEP 1.2 rubric: 8 competitor rows, a primary-source MCP-spec read, valid witness, executable recommendations, under 1500 words.

Two confirmed, distinct findings from the ledger check itself:

• 2026-08-24 → 2026-09-01 (9 nights): ledger-append bug. All 9 nights ran the full pipeline (branch, issue, evaluated draft PR each — verified via git ls-remote and a GitHub PR listing, not inferred) but docs/dream-cycle/LEDGER.md was never appended. Same failure mode as the 08-14..08-19 gap recovered on 08-19. Backfilled tonight; root cause of the append step itself remains undiagnosed — flagged as a standing candidate for a future automation/meta scan.
• 2026-08-20 → 2026-08-23 (4 nights): real pipeline gap, not a ledger bug — no dream/* branches, issues, or PRs exist for these dates at all.

No intelligence-surface finding repeated 3+ times in the trailing window, so no reject-duplicate trigger.

3. Deep Dive Findings (intelligence)

5 parallel research roles (Deep Researcher, Scan A/capabilities, Scan B/memory, Competitor Analyst, Ruflo Architecture Reviewer) converged on: LearningBridge.consolidate() (v3/@claude-flow/memory/src/learning-bridge.ts) hardcoded completeTask(trajectoryId, 1.0) for every completed trajectory, regardless of the underlying insight's actual quality — a reward-blind learning path. Full write-up: docs/dream-cycle/dream-gist-2026-09-02.md.

Separately (not tonight's candidate, a governance finding): PR #3110 (2026-08-27, EWC-gate wiring fix, already evaluated ACCEPT-scoped) was never merged to main — that bug is still live in production. Recommend prioritizing its merge.

4. Hypothesis

Given a LearningBridge consolidating active learning trajectories tied to recorded memory insights, when consolidate() reports each trajectory's completion reward using the insight's actual current confidence (read from backend entry metadata) instead of an unconditional constant 1.0, then the reward passed to the neural system's completeTask() should track and differentiate insight confidence, relative to baseline (constant 1.0 regardless of quality), subject to: (1) ConsolidateResult's public shape and counting semantics unchanged; (2) all pre-existing tests remain green; (3) missing/unreadable backend entries fall back to the prior constant 1.0; (4) fully deterministic, $0 test coverage.

5. Evaluation Receipt

Real evaluator: vitest run on v3/@claude-flow/memory/src/learning-bridge.test.ts, deterministic, $0.

• Baseline (pre-diff, stashed): 56/56 passing.
• Candidate (final, post-critique-fix): 62/62 passing (56 unchanged + 6 new).
• Full package regression: 464/465; the 1 failure is a pre-existing, environmental read-only-file-permission test (root bypasses chmod in this sandbox) — reproduced identically on baseline.

6. Darwin Results

Not run, deliberately — this is a binary correctness fix with no tunable parameter, same reasoning as PR #3152's precedent.

7. Flywheel Evidence

No .claude-flow/flywheel/ state in this repo. Evidence retained as: the diff, this issue, the PR, and docs/dream-cycle/dream-gist-2026-09-02.md.

8. Reward Hack Check

Manual checklist (no generic diff/benchmark-scanning CLI reachable this session). No test weakened — diff only adds coverage (confirmed by the clean 56→62 baseline/candidate comparison). No gold data touched. No seed/metric manipulation. Minor added cost: one backend.getByKey() call per active trajectory in consolidate(), gated by consolidationThreshold (default 10) — not hidden, noted here.

9. Security Review

Not security-sensitive: internal learning-signal correctness fix, no new attack surface, no credential/network/filesystem exposure change.

