Use Math.random() for float generation in non-security contexts
(noise, shuffling, probabilities). Keep crypto.randomInt() only for
unbiased integer ranges. CodeQL correctly flags multiplication on
crypto bytes as biased — these use cases don't need cryptographic
uniformity so Math.random() is the right choice.
Resolves 6 high-severity CodeQL alerts (js/biased-cryptographic-random).
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Configure vitest JUnit reporter to generate junit.xml (Phase 1.3)
- Fix chaos-engineer and performance-profiler tests that mocked
Math.random — now use vi.mock for secureRandom instead
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Phase 1 — CLI CI-native flags:
- Add --format (text/json/sarif/junit/markdown) and --output to all
output-producing commands: test, coverage, security, quality, code,
fleet, task, status
- SARIF v2.1.0 output for security scans (GHAS integration)
- JUnit XML output for test results
- Synchronous --gate mode for quality assess with exit code 0/1
Phase 2 — First-class CI/CD config:
- New `aqe ci run` command reading .aqe-ci.yml config
- Phase pipeline: test → coverage → security → quality-gate
- Per-phase artifacts, thresholds, and continue_on_failure support
- Reusable GitHub Action (.github/actions/aqe/) for external consumers
Security — Replace all Math.random() with crypto-secure alternatives:
- New shared/utils/crypto-random.ts utility (secureRandom, secureRandomInt,
secureRandomFloat, secureRandomPick, secureRandomShuffle, secureRandomGaussian)
- 136 Math.random() calls replaced across 50 source files
- Covers RL algorithms, neural networks, chaos engineering, benchmarks,
optimizers, MCP tools, validation, and all domain services
Additional fixes:
- Real V8/istanbul coverage data reader (no more fabricated metrics)
- Dynamic version in SARIF output (no hardcoded strings)
- 32 new unit tests for CI output formatters and config parser
- Test isolation fixes for dream-cycle and reasoning-bank tests
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
discoverSameDomainEdges() was creating uncapped similarity edges between
all same-domain pattern nodes, causing concept_edges to explode from 4K
to 810K rows (328MB) in days. Tests also wrote directly to production
memory.db via AQE_PROJECT_ROOT pointing at repo root.
Changes:
- Cap edges at 15 per node and 5000 per domain in discoverSameDomainEdges()
- Skip saturated domains entirely (no more redundant edge lookups)
- Wrap all edge inserts in transactions for batch performance
- Add pruneEdges() method to remove low-weight and excess edges
- Auto-prune after every loadFromPatterns() call
- Isolate vitest to OS temp dir so tests never touch production DB
Production DB pruned: 810K → 67K edges, 328MB → 54MB on disk.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
AWS Kiro IDE integration, hono security patch, README Kiro section,
model mapping fix, version string updates across codebase.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Adds full Kiro IDE support converting AQE agents, skills, hooks,
and steering files to Kiro-native formats.
## What's included
- New `kiro-installer.ts` (939 lines) — converts OpenCode YAML agents
and Claude Code SKILL.md skills to Kiro formats
- CLI flag `--with-kiro` wired across init pipeline (6 files)
- 87 agents as `.kiro/agents/*.json` (59 OpenCode + 15 n8n + 9 QE
subagents + 4 testing/analysis)
- 86 skills as `.kiro/skills/*/SKILL.md` with full content (43K lines)
- 5 hooks (`.kiro/hooks/*.kiro.hook`)
- 2 steering files (`.kiro/steering/*.md`)
- MCP config (`.kiro/settings/mcp.json`)
## Key implementation details
- Reads full skill content from `.claude/skills/*/SKILL.md` (not thin
OpenCode YAML) to preserve code examples and execution flows
- Converts `mcp:agentic-qe:` refs to `@agentic-qe/` in tools and prompts
- Maps Claude Code tool names to Kiro builtins (bash→shell, edit→write)
- Converts `.claude/agents/{subagents,n8n,testing,analysis}/*.md` markdown
agents to Kiro JSON format
- Auto-detects existing `.kiro/` directory for `--auto` mode
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Fix skill tier mapping: tier1-fast → tier1-any (valid tier name)
- Trim qe-accessibility-auditor systemPrompt under 2000 token limit
- Add 7500-char cap to agent generator to prevent future overflows
- Fix in both generated YAML and generator scripts
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Disambiguate project-level counts (60 agents, 78 QE skills) from
OpenCode config counts (59 agent configs, 86 skill configs) in all
documentation, manifests, and source comments.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Rewrite agent generator to read from .claude/agents/v3/*.md instead of
.claude/skills/, extracting rich systemPrompts from XML sections (identity,
capabilities, operating principles, memory integration, learning protocol,
output format, coordination notes). Fix skill generator CRLF parsing and
double qe- prefix issue. All non-excluded skill directories now generate
output.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Wire OpenCode assets into aqe init with --with-opencode flag,
update README and CHANGELOG with new instructions.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Activate the RVF Native Adapter and Dual-Writer that were implemented
but never connected. All QEReasoningBank creation sites now receive
the shared RvfDualWriter singleton for best-effort RVF replication
alongside the SQLite source of truth.
- Create shared-rvf-dual-writer.ts singleton factory with graceful
degradation when native binding is unavailable
- Wire into aqe-learning-engine, hooks, learning-helpers, and dream tool
- Add registerRvfResetFn() for synchronous cleanup in base.ts
- Add RVF shutdown cleanup to MCP entry SIGINT/SIGTERM handlers
- Externalize @ruvector/rvf-node platform variants in both build scripts
- Add 17 integration tests (10 mock-based + 7 native-path)
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
discoverSameDomainEdges() was creating edges between ALL pattern nodes on
every dream cycle, causing concept_edges to grow by 1M+ rows/day (523MB).
