What it catches
- false closure
- weak verification
- context bloat
- environment mismatch
- loop control failure
Local-first session evidence
AI Workflow Coach
A local-first CLI and agent skill that reads recent coding-agent sessions, builds an inspectable evidence bundle, and returns one short coaching note with real quotes, conservative attribution, and a next-session check.
evidence_bundlevalidated
candidatefalse-closure-after-failed-tool
spanturn-47: tool_result failed
claimDone. Tests pass.
surfaceskill checklist, not AGENTS.md
surface
CLI + skill
evidence
.ai-coach/
output
one note
What it emits
- evidence_bundle from deterministic parsing
- candidate_patterns before coaching judgment
- a short note with a quote and one experiment
What it avoids
- not a transcript browser
- not a productivity dashboard
- not a workflow framework
v0
- Consent-first /coach flow
- Deterministic candidate patterns
- One short coaching note, not a dashboard
v1
- Fixture-backed adapter parity
- Repo-local versus cross-tool findings
- Clear recommended surfaces for docs, skills, hooks, and checklists