WhenLabs-org/when

Major Changes

• 8019f9f: Trim stale: drop AI analyzers, fix, watch, sarif.

  Per the audit, stale's AI-powered analyzers (--deep / semantic /
  completeness / examples) had real Claude API calls wired up but zero
  test coverage. The fix command (580 LOC of auto-rewrite logic) and
  watch command (chokidar-based re-scan) both overlap with standard
  linter/IDE workflows.

  Removed:

  • src/analyzers/ai/ (semantic, completeness, examples + client)
  • src/commands/fix.ts, src/commands/watch.ts
  • src/reporters/sarif.ts
  • --deep and sarif format flags from CLI + Tool API
  • DriftCategory: dropped semantic, completeness, example,
    architecture, response-shape (all AI-only)
  • AiAnalyzer type, getAiAnalyzers()
  • config.ai.* fields + merging
  • Dependencies: @anthropic-ai/sdk, chokidar, handlebars
  • Corresponding tests + snapshots + GitHub Action deep input

  The wrapper's stale_scan MCP tool no longer advertises deep or
  sarif format (@whenlabs/when patch bump).

  Kept: scan, init, 10 static analyzers, terminal/json/markdown
  reporters, --git flag, the GitHub Action.

  Major version bump for stale.

Patch Changes

• 4360845: Post-trim cleanup: sync READMEs with the actual command surface, drop a dead chokidar dep, and fix stale's summary passed count.

  • READMEs rewritten for the 5 trimmed tools so they match what the CLI actually ships. Removed references to commands and flags that no longer exist (vow fix|hook|audit|diff|policy, vow --offline|--api-key|ANTHROPIC_API_KEY; stale fix|watch, stale --deep, STALE_AI_KEY, SARIF format; aware watch|validate|doctor|add, --exit-code; envalid onboard|hook|export|watch|fix|migrate, plugins, secret providers, framework adapters). Documented the flags each command actually accepts today (e.g. aware diff --check|--json|--target|--quiet, vow check --ignore).
  • aware: removed unused chokidar dependency (carried over from the dropped aware watch command — grep chokidar src/ had zero hits).
  • stale: fixed summary.passed going negative on reports with many issues. buildSummary was computing totalChecks - errors - warnings - infos, where totalChecks was per (doc × analyzer) but issues are per finding, so a heavy report trivially overflowed it. totalChecks now counts analyzers run, and passed counts analyzers whose category produced zero issues. Per-category passed is now 1 when that analyzer ran and produced no issues, 0 otherwise. Test fixture + snapshot updated for the post-trim DriftCategory set.
  • vow: deleted docs/workflows/ — the three example workflow YAMLs and their README referenced vow check --offline, vow diff, vow policy compile, ANTHROPIC_API_KEY, and the archived whenlabs-org/vow@v1 composite action, none of which exist anymore.

• dbbb266: Drop stale GitHub Action inputs that no longer do anything.

  packages/stale/action/action.yml still advertised deep ("Enable AI-powered deep analysis (requires STALE_AI_KEY secret)") and listed sarif as a valid format value, but both were removed from the tool in the trim — run.ts no longer reads deep and parseFormat silently falls back to terminal when passed sarif. Removed the dead input and trimmed the format description to terminal, json, markdown so the Action surface matches reality.