v0.21.0
neo learns from AI-tool transcripts, and gains optional CAR-first inference.
Added
- Transcript learning (#97): the observer mines AI-tool session transcripts each cycle → verified, generalizable PATTERN/FAILURE facts (verbatim-evidence + adversarial-judge gates; probationary, capped confidence; bounded per cycle).
- Multi-source ingestion (#98, #99): a
TranscriptSourceadapter interface — Claude Code + Codex (project-scoped), CAR (global-scoped). - CAR-first inference (#100):
inference_mode="auto"routes through CAR's router with graceful fallback; defaults to"static"(gpt-5.5) until a CAR release verifies router quality.
Changed
add_factdomainparam →Fact.domain(domain-filtered retrieval); observercar-runtimeref alignment (#96).
Full notes in CHANGELOG.md.