v3.9.30: Six MCP fixes (#467, #469, #470, #472, #473, #474)

What's New

Six user-reported MCP issues fixed, each verified end-to-end by replaying the reporter's exact reproduction against the freshly-built bundle.

Fixed

• qe_requirements_quality-criteria and qe_workflows_browser-load no longer crash with __dirname is not defined (#467). Added ESM polyfills across 7 files that were still using the CommonJS __dirname global.

• qe_embeddings_search now returns real semantic matches (#469). Replaced the random-vector mock with @xenova/transformers' all-MiniLM-L6-v2. Cosine similarity of "apple pie" ↔ "apple fruit pie recipe" is now ~0.78 (was -0.07). Also fixed threshold || 0.5 → ?? so threshold: 0 is honored.

• qe_coherence_collapse no longer returns contradictory riskLevel: "critical" with isAtRisk: false (#470). NaN from missing input fields no longer falls through categorizeRisk to "critical". isAtRisk is now derived from riskLevel so the two can never disagree.

• qe_tests_schedule no longer crashes with phases is not iterable (#472). Falls back to DEFAULT_TEST_PHASES (unit → integration → e2e) when callers don't specify a phase list.

• session_cache_stats now reports real hit rates (#473). Wired SessionOperationCache into the MCP tools/call dispatch for isConcurrencySafe tools. Repeated identical calls now short-circuit through the cache. Disable with AQE_SESSION_CACHE=off.

• test_generate_enhanced now honors the language parameter (#474). Unsupported languages are rejected with a clear error and the list of supported languages. Framework auto-derives from language (python → pytest, go → go-test, etc.) instead of always defaulting to vitest. Framework/language mismatches are rejected. The MCP schema enum exposes all 16 generators we ship.

Getting Started

npx agentic-qe init --auto

See CHANGELOG and v3.9.30 release notes for full details including failure modes and mitigations.