Anti Skill v2.2.0 — Security Hardening & Release Readiness
Security Fixes
- Reflection files now written at 0600, directory at 0700; legacy world-readable files migrate on next write
- Reflection tests isolated to tempdir; no more production data pollution
Bug Fixes (External Agent Report)
- Plan scope grounding: language/framework profile injected into plan prompts so models stop guessing TypeScript on Python repos
- Chunk-cap fail-fast ordering: panel review pre-checks chunk overflow before spending any model call
- VALIDATION_REQUIRED 403 recovery: Google re-auth URL surfaced with [ACTION REQUIRED] guidance; query-param tokens redacted
- Error-path run records: failed runs populate models, prompt_chars, output_chars, failed_lanes
New Features
runs clean --older-than Nnow prunes stale reflection files too- Smoke version-drift check works on Python 3.10 via tomli/regex fallback chain
- Polyglot repos report all matching language manifests
Documentation
- SKILL.md documents
runs reflectionsand passive tracking semantics
Tests
- 719 tests + 204 subtests (up from 702), full CI matrix green (Python 3.10-3.14, Ubuntu + Windows)
Full changelog: https://github.com/Reedtrullz/codex-antigravity-auth/blob/main/CHANGELOG.md