ADR-064: Stub Remediation & Full Implementation — v3.5.22

Date: 2026-03-17 Status: Implemented (22 stubs fully implemented in v3.5.43, PR #1438) Context: Deep capability audit (ADR-063) identified stub implementations and broken features. This ADR documents the remediation plan and tracks which items were already functional vs genuinely needing fixes.

Audit Correction

The ADR-063 audit flagged several items as "stub" or "missing" that are in fact fully implemented. The audit ran in Docker where some capabilities were degraded due to pruned dependencies, leading to false negatives.

Already Implemented (No Changes Needed)

Item	Finding	Evidence
Session restore --latest	Fully implemented with MCP integration	`session.ts:274-432` — interactive selector when no ID given
Workflow persistence	Saves to `.claude-flow/workflows/store.json`	`workflow-tools.ts:52-78` — `loadWorkflowStore()`/`saveWorkflowStore()`
Daemon scheduler	Working process management with PID files	`daemon.ts:234-458` — detached process, signal handling
Flash Attention JS	CPU-optimized block-wise implementation	`flash-attention.ts:67-150` — top-K sparse, fused softmax-matmul
MCP path resolution	Properly sanitized with `process.cwd()`	`session-tools.ts:34-39` — path sanitization regex
Hook handler stdin	Correct timeout + pause cleanup	`hook-handler.cjs:40-58` — graceful degradation

Genuinely Needs Implementation (7 Items)

Priority	Item	Current State	Target
P0	@ruvector/learning-wasm integration	Not wired in CLI	Wire `@ruvector/[email protected]` into neural commands
P0	Consensus vote counting	Proposals queued but no quorum validation	BFT 2/3 majority, Raft leader election, configurable quorum
P1	Memory delete HNSW cleanup	SQL row deleted, vector remains	Delete embedding row + rebuild search index
P1	SONA reinforcement learning	Signal recording only	Trajectory → reward → weight update loop
P1	EWC++ consolidation execution	Algorithm defined, not executed	Fisher matrix computation + penalty application
P2	Coverage hooks test integration	Returns 0% always	Read lcov.info / jest coverage-summary.json from disk
P2	Intelligence stats → neural bridge	Stats counter disconnected	Read from ReasoningBank `.claude-flow/neural/patterns.json`

Implementation Plan

P0-1: @ruvector/learning-wasm Integration

Wire the published @ruvector/[email protected] package into the CLI:

neural train → ruvector.initTraining() with WASM backend
neural predict → ruvector.predict() with Flash Attention
neural optimize → ruvector.quantize() for Int8

Files: src/commands/neural.ts, src/ruvector/index.ts Test: neural train --wasm should use WASM backend, report >2x speedup

P0-2: Consensus Protocol Implementation

Replace stub vote recording with real consensus algorithms:

BFT (Byzantine Fault Tolerance):

Require 2f+1 votes where f = floor((n-1)/3)
Validate vote signatures (hash-based)
Detect conflicting votes from same agent

Raft (Leader Election):

Term-based leader election
Log replication to followers
Heartbeat timeout triggers new election

Quorum:

Configurable quorum threshold (default: majority)
Timeout-based auto-resolution

File: src/mcp-tools/hive-mind-tools.ts Test: Create proposal → submit votes → verify quorum reached/rejected

P1-1: Memory Delete HNSW Cleanup

When deleting a memory entry:

Delete SQL row (existing)
Delete embedding from vector store
Mark HNSW index as dirty for rebuild on next search

File: src/memory/memory-initializer.ts Test: Store entry → search finds it → delete → search no longer finds it

P1-2: SONA Reinforcement Learning Loop

Implement the 4-step intelligence pipeline for real:

RETRIEVE: Find similar patterns via ReasoningBank (existing)
JUDGE: Score trajectory outcome (success/failure/partial)
DISTILL: Compute LoRA weight delta from successful trajectories
CONSOLIDATE: Apply EWC++ penalty to prevent catastrophic forgetting

Implementation:

recordTrajectoryStep() → accumulate steps with outcomes
endTrajectory(verdict) → compute reward signal
distillLearning() → update pattern confidence scores based on outcomes
consolidate() → apply EWC++ Fisher penalty

Files: src/memory/intelligence.ts, src/memory/ewc-consolidation.ts Test: Train patterns → use them → record outcome → confidence scores change

P1-3: EWC++ Consolidation Execution

The Fisher Information Matrix and consolidation formula are defined but not executed.

