run-supervised-checkpoint
Approve, execute, verify, repair, continue, independently review, preview, or explicitly finalize the current CEWP supervised checkpoint through Core gates. Use when the user asks to run or continue a
Files1
View on GitHubSKILL.mdplugins/cewp/skills/run-supervised-checkpoint/SKILL.md
GitHubRun A Supervised Checkpoint
Use canonical JSON from cewp supervise; generated Markdown is a view, not mutable state. Keep the ChatGPT/Codex host in the operator role while CEWP owns the isolated child worktree.
- Run
cewp supervise status [run-id] --jsonand state the current run/checkpoint status, plan revision, owner/backend, latest evidence, budget consumption, warnings, blocker, and next safe action. - Perform no more than one CEWP-controlled model operation per invocation. Local verification commands do not count as model operations, but their approved limits still apply.
- Follow only the transition matching canonical state:
proposed: show the plan and runcewp supervise approve <run-id> --yes --jsononly after explicit approval. Stop before dispatch.approved/ready: if the operator explicitly requested a Codex task class, model, or effort, record it first withcewp supervise effort <run-id> --operation implementation --task-class <class> [--model <model>] [--effort <effort>] --yes --json. Never infer model or effort from the class. Then runcewp supervise execute <run-id> --yes --jsononly after explicit execution approval. If dispatch reachesverifying, runcewp supervise verify <run-id> --json; then stop at the verified, repair, paused, or blocked result.needs-repair/repair-ready: explain the failure signature and remaining repair allocation. Record any explicitly requested repair-specific Codex selection throughcewp supervise effort ... --operation repair ... --yes --json; do not reuse or change implementation settings implicitly. Runcewp supervise retry <run-id> --yes --jsononly when the user explicitly chooses retry, then run local verification and stop.checkpoint-complete/verified: offer final review or a safe pause for one manually bounded next checkpoint. For final review, record any explicitly requested independent-review Codex selection throughcewp supervise effort ... --operation reviewer ... --yes --json, recordcewp supervise continue <run-id> --json, runcewp supervise review <run-id> --yes --json, and stop at the reviewer decision. For more work, hand off to the resume workflow without dispatching another model operation.review-passed: runcewp supervise receipt <run-id> --jsonto preview the receipt. Do not finalize in the same step unless the user already gave explicit finalize intent after seeing equivalent receipt facts.ready-to-finalize: runcewp supervise finalize <run-id> --yes --jsononly after explicit finalization approval.- any
paused-*orblocked: do not dispatch. Present Core recovery actions and hand off to the resume workflow.
- Treat any nonzero command, scope warning, failed verification, missing reviewer decision,
REQUEST_CHANGES,BLOCK, exhausted allocation, host limit, or ownership mismatch as a closed gate. - Report the resulting canonical status and exact next safe command. Never label dispatch completion, partial files, or a reviewer request as checkpoint success.
Do not edit the managed worktree from the host task, bypass --yes, borrow protected allocations, automatically add budget, switch backend/provider, merge, push, publish, tag, or release