ADR-074 — Self-Learning Wiring + Honest Multi-Path Output (#2245)

Status: Accepted — Implemented in ruflo 3.10.14 Date: 2026-05-30 Tracking issue: #2245 Companion: ADR-143 (deterministic Tier-1 codemods — same claims-vs-reality pattern), ADR-073 (agentdb SOTA — also a claims-vs-reality round), ruvnet/ruflo#2241 Dream-Cycle research (Structured Distillation is the natural Round-2 enhancement to this work)

Context

The reporter on #2245 found that the self-learning subsystem reports success but persists nothing queryable. Specifically:

1. signalsProcessed is initialized in three places, read once, incremented zero times anywhere in the codebase. Pure dead metric posing as a health indicator.
2. hooks_task-completed {trainPatterns: true} is a stub: handler returns success: true, patternsLearned: 0 with hardcoded zeros, ignores trainPatterns entirely, and the inline note says the work is "delegated to the intelligence pipeline (#1916 follow-up)."
3. hooks_pretrain does real file scanning but bundles results into one memory-bridge row in the pretrain namespace and never writes per-pattern rows to the neural store. Result: neural_patterns list returns [] immediately after pretrain "succeeds" with "47 patterns extracted."
4. Persistence reset bug: loadPersistedStats() only restored trajectoriesRecorded — patternsLearned and signalsProcessed were reset to 0 on every process restart, masking real learning history.
5. Four disjoint stat aggregators (globalStats / memory_bridge / hooks metrics / neural_patterns) never agree because none of them share a source of truth.

The reporter correctly identified that the system has one path that actually works: hooks_intelligence_trajectory-start → -step → -end. Everything else is either a stub, a write to a different store than the dashboard reads from, or a permanently-zero counter.

This is the same family as ADR-143 (Agent Booster Tier-1 was dead/mislabeled) and ADR-073 §A (recall@10 was un-measured). A real engine exists; convenience surfaces advertise capabilities they never actually invoke.

Decision

Ship ruflo 3.10.14 with the three minimal wirings plus honest multi-path output. The full unification (one source of truth for all 4 aggregators) is a multi-PR program tracked separately — not crammed into this round.

§1 — Wire hooks_task-completed to the trajectory pipeline

When trainPatterns: true, synthesize a one-step trajectory from { taskId, success, quality, content } and call the existing recordTrajectory() helper in intelligence.ts — the same code path hooks_intelligence_trajectory-end invokes. The handler now:

• Returns learningPath: 'trajectory-pipeline' | 'recorded-only' so callers can tell which path ran.
• Returns honest patternsLearned and trajectoriesRecorded counts (deltas measured from getIntelligenceStats()).
• Falls back to "recorded-only" with an honest learningError field if the pipeline fails (e.g. SONA unavailable).
• Surfaces a clear note either "Trained via SONA + EWC++ trajectory pipeline" or "Completion recorded only. Pass trainPatterns:true to feed the learning loop."

§2 — Wire signalsProcessed

Add recordSignalProcessed() exported from intelligence.ts. Increment + throttled persist (every 16th signal). Call it from bridgeStoreEntry() so every memory-bridge write counts as a real signal. The dead-zero metric is now an actual health indicator.

Also fix loadPersistedStats() to restore patternsLearned and signalsProcessed alongside trajectoriesRecorded — so a process restart no longer zeroes the learning history.

§3 — hooks_pretrain writes per-pattern rows + honest output

Add storeNeuralPatterns(items) exported from neural-tools.ts. Pretrain now writes:

• One bundle row to memory-bridge pretrain namespace (existing behavior, back-compat).
• N per-pattern rows to the neural store via storeNeuralPatterns — so neural_patterns list actually returns them.

The response surfaces both patternsBundled (bundle row) and patternsIndexed (per-pattern rows) separately, plus a sources.stores field naming exactly which stores got written. Callers can no longer be misled by a single "47 patterns extracted" number.

§4 — Honest multi-path messaging

Per the goal-condition "soften where possible to indicate multiple paths":

• Every learning-adjacent handler that returns success: true also returns a field naming the path it took and the store(s) it wrote to.
• The hooks_task-completed description explicitly lists the three paths: (a) trainPatterns:true for one-step learning, (b) hooks_intelligence_trajectory-* for multi-step, (c) memory_store for storage without learning.
• The note strings tell callers what didn't happen, not just what did. ("recorded only — pass trainPatterns:true to feed the learning loop", not "[OK] Outcome recorded.")

§5 — Adversarial input hardening (OWASP ASI06)

Per #2241's OWASP Top-10-for-Agentic-Apps scan: task-completed's content is user-supplied and now feeds the SONA learning model. Add basic sanitization:

• Strip ASCII control chars (except newline + tab).
• Cap to 4 KB (way over a real trajectory step, well under a memory bomb).

This is defense-in-depth against context-poisoning attempts that try to inject control sequences into the learned-pattern store.

Deliberately NOT in this round

Tracked in #2245 / #2241 for future PRs:

• Unify the 4 stat sources — designate globalStats as authoritative, derive memory_bridge_status, hooks metrics, and neural_patterns count from it. Multi-PR design work.
• Wire post-edit / post-command to feed the trajectory pipeline — design call (which store wins).
• Structured distillation of trajectory content (#2241 §Structured Distillation) — 11× compression + +1.4 MRR. Natural Round-2 enhancement.
• Schedule consolidation worker — background NightlyLearner instead of on-demand only.
• Wire MCP trajectory-end to globalStats too — currently feeds sonaCoordinator only.

Consequences

• hooks_task-completed {trainPatterns:true} now does what the schema advertised. Agents that use this hook for completion-driven learning actually train the model.
• signalsProcessed is now a real health indicator. Long-running daemons can use it to verify the memory bridge is actually receiving traffic.
• neural_patterns list reflects what pretrain claims — no more "47 extracted, 0 listed" gap.
• Honest output (learningPath, patternsLearned, learningError, sources.stores) means downstream agents can make routing decisions on the real outcome instead of trusting success: true.
• The "multiple paths" messaging is explicit: callers are told upfront that task-completed is one path, trajectory-* is another, and memory_store is a third. No more single "[OK] Outcome recorded" theater.

Verification

• __tests__/self-learning-2245.test.ts — 9 tests across EASY (primitives), MEDIUM (MCP surfaces), and COMPLEX (multi-step + persistence) categories. CI gate breaks if any of the three fixes regresses.
• scripts/benchmark-self-learning.mjs — proof harness with 5 sections (A–E), writes a committed run JSON. Reproduces with N=20 node scripts/benchmark-self-learning.mjs from a fresh checkout.
• Build clean (tsc -b); full CLI test suite still green.
• Manual end-to-end:
  • recordSignalProcessed × 10 → signalsProcessed = 10 ✅
  • hooks_task-completed {trainPatterns:true} × 10 mixed-verdict → 10 trained via trajectory pipeline, ~18 ms/call avg ✅
  • hooks_task-completed (no trainPatterns) × 10 → trajectories+0 (correct negative control) ✅
  • storeNeuralPatterns × 10 → neural_patterns list returns ≥10 ✅