QE fleet reports generated using Claude Opus 4.6 (11 reports), GLM-5 (9 reports),
and Qwen 3.5 (9 reports) analyzing the same codebase. Includes a cross-model
analysis with devils-advocate verification of claims against actual codebase.
Key finding: model choice fundamentally determines report reliability, with
factual accuracy ranging from 88% (Opus) to 25% (Qwen). Qwen fabricated
27 CRITICAL vulnerabilities by confusing SQLite db.exec() with child_process.exec().
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Brain export v3.0 (25 tables), witness chain v3 with Ed25519 signing,
streaming for large tables, RVF freeze/derive/indexStats, npm packaging fix.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The massive file count (5473 files / 63MB) triggered npm's ENOTEMPTY race
condition during global install upgrades. Narrowed package.json `files` field
to exclude 2144 source maps and 36 dev-only scripts while keeping all 8
consumer-facing scripts (lifecycle hooks, skill validators, fetch-content).
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
CodeQL flagged `parseSkillFrontmatter` regex `/^---\s*\n([\s\S]*?)\n---/`
as polynomial-time on crafted input. Replaced with indexOf-based parsing
that avoids backtracking entirely.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Implements three improvements from Claude Blog skill-creator analysis:
- Trigger Optimizer: analyzes skill descriptions/tags for false positive/negative activation risks
- Version Comparator: A/B testing between skill versions with Cohen's d effect size
- Skill Intent: classify skills as capability_uplift, encoded_preference, or hybrid
Includes 52 unit tests (23 + 29), schema sync across all 3 copies, and ADR-056 update.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
TypeScript was an external ESM dependency loaded at bundle startup,
crashing `aqe --version` when typescript isn't installed (global
installs). Now lazy-loaded via createRequire Proxy — only triggers
when AST parsing features are used.
Also adds isolated install verification to /release skill (steps 8e
and 15) to catch missing externals before they reach users.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
All governance modules now properly load AND wire their guidance
counterparts using a local-first pattern where AQE logic is
authoritative and guidance supplements:
- continue-gate: maps ContinueDecision with full StepContext
- memory-write-gate: local-first evaluation order, proper WriteDecision mapping
- adversarial-defense: ThreatDetector + CollusionDetector wiring
- deterministic-gateway: gateway.evaluate() supplements tool-call gating
- proof-envelope: parallel audit trail via ProofChain.append()
- shard-retriever: embedding-based relevance boosting via retrieve()
- evolution-pipeline: signed change proposals on promote/demote
- trust-accumulator: guidance trust scoring integration
Also adds ./governance subpath export to package.json for consumers.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Replace weekly downloads badge with monthly and total download
badges for better visibility of package adoption.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Replace unindented heredoc (<<EOF) with echo statements in the
step summary. The heredoc content at column 0 broke YAML block scalar
parsing since it was inside a `run: |` block that requires indentation.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Add MAX_REGEX_LINE_LEN (1000) guard before every regex .match() call
in multi-language-parser.ts. This is the CodeQL-recommended mitigation
for polynomial ReDoS — caps worst-case at O(1M) which is sub-millisecond.
Source code lines above 1000 chars are not valid function/class declarations
and are safely skipped. Brace-depth tracking is preserved for Go/Kotlin
parsers that need it.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Line 356: split(/ +as +/) → split(' as ') to avoid polynomial scan
- Lines 502, 633: import/using \s+ → space-only, capture starts with \S
- Line 891: use + capture starts with \S to prevent overlap with leading spaces
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Fix 16 polynomial ReDoS in multi-language-parser.ts by eliminating
overlapping quantifiers between [^X]+ capture groups and \s* delimiters
- Fix incomplete sanitization in language-detector.ts: replace('*','')
only removes first occurrence, changed to replace(/\*/g, '')
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The release skill is a project-internal workflow orchestrator, not a
QE skill that produces JSON output. Tier 3 requires schemas/output.json
which doesn't apply here.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Fix MCP entry path resolution in mcp.ts (1 level up, not 2)
- Remove stale v3/ directory references in 10-workers.ts
- Fix CRLF line endings in 170+ skill files breaking frontmatter parsing
- Add CRLF→LF safety net in prepare-assets.sh
- Rewrite README.md for clarity (1097→280 lines, outcome-focused)
- Sync assets from .claude/skills via prepare-assets.sh
Fixes issues reported in v3.7.9 upgrade (skill-lint failures, MCP errors)
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Matches local dev environment (Node 24 / npm 11) to fix lockfile
compatibility issues causing `npm ci` failures (jose@5.10.0 missing).