Independent adversarial-critic subagent (separate context) — two rounds:

• Round 1: found a real bug in the fix itself. entryId flowing through LearningBridge is the caller-assigned key string (from AutoMemoryBridge.storeInsightInAgentDB), not the backend's internal entry.id (an independently-generated UUID). The original fix called backend.get(entryId), which is indexed by id, so it would almost always miss in production and silently fall back to the same 1.0 the fix claimed to eliminate — passing unit tests only because the mock backend's get() didn't validate id/key correspondence.
• Fixed: switched to backend.getByKey(namespace, key) (traced and confirmed correct across all three real backends — AgentDB, SQLite, hybrid), added a configurable insightNamespace (default 'learnings'), and rewrote the mock backend's getByKey to do a genuine namespace+key lookup so the tests exercise the real id≠key distinction via backend.store() + onInsightRecorded(insight, entry.key), not a permissive stub.
• Round 2 verdict: CONFIRMED, independently re-traced against agentdb-backend.ts, sqlite-backend.ts, and hybrid-backend.ts. One non-blocking nit: controller-registry.ts instantiates a second LearningBridge that doesn't thread insightNamespace through, but that instance is never fed any insights (dead/unwired in that path) — pre-existing, out of scope tonight.

10. Scan Findings: capabilities

Ruflo's plugin system has a fully-typed PluginPermission capability manifest (network/filesystem/execute/etc.) plus install-policy shape (allowedPermissions, minTrustLevel) — but grepping the actual loader (plugins/manager.ts, plugins/store/index.ts) finds zero consumers of any of it; manager.ts:341 explicitly logs "Plugins run with full process access." The trust-anchor Ed25519 key is still a literal PLACEHOLDER. This is ahead of bare MCP (which has no native per-tool authorization at all per current OWASP/MCP-gateway guidance) in shape, but the enforcement layer is missing — worth a HIGH-severity flag in a future metaharness mcp-scan, distinct from the already-tracked HIGH-04 unsandboxed-execution finding.

11. Scan Findings: memory

Ruflo's memory layer is more advanced than expected going in: RRF fusion, recency-decay, and MMR diversity re-ranking already exist in smart-retrieval.ts. The concrete gap: filtering (hnsw-index.ts:388, agentdb-backend.ts) is post-filter/over-fetch, not graph-aware — ACORN (SIGMOD 2024, Grade A) and its 2025-26 follow-ons (RACORN-1, Compass) show graph-native filtered HNSW can be 2-1000x faster at fixed recall under selective filters. Weaviate shipped ACORN-style filtering (v1.27+, default-on v1.34); Qdrant ships a competing filterable-HNSW approach. Not proposed as tonight's candidate — flagged for a future night.

12. Competitors Reviewed

LangGraph, AutoGen/MS Agent Framework, CrewAI, OpenAI Agents SDK/AgentKit, Qdrant, Weaviate, Milvus, LanceDB, Vespa, MCP (protocol). Full table + "why the evolutionary-loop column is uniformly empty" analysis in the gist — headline: CrewAI explicitly declined this design (issue #3015, closed not-planned) and OpenAI is actively retreating from it (Agent Builder/Evals wind-down announced 2026-06-03), both citing/evidencing safety concerns (DGM reward-hacking, documented "capability erosion under self-evolution") that Ruflo's Darwin/Flywheel governance shape is specifically built to address.

13. Gist

docs/dream-cycle/dream-gist-2026-09-02.md (committed to the repo — no gist-creation tool was reachable in this execution environment, only GitHub issue/PR/repo MCP tools; noting this honestly rather than fabricating a gist URL).

14. Witness

Field	Value
Session commit	4d0134e59b4fa5e8552cb7b98c6b9846f08b0c82
Report SHA-256 (pre-witness content)	29c9189c833c3d56c378f9a1329b3baf6f16543493ad28e4e62ec76b14c0fdb9
Witness stamp	bead6402cb018f0e57ec6695c82e007365b5f2ae91d530a74c82a2eb8a1d362b

15. Recommendation

evaluated: accepted

1. Merge this fix.
2. Merge PR dream(intelligence): #3109 wire computeConfidencePenalty/updateFisherFromConfidences into distillLearning's EWC gate (evaluated, ACCEPT-scoped, corrected post-review) #3110 — an already-evaluated, still-open EWC-gate fix with a live production defect.
3. Diagnose the ledger-append failure mode (9 recurrences now) as a future automation/meta finding.