All Agentic QE release notes organized by version.
| Version | Date | Highlights |
|---|---|---|
| v3.14.8 | 2026-10-04 | Braces CVE removed, reliable memory, learning, CI gates and caches. |
| v3.14.7 | 2026-10-02 | Opt-in browser setup, clean MCP shutdown, self-healing pattern search. |
| v3.14.6 | 2026-09-29 | Audit verify crash fix, audit repair, converging init, live GOAP state. |
| v3.14.5 | 2026-09-29 | Trustworthy exit codes, working platform setup, stable audit chains, quieter CLI. |
| v3.14.4 | 2026-09-27 | Current model generation, request-failure fixes, and honest task/workflow outcomes. |
| v3.14.3 | 2026-09-22 | Honest test, security, coherence, and quality-gate failures; Vitest 5 support. |
| v3.14.2 | 2026-09-13 | Reliable foreground daemon lifetime, conservative MCP safety metadata, and patched Hono/Sharp dependencies. |
| v3.14.1 | 2026-09-07 | Node 22+ support and stronger verification, recovery, and learning evidence. |
| v3.14.0 | 2026-08-30 | Qualified learning evidence, embedding provenance, fail-closed RVF recovery, and packaged QE Court. |
| v3.13.12 | 2026-08-21 | External LLM provider registration for downstream hosts. |
| v3.13.11 | 2026-08-14 | Reliable diagnostics, native ESM exports, and learned pattern visibility. |
| v3.13.10 | 2026-08-06 | Reliable learning consolidation, exports, and PageRank fallback. |
| v3.13.9 | 2026-08-04 | Broad Ruflo development-lock security remediation. |
| v3.13.8 | 2026-08-04 | Stable coverage gates and isolated learning-engine tests. |
| v3.13.7 | 2026-08-04 | Explicit semantic-index status, stable ANN top-K results, and critical dependency fixes. |
| v3.13.6 | 2026-08-03 | Reliable namespaced search, learning persistence, and embedding repair. |
| v3.13.5 | 2026-08-03 | Measured quality gates and trustworthy cross-framework test execution. |
| v3.13.4 | 2026-08-02 | Reliable Codex fleet setup, upgrades, hooks, and verification. |
| v3.13.3 | 2026-07-29 | Learning capture no longer stops silently under AQE_DISABLE_WAL (it now refuses to write in an unsafe journal mode and names what is holding the DB), and QE-Court's shipped panel no longer violates its own writerIsNeverJuror rule (#576). |
| v3.13.2 | 2026-07-24 | Stops hook-driven brain.rvf/patterns.rvf corruption loops by closing native stores after every hook and protecting same-process live locks (#574). |
| v3.13.1 | 2026-07-23 | Honesty release: coverage analysis no longer reports numbers it never measured (real cargo llvm-cov for Rust; uncollected metrics are null, not 0 or 100), test_generate_enhanced actually calls your LLM (#567), and a HIGH advisory no longer ships via an optional peer (#565). Adds on-disk per-agent LLM routing (#568) and Codex CLI QE skills. BREAKING: coverage metrics are now number | null. |
| v3.13.0 | 2026-07-18 | QE-Court: adversarial review where a change is prosecuted by independent cross-vendor AI reviewers and a SHIP verdict must survive an escalating deeper review — it reviewed this release and blocked it, catching 3 data-loss/crash bugs. Adds a Codex (GPT) provider, a significance gate for the flywheel, automatic memory.db recovery, per-agent cost tracking, and clears 4 HIGH CVEs from consumer installs (#565). |
| v3.12.3 | 2026-07-17 | Fixes vector search silently switching off for good: an interrupted RVF export left a store AQE could neither open nor replace, degrading to SQLite on every later run. Exports are now atomic; unusable stores quarantine and rebuild themselves — no manual cleanup. Flywheel receipts signed with the Cognitum platform identity. |