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Create eval configs (evals/*.yaml, schemas/output.json, scripts/validate-config.json)
for: qcsd-cicd-swarm, qcsd-development-swarm, enterprise-integration-testing,
observability-testing-patterns, middleware-testing-patterns
- Set debug-loop and pr-review to tier 0 (advisory-only, no eval needed)
- Update trust-tier-manifest.json with correct tier assignments
- Fix prepare-assets.sh to include qcsd-*, n8n-*, and 15 additional QE skills
that were in assets/ but not matched by the copy filter
- Sync all eval infrastructure to assets/skills/ for npm distribution
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Update default embed model from nomic-embed-text to all-MiniLM-L6-v2
in Ollama LLM provider
- Fix 15 test files with hardcoded 768-dim references
- Update mock model names in Ollama health checks
- Update comments referencing old model name
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Previously had a mix of 768 (nomic-embed-text) and 384 (MiniLM-L6-v2)
dimensions across the codebase, causing dimension mismatches between
KG vectors and pattern store vectors. Standardize everything to 384
since all-MiniLM-L6-v2 is the model we ship (no Ollama dependency).
Changed 24 files: EMBEDDING_CONFIG.DIMENSIONS, DEFAULT_VECTOR_DIMENSIONS,
HNSW configs, coverage analysis, governance flags, test generation,
decision transformer, and all hardcoded 768 references.
Also cleaned 22,082 stale 768-dim vectors from memory.db (147MB -> 54MB).
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Add test generators for Go, Rust, Kotlin, Swift, Flutter, JUnit5,
xUnit, Jest-RN, and pytest with language detection, multi-language
parser, compilation validator, test file resolver, mobile/Spring
detectors, and Rust ownership analyzer.
Also includes init phase fixes, dream engine updates, routing
type extensions, and config sync.
Note: TODO comments in generators are template placeholders output
in generated user code, not actual implementation gaps.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Update ADRs 075-079 with implementation details, revise the
multi-language GOAP implementation plan, add test generators and
parser test fixtures for Go, Rust, Java, C#, Swift, Python.
Include skill validation schema updates and v2→v3 migration skill.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Remove redundant /v3/ segment from agent memory namespace paths
(e.g., aqe/v3/accessibility/outcomes/* → aqe/accessibility/outcomes/*).
Namespace identifiers in memory.db remain unchanged — this only
affects the documentation references in agent markdown files.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Pattern count was stagnant because the LearningConsolidationWorker
only runs via the background daemon (30min interval), which wasn't
active. Added lightweight consolidateExperiencesToPatterns() to the
session-end hook so captured_experiences are converted to qe_patterns
at the end of every Claude Code session.
- Aggregates unprocessed experiences by domain+agent
- Creates new patterns for clusters meeting quality thresholds
- Reinforces existing patterns instead of creating duplicates
- Marks processed experiences to ensure idempotency
- Verified: 9 new patterns created from 3,622 pending experiences
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Replace exponential-backtracking regex with two-phase approach:
1. Match block structure using [^}]* (linear time)
2. Strip comments separately to check if body is empty
Fixes CodeQL high-severity alerts for polynomial regex backtracking.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Add comprehensive plan to extend AQE test generation from 2 languages
(TS/JS) to 11 (adding Java, C#, Go, Rust, Swift, Kotlin, Dart, Python
AST fix, React Native, Flutter). Includes GOAP execution plan optimized
for parallel agent swarm execution (up to 5 concurrent agents).
ADR-075: Unified TestFramework type system (single source of truth)
ADR-076: tree-sitter WASM multi-language parser integration
ADR-077: Compilation validation loop for generated tests
ADR-078: Backward-compatible multi-language API extension
ADR-079: Language-specific test file path resolution
GitHub Issue: #319
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Add anti-sycophancy scoring, test quality gates, blind review,
EMA calibration, edge-case injection, complexity-driven composition,
and auto-escalation. All features enabled by default (opt-out).
178 tests, 14 new files, 6 config flags.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Resource blocking with presets (functional/visual/performance) and tracker/ad domain lists
- Adaptive locator with similarity-based element matching for resilient E2E selectors
- Browser page pool for concurrent page management with async acquire/release
- Stealth browser client via Patchright for bot-protected test environments
- Integrated adaptive locator into click/type step executors with fallback chain
- Wired resource blocking into browser orchestrator launch lifecycle
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Profa