stage-coach

Use this skill for queued per-stage coach reviews from .cocoplus/sentinel/coach-requests.jsonl.

Objective

Evaluate completed CocoFlow stages with a separate coach model so generator and critic remain structurally separated.

Workflow

1. Verify .cocoplus/ exists.
2. Read .cocoplus/sentinel/coach-requests.jsonl.
3. For each queued request, collect:
   • stage role and task brief from .cocoplus/flow.json
   • declared artifacts.reads and artifacts.writes
   • produced artifact paths under .cocoplus/flow/artifacts/<run-id>/
   • downstream missing artifact status
4. Verify coach_model differs from the executor model. If not, write .cocoplus/sentinel/known-gaps.jsonl and skip the request.
5. Run an external coach evaluation using the configured coach model.
6. Write .cocoplus/sentinel/stage-quality.jsonl:

{
  "ts": "...",
  "stage_id": "function-designer",
  "score": 7,
  "reasoning": "Specific evidence-backed rationale.",
  "attributed_to": "function-designer",
  "coach_model": "haiku"
}

1. Mark consumed requests as processed by writing .cocoplus/sentinel/coach-requests.processed.jsonl.

Scoring

• 8-10: usable output with clear artifact handoff
• 5-7: core output exists but a declared contract, evidence item, or handoff is incomplete
• <5: missing required output, ungrounded result, or downstream blocker attributable to this stage

Exit Criteria

• Each score cites produced or missing artifacts.
• The coach never evaluates its own generation.
• Missing artifacts are attributed to the stage that declared them.
• Console-readable stage-quality.jsonl is updated.

Anti-Rationalization

Shortcut / Temptation	Why It Fails
Treat the skill as complete because the file exists	Skill contracts must describe observable behavior and verification, not just command names.
Skip artifact and safety checks for a small command	Small commands still mutate state or guide execution; preserve the same gates.