| v3.12.2 | 2026-07-13 | Billing-aware LLM execution: run QE on your Claude subscription (AQE_LLM_PROVIDER=claude-code), enforced cross-process spend caps (--max-budget-usd), Cognitum gateway provider, and billing transparency in aqe health. Fixes aqe eval run scoring fabricated mock responses — now real calls by default. ADR-123. |
| v3.12.1 | 2026-07-12 | Non-destructive aqe init: merges (never overwrites) .claude/settings.json, preserves custom statusLine / includeCoAuthoredBy / AQE_* overrides / other tools' hooks (ruflo, Claude Flow), and backs up the original before changes. |
| v3.12.0 | 2026-07-10 | Learning-integrity layer: promotions gated by evidence quality (test-exec > LLM-judge > structural), frozen hash-pinned oracle benchmark, Ed25519-signed replayable receipts, and frozen-rule re-execution before promotion. New two-gate aqe quality-gate CLI + qe/quality/gate MCP tool. Designed-experiment (DoE) tool with real ANOVA. ADRs 117–122. Additive. |
| v3.11.5 | 2026-07-07 | System-integrity sweep of the self-learning loop: dream-cycle insights now genuinely become reusable patterns (gate + fake-counter fixes); pattern usage stats fixed (double-write, id-desync lookup failures, mid-session regression); witness-chain signing/archival, agent-topology tracking, GOAP real execution, and SONA cold-start bugs fixed. |
| v3.11.4 | 2026-07-01 | Free-tier local test-gen defaults to code-tuned qwen3-coder:30b + pattern-cache skip; unified AQE_OLLAMA_URL + adaptive timeouts + capability-aware tool routing; opt-in proof-gate for tamper-evident memory.db integrity (ADR-116); self-running dependency-security system (daily audit + weekly in-range PRs, ADR-115); grpc-js HIGH CVE fixed, moderates 29→7. Additive. |
| v3.11.3 | 2026-06-27 | Oracle evals (run + mutate generated tests, ADR-113), durable-first test generation, mutation-score/regenerability gate, conservation guard for CLI/output/skill surfaces (ADR-114), cheaper-model lanes; fixed OpenRouter + Gemini providers. Additive. |
| v3.11.2 | 2026-06-27 | First-class C4 architecture diagrams from your codebase: aqe code c4 CLI + qe/code/c4 MCP tool generate Context/Container/Component Mermaid with a deterministic confidence gate, real Knowledge-Graph-derived relationships, and semantic diagram search. Additive. |
| v3.11.1 | 2026-06-25 | Benchmark-driven deepening of the cheap-first lane (opt-in): best-of-k + cross-model generation (~+6 quality pts), self-test Goodhart guard, @ruvector/adversarial-verify + verify gate, MCP default-deny self-governance, cost-Pareto value scoring, witnessed finding delivery, graceful optional-module loading. Default free-tier model → qwen3:30b-a3b. |
| v3.11.0 | 2026-06-20 | Free-tier local test generation (opt-in, AQE_FREE_TIER=1): cheap-first on local/OpenRouter/OpenAI-compatible models with a repair loop, auto-escalation to paid tiers, and self-learning routing feedback. PACT → PACTS framework (adds Structured principle, #545). Restores Dependabot-dropped @ruvector/ruvllm-* deps (#541). |
| v3.10.9 | 2026-06-17 | MCP reliability + db-free completeness (@nagoodman): agentic-qe mcp runs in-process so real clients get tools/list (#528); memory_delete no longer masked by stale session cache (#535); --no-database reaches all editor MCP configs (#533); new --no-claude exclusive install (#532). |