Implementation:

After each successful trajectory, compute diagonal Fisher approximation
Store Fisher values per pattern dimension
On consolidation, apply penalty: L_total = L_new + (lambda/2) * sum(F_i * (theta_i - theta_old_i)^2)
Persist Fisher matrices to .swarm/ewc-fisher.json

File: src/memory/ewc-consolidation.ts Test: Train → consolidate → retrain different domain → verify original patterns preserved

P2-1: Coverage Hooks Test Integration

Read real coverage data from common test tools:

Supported formats:

coverage/coverage-summary.json (Jest/Istanbul)
coverage/lcov.info (lcov)
.nyc_output/ (NYC)

Implementation:

coverage-gaps reads coverage files, computes per-file gaps
coverage-route uses coverage data to prioritize test-related tasks
Falls back to "no coverage data found" message (not 0%)

File: src/commands/hooks.ts (coverage-gaps, coverage-route, coverage-suggest)

P2-2: Intelligence Stats → Neural Bridge

Connect hooks intelligence stats to ReasoningBank data:

Read pattern count from .claude-flow/neural/patterns.json
Read trajectory count from .claude-flow/neural/stats.json
Read SONA adaptation metrics from intelligence system

File: src/commands/hooks.ts (intelligence stats handler)

Non-Functional Bug Fixes (8 Items)

Issue	Fix	File
#1335 Daemon scheduler	Verify interval comparison uses consistent timestamps	`daemon.ts`
#1122 Memory delete HNSW	See P1-1 above	`memory-initializer.ts`
#1117 Orphan processes	Kill child process on timeout rejection	`src/utils/`
#1116 daemon.log 0 bytes	Use `import('fs')` instead of `require('fs')` in ESM	`daemon.ts`
#1113 Scheduling intervals	Fix interval multiplier calculation	`daemon.ts`
#1333 MCP from /	Already fixed — path sanitization is correct	N/A
#1342 Post-bash junk files	Quote/escape tool output in hook args	`hook-handler.cjs`
#1331 stdin hang	Already fixed — timeout + pause cleanup works	N/A

Dependencies

Package	Version	Purpose
`@ruvector/learning-wasm`	0.1.29	MicroLoRA, Flash Attention WASM, Int8 quantization
`@claude-flow/memory`	3.0.0-alpha.12	AgentDB + ControllerRegistry

Success Criteria

neural train --wasm uses WASM backend and reports >2x speedup
hive-mind consensus validates votes with real quorum logic
memory delete removes both SQL row and embedding vector
neural train → neural predict → hooks post-task → confidence scores update
coverage-gaps reads real coverage files (or reports "no data" instead of 0%)
hooks intelligence stats shows real pattern/trajectory counts
All 8 bug fixes verified
Build clean, all 301+ command endpoints functional

Consequences

Positive

Eliminates all known stub implementations
WASM integration unlocks real Flash Attention and Int8 quantization performance
Consensus protocols become functional for multi-agent coordination
RL loop makes the learning system genuinely adaptive

Negative

@ruvector/learning-wasm adds ~2MB to package size
Consensus overhead adds latency to hive-mind operations
Real RL requires more computation than stub signal recording

Risks

WASM module may not load in all environments (fallback to JS required)
Consensus with slow/unreachable agents needs timeout handling
EWC++ Fisher matrices grow with pattern count — need size limits