cowork-review

Cowork — prompt lifecycle + review gates

Worker-executable work lives as markdown prompt files under tasks/{todo,run,done} (gitignored). Move with plain mv — except run/→done/, which only promote may perform.

Prompt frontmatter (MANDATORY)

---
status: draft        # draft | in-review | changes-requested | approved | blocked-review
review_round: 0      # incremented on each resubmission
# --- output-gate fields (populated after the Worker runs) ---
output_status: none  # none | pending | approved | changes-requested | blocked-review
output_round: 0      # incremented by review-output on each verdict
verification:        # filled by `promote` (build/tests) and by the Worker (evidence)
  build: unknown     # unknown | pass | fail
  tests: unknown     # unknown | pass | fail
  evidence: ""       # path to screenshot/log; required for UI tasks
---

A Worker is spawned ONLY when status: approved. A prompt reaches done/ ONLY when output_status: approved + build/tests green + evidence satisfied. Filesystem + frontmatter are the single source of truth.

Review gate (before any Worker execution)

1. Architect drafts the prompt in tasks/todo/ (status: draft, review_round: 0).
2. Architect submits it (status: in-review) to the Analyst (a review step — in the single-agent chatboard this is a dedicated review pass / a security-or-review agent like Argo, not a third session-role).
3. Analyst writes a verdict (approved / changes-requested) into the frontmatter and APPENDS one round to ## Review log (never overwrites prior rounds).
4. changes-requested → Architect revises in place, review_round++, status: in-review, resubmit.
5. approved → proceed to execute.
6. Loop cap: after 3 rounds without approved → status: blocked-review, escalate to the developer.

Analyst defaults to changes-requested when the prompt is ambiguous/under-specified/unverifiable/risky. Approval is earned. Analyst never writes code or moves files.