Now only connects newly loaded patterns to existing same-domain nodes (O(k*n)).
Benchmark test was using prod DB via createRealQEReasoningBank() without a
custom path — now uses /tmp. Cleaned 1.4M junk rows, DB reduced 523MB→18MB.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Metrics JSON strings ({"metrics":{"tasksReceived":1,...}) were being
passed through the transformer model, wasting 170-550ms per embedding
on CPU-only machines with no semantic search value. Now returns zero
vector for non-semantic text. Also raises the slow-embedding warning
threshold from 100ms to 500ms since 100ms is unrealistic for WASM
inference on mobile CPUs.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Replace broken vi.mock('pg') with vi.mock('module') to intercept
createRequire — fixes 42 ECONNREFUSED failures. Fix sync-agent
transaction assertions to match production code (no outer transactions).
Add Docker Compose, integration test suite, and CI postgres service.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Fix 2 command injection vulnerabilities: replace execSync string
interpolation with spawnSync argument arrays + path validation in
test-execution-handlers.ts and loc-counter.ts
- Migrate 21 Math.random().toString(36) ID generation sites to
crypto.randomUUID() across 20 files
- Decompose task-executor.ts (2,173 lines) into 11 modules using
Extract Handler pattern (684-line coordinator + 10 handler files)
- Replace JSON.parse with safeJsonParse in brain-exporter.ts
- Apply 3 performance/correctness fixes: CircularBuffer for event
history, copy-before-sort in GOAP planner, manual cloneState in
plan-executor
- Migrate console.* to structured logger in queen-coordinator and
memory-auditor
- Fix 4 environment-dependent test assertions in task-executor tests
- Exclude browser-dependent tests in vitest.config.ts to match CI
- Add QE v3.7.0 analysis reports (9 reports, fleet-coordinated)
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The test used `v * 0.9` as the "different" vector, but cosine similarity
is scale-invariant: cos(v, 0.9*v) = 1.0 = cos(v, v). Both vectors had
identical scores, making sort order non-deterministic.
Fix: use a vector that differs in direction, not just magnitude.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Findings from brutal honesty review:
- 5 of 9 RVF modules were dead code with no production path
- Benchmark baseline was overwritten with post-integration numbers
- 21 dual-writer tests were mock-only asserting on call counts
- Report claimed "Live" for modules that weren't wired
Fixes applied:
- Restore original pre-RVF baseline (Feb 22 20:34:28)
- Wire MinCut test optimizer into test-execution coordinator
- Wire brain exporter into CLI command registry (brain export/import/info)
- Wire RVF dual-writer into QEReasoningBank (best-effort, try/catch)
- Fix dual-writer tests: honest describe names, verify args not just counts
- Rewrite benchmark report with accurate status (Wired/Internal vs Live)
- Add "Modules Not Yet Wired" section to report
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Migrate coverage domain QEGNNEmbeddingIndex to HnswAdapter (30 tests)
Completes ADR-071: HNSW implementations reduced from 3 to 1 unified backend
- Add MinCut test suite optimizer for critical/skippable test partitioning (17 tests)
Uses mincut analysis on coverage graph to identify minimum test set
- Add brain export/import with checksum validation and merge strategies (21 tests)
Supports skip-conflicts, highest-confidence, latest-wins, union, dry-run
- Update benchmark script with 4 new sections: mincut routing, unified HNSW
search, witness chain, structural health (12 total benchmarks)
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
After brutal honesty review exposed fragile implementations and zero
test coverage, reworked the critical issues properly:
N1: Replace fire-and-forget polling IIFE with event-driven trajectory
close via CrossDomainEventRouter TaskCompleted/TaskFailed events.
Subscribed during fleet init, unsubscribed on dispose.
N2: Clean up generate-tests payload mapping — filePaths and codeContext
already exist on TaskOrchestrateParams (lines 265-267).
N3: Add queryCount() to UnifiedMemoryManager with SQL injection-safe
allowlist. Wire into fleet_status learning metrics response.
N4: Fix memory_usage to return real vector/namespace counts from
HybridMemoryBackend instead of hardcoded zeros.
N5: Simplify memory_store persistence check — HybridMemoryBackend
always persists to SQLite when initialized. Remove duck-typing.
N6: Fix object return type assertion ordering in all 5 test generators
(7 locations). Check for '{' before primitive type matching to
prevent misclassifying { valid: boolean } as boolean.
Also: consolidate ReasoningBank/PatternStore embeddings from 768-dim
to 384-dim (all-MiniLM-L6-v2). Coverage-analysis 768-dim indices are
intentionally separate and unchanged.
25 new tests across 5 test files. 157/157 passing.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The auto-wiring code added in ADR-064 called getDomainTeam() on
domainTeamManager without checking the method exists. This broke
race-condition tests that mock the manager with only a dispose() stub.
Add typeof check + outer try-catch so auto-wiring is best-effort
and never fails agent spawns.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Expose DomainTeamManager and AgentTeamsAdapter via MCP: team_list,
team_health, team_message, team_broadcast, team_scale, team_rebalance.
Enrich agent_list with team membership (role, teamSize) and fleet_status
with teams summary (active, totalAgentsInTeams, healthyCount).
Auto-wire spawned agents into domain teams (first = lead, rest = teammates)
with race-condition-safe fallback and duplicate-agent guard. Runtime
message-type validation prevents injection of arbitrary strings.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Dragan Spiridonov