| v3.10.8 | 2026-06-16 | Database-free OpenCode install (--no-database/--memory memory, #530/#534) + MCP stdio client-compat fixes (#527 stdout purity + negotiated protocol, #529 loadable OpenCode assets) from @nagoodman; cleared protobufjs + ws production CVEs; vite → 8.0.16. |
| v3.10.7 | 2026-06-12 | Learning-integrity + supply chain: fix silent self-learning loss (hook swallowed native-binary errors; SONA reward feedback was inert, fixed via @ruvector/sona 0.1.7); shift-left consumer-tarball audit gate; refresh QE skill evals to current models. |
| v3.10.6 | 2026-06-11 | Pattern Space cross-pollination (#522): evidence-class labels on findings, pass/fail safety eval for data-protection rules, shipped-agent invariant CI, benchmark lineage + interaction benchmark, kept-nulls learning. ADR-105–110. |
| v3.10.5 | 2026-06-10 | Fable 5 parity (#520): CLI fleet/memory commands fixed, prompt caching wired, reasoning-scrub for clean learning, adversarial QCSD reviews, aqe arena tournaments. ADR-099–104. |
| v3.10.4 | 2026-06-08 | Keep project learning in its own .agentic-qe/ (#516: nearest-root + RVF resolution); stop orphaned aqe-mcp busy-loop / SIGTERM-ignore (#513); opt-in local Nagual hub + LLM judge. |
| v3.10.3 | 2026-06-03 | TS/JS code intelligence with no typescript dependency via bundled tree-sitter WASM grammars (#511); fixes MCP domain tools hanging on fresh install, code deps/search returning empty, CLI↔MCP graph namespace; repaired stale WASM grammars; lint cleanup. |
| v3.10.2 | 2026-06-02 | Self-learning hardening: fixes #508 (post-task hook recorded every trajectory success=0/agent=unknown) + #509 (stale dream readout); EWC++ forgetting protection, contradiction detection, MCP/learning verification harnesses, ADR-098. |
| v3.10.1 | 2026-05-22 | ADR-043 LLM router actually wired: kernel singleton threaded through 11 domains + MCP tools, persistent aqe llm config --set, Gemini GOOGLE_API_KEY alias, 2 fallback-chain bug fixes, 10-finding audit hardening. |
| v3.10.0 | 2026-05-20 | Optional external embedder endpoint (#503, ADR-097): OpenAI-compatible /v1/embeddings over HTTP or Unix socket, identity fingerprint + drift detection, hard-fail (no hash poisoning), TLS knobs. |
| v3.9.36 | 2026-05-19 | Close the self-learning routing Q-loop (#499): post-route now trains rl_q_values, post-task fires Q-update even when pre-task bridge is absent, Q-state collapsed from 4-dim to 3-dim for ~11× faster convergence. ADR-096. |
| v3.9.35 | 2026-05-19 | Restore ESM subpath import of HnswAdapter (#497) + dismantle cumulative-growth pattern in vitest fork workers (#448 steps 2+3): split heavy learning tests, release module singletons between tests |
| v3.9.34 | 2026-05-18 | Stop unbounded patterns.rvf/brain.rvf growth — fixes #495 sibling init runaway + adds post-dream compaction, boot-time size guard, bounded SQLite→RVF backfill |
| v3.9.33 | 2026-05-18 | Supply-chain cleanup: 0 consumer-side CVEs (was 4 high + 1 critical), ~480 fewer transitive packages, zero ERESOLVE warnings; new consumer-side audit gate prevents regressions |
| v3.9.32 | 2026-05-17 | Fix #491 chain — daemon/kernel wiring, namespace-aware reads, timestamp rehydration, loop-health try/finally. Adds daemon-runtime release gate (test:integration:fast + A23) |
| v3.9.31 | 2026-05-15 | Self-learning loop ships end-to-end: kernel-side dream cycles (ADR-094), three-signal routing with Q-blend + ε-greedy + mincut gate (ADR-095), aqe learning loop-health dashboard, retention pruner, daemon pidfile fix (#480, #486, #487, #488) |
| v3.9.30 | 2026-05-14 | Six MCP fixes: ESM __dirname polyfill, real ONNX embeddings, coherence_collapse NaN guard, default test phases, session cache wired, test_generate language validation (#467,#469,#470,#472,#473,#474) |
| v3.9.29 | 2026-05-14 | Hotfix: bridge publishes learning.ExperienceCaptured with the canonical nested payload shape so handler destructure resolves and recordExperience runs (#482 round 3) |
| v3.9.28 | 2026-05-14 | Hotfix: bridge eagerly loads domain plugins before starting so subscribers are wired when the first drain publishes (#482) |
| v3.9.27 | 2026-05-14 | Two architectural fixes: AbortSignal bounds session-start init (no more 43 GB patterns.rvf leak, #478), CapturedExperienceBridge wires hook activity to domain plugins (#479) |
| v3.9.26 | 2026-05-13 | Ten learning-pipeline fixes: post-edit stdin file_path, getStats DB fallback, pattern promotion, dream apply, Q-learning agent key, concurrent-dream guard, RVF orphan purge, tokens_saved, cli-hook consolidation, stale sentinel sweep (#453–#465) |
| v3.9.25 | 2026-05-12 | Fix: new post-route Stop hook closes route sentinels 1:1 — routing_outcomes no longer accumulates quality_score=-1 rows (#451) |
| v3.9.24 | 2026-05-12 | Fix: post-task no longer skips Stream B/D/F when --task-id is missing — rl_q_values populates on every invocation (#449) |
| v3.9.23 | 2026-05-11 | Two self-learning fixes: bootstrap patterns now reach HNSW (#445), qe_pattern_usage no longer permanently empty after ON CONFLICT upserts (#447) |
| v3.9.22 | 2026-05-09 | Non-destructive consolidation safety valve (no more silent Exp counter shrinkage) + CLI-only aqe init flow now persists hooked experiences |
| v3.9.21 | 2026-05-08 | Windows install unblocked (#439), RVF FsyncFailed root-cause fix, hypergraph untested / impacted queries return useful results |
| v3.9.20 | 2026-05-08 | Second wave wire-gap fixes: HNSW A in routing, hook-side embeddings inline, dream cycle busy_timeout, pattern quality_score moves on hook flow, TrajectoryBridge → unified memory |
| v3.9.19 | 2026-05-05 | Self-learning loop wire-gap fixes: workers tick, HNSW loads on boot, hook outputs persist, multi-dim quality |
| v3.9.18 | 2026-04-30 | Four MCP fixes (governance throttle, jest output, temp-path leak, coverage error msg) + agentic-qe-fleet plugin |
| v3.9.17 | 2026-04-27 | Fix: UserPromptSubmit hook now reads stdin event JSON — closes the routing learning loop for fresh aqe init |
| v3.9.16 | 2026-04-24 | Brain-export tooling: aqe brain diff/search, native-binding advisor aqe upgrade |
| v3.9.15 | 2026-04-22 | qe-browser Implemented (ADR-091): aqe eval run CLI, CI gate, Linux ARM64 browser hint |
| v3.9.14 | 2026-04-20 | Security + supply-chain: 15 critical vulns fixed, tarball -52%, command injection patched |
| v3.9.13 | 2026-04-17 | Opus 4.7 migration (ADR-093): Sonnet 4.6 default, xhigh fleet-wide, cyber agents pinned |
| v3.9.12 | 2026-04-16 | Fix: hook duplication after ruflo init, init hang from pretrain, faster CF detection |
| v3.9.11 | 2026-04-13 | Fix: aqe init --auto now correctly updates agents and helpers on version upgrade |
| v3.9.10 | 2026-04-13 | Multi-provider advisor routing: use any LLM for QE tasks, auto-failover, PII redaction |
| v3.9.9 | 2026-04-09 | qe-browser fleet skill: Vibium engine (10MB vs 300MB Playwright), typed assertions, visual diff, injection scan (ADR-091) |
| v3.9.8 | 2026-04-08 | Release-process maturity: codeload mirror, PR template CI enforcement, chaos workflow, VERIFICATION.md (#401 follow-ups) |
| v3.9.7 | 2026-04-07 | Release-gate corpus + aqe init --json + phase 06 stops lying about success (#401) |
| v3.9.6 | 2026-04-06 | Native HNSW works again — replaced @ruvector/router with hnswlib-node, no more vectors.db cruft |
| v3.9.5 | 2026-04-06 | Root-cause fix: disable native HNSW (deadlocks on certain inputs), faster JS backend by default |
| v3.9.4 | 2026-04-06 | Hotfix: governance phase chunk-split regression, --skip-code-index escape hatch, per-file logging |
| v3.9.3 | 2026-04-06 | Proper fix: aqe init watchdog, lazy bootstrap, remove daemon spawn, MCP entry regression |
| v3.9.2 | 2026-04-06 | Hotfix: aqe init --auto RVF lock deadlock, CLI exit hang |
| v3.9.1 | 2026-04-05 | RVF persistent vectors, agent COW branching, HNSW unification, StrongDM |
| v3.9.0 | 2026-04-02 | 11 CC-internals improvements, 4-tier compaction, plugin system, QE daemon |
| v3.8.14 | 2026-03-31 | SQL injection fix, remove faker from generators (~6 MB), 3 P0 blockers resolved |
| v3.8.13 | 2026-03-30 | Code intelligence CLI (complexity, --incremental, --git-since), security fix, doc cleanup |
| v3.8.12 | 2026-03-29 | RuVector Phase 5: pattern intelligence, graph learning, scale optimization |
| v3.8.11 | 2026-03-27 | YAML pipelines, heartbeat CLI, economic routing, session cache (Tier 3) |
| v3.8.10 | 2026-03-26 | Fix coverage data pipeline — eliminate fabricated scores, unify key ecosystem |
| v3.8.9 | 2026-03-25 | Multi-language coverage parsers, SHA-256 witness security, P1 scale benchmarks |
| v3.8.8 | 2026-03-24 | MCP-free agents, aqe memory CLI, WASM parsers, skill description improvements |
| v3.8.7 | 2026-03-23 | Unified hypergraph persistence, CLI/MCP query tools, code index extractor |
| v3.8.6 | 2026-03-23 | Security ReDoS fixes, flaky timer elimination, module decomposition |
| v3.8.5 | 2026-03-21 | Both P1 items resolved, structured logging, hooks decomposition, test stability |
| v3.8.4 | 2026-03-19 | P0 security fixes, DB corruption repair, 30-40% smaller bundles, CI parallelization |
| v3.8.3 | 2026-03-18 | ADR-086 skill overhaul (84 skills), 5 new skills, 7 bug fixes |
| v3.8.2 | 2026-03-17 | 8 bug fixes from user testing, validation pipeline helper, ruflo rebrand |
| v3.8.1 | 2026-03-17 | Fix MCP tool prefix mismatch in 8 agent files, permission pattern fix |
| v3.8.0 | 2026-03-16 | RuVector: native HNSW (150x faster), neural routing, coherence safety gates |
| v3.7.22 | 2026-03-14 | Hook path resolution fix, SQLite corruption prevention, v2 migration removal |
| v3.7.21 | 2026-03-13 | Agent dependency intelligence, shell injection fixes, semgrep parallel SAST |
| v3.7.20 | 2026-03-12 | Fix duplicate brain-checkpoint hooks and governance time budget blocking tools |
| v3.7.19 | 2026-03-12 | YAML pipelines, validation pipeline, cross-phase signals, heartbeat scheduler, context sources |
| v3.7.18 | 2026-03-11 | Hotfix: agents not installed on aqe init --auto upgrade |
| v3.7.17 | 2026-03-11 | BMAD-inspired: adversarial review, agent overlays, validation pipelines, branch enumerator |
| v3.7.16 | 2026-03-10 | Critical: test DB isolation fix, Tier 3 baselines, MCP persistence pipeline |
| v3.7.15 | 2026-03-09 | Six Hats learning improvements, Proof-of-Quality CLI, MCP tool scoping, test stability |
| v3.7.14 | 2026-03-08 | Brain export v3.0 (25 tables), witness chain v3, streaming, npm packaging fix |
| v3.7.13 | 2026-03-07 | Trigger optimizer, version comparator, skill intent classification |
| v3.7.12 | 2026-03-06 | Fix CLI crash on global install (lazy-load typescript), release verification hardening |
| v3.7.11 | 2026-03-06 | Full @claude-flow/guidance governance integration across all 8 modules |
| v3.7.10 | 2026-03-05 | Fix MCP path resolution, CRLF skill parsing, stale v3/ refs; README rewrite |
| v3.7.9 | 2026-03-05 | Multi-language test gen (8 langs), 384-dim embedding standardization, trust tier evals |
| v3.7.8 | 2026-03-04 | Loki-Mode: 7 adversarial quality gates (anti-sycophancy, test gates, EMA, auto-escalation) |
| v3.7.7 | 2026-03-02 | Scrapling-inspired browser: adaptive locators, resource blocking, stealth client, page pool |
| v3.7.6 | 2026-03-02 | 30 MCP tools wired, 9 security fixes, dependency cleanup, dead code removal |
| v3.7.5 | 2026-03-01 | Flat project structure, RVF binary brain export, native module crash fix |
| v3.7.4 | 2026-02-28 | 8 new platform integrations (Copilot, Cursor, Cline, Kilo, Roo, Codex, Windsurf, Continue.dev), 202 new tests |
| v3.7.3 | 2026-02-27 | 17 bug fixes (Phase 0), CI-native exit codes, Phase 1 complete |
| v3.7.2 | 2026-02-25 | AWS Kiro IDE integration, hono security patch |
| v3.7.1 | 2026-02-24 | OpenCode integration, RVF production wiring, 8 bug fixes |
| v3.7.0 | 2026-02-23 | RVF Cognitive Container integration: MinCut routing, witness chain, brain CLI, HNSW unification |
| v3.6.19 | 2026-02-22 | 6 learning pipeline fixes, event-driven trajectories, 25 new tests |
| v3.6.18 | 2026-02-21 | 6 new Agent Teams MCP tools, auto-team-wiring, team enriched responses |
| v3.6.17 | 2026-02-21 | QE Queen Coordinator 7 fixes, HNSW semantic search, hierarchical topology |
| v3.6.16 | 2026-02-21 | Node.js node:test generator, smart assertions, 12 bug fixes |
| v3.6.15 | 2026-02-20 | Fix 11 test generation bugs (#295), fix cloud sync pg loading |
| v3.6.14 | 2026-02-20 | Fix test OOM crashes, CI artifact upload, flaky test stabilization |
| v3.6.13 | 2026-02-19 | Fix test stubs, MCP crash, vector mismatch, coverage, security scan |
| v3.6.12 | 2026-02-19 | Fix HNSW crash, real security scanning, hooks directory, 5 issues fixed |
| v3.6.11 | 2026-02-18 | MCP stability, multi-language support, smart init merge, 9 issues fixed |
| v3.6.10 | 2026-02-18 | QCSD Production Telemetry, eval-driven workflow, KG test generation |
| v3.6.9 | 2026-02-17 | Code quality overhaul, self-learning feedback loop, P0 security fixes |
| v3.6.8 | 2026-02-15 | MCP auto-init fleet, unified experience persistence, fix #262 |
| v3.6.7 | 2026-02-14 | Fix DB corruption, remove VACUUM, eliminate dual-DB split-brain |
| v3.6.6 | 2026-02-13 | Fix Node 22+ crash, restore self-learning hooks, single-DB resolution |
| v3.6.5 | 2026-02-13 | Fix fleet_init dimension mismatch (#255), agent asset sync, README fixes |
| v3.6.4 | 2026-02-12 | Security hardening, Dream Scheduler, QE pattern seeding, delta scanning |
| v3.6.3 | 2026-02-11 | QX Analysis, SFDIPOT product factors, cross-platform compatibility |
| v3.6.2 | 2026-02-10 | Init stability fixes (YAML parser, agent helper file placement) |
| v3.6.1 | 2026-02-09 | Agent Teams Integration, Distributed Tracing, Dynamic Scaling |
| v3.6.0 | 2026-02-08 | Enterprise Integration Domain, 8 new agents, Pentest Validation |
| v3.5.0 | 2026-02-04 | Governance by default, QCSD 2.0 Lifecycle, Self-Healing Enterprise |
| v3.4.2 | 2026-02-02 | Skill Validation System, Trust Tiers |
| v3.4.0 | 2026-02-01 | AG-UI, A2A, A2UI Protocols |
For the full changelog with all minor and patch versions, see v3/CHANGELOG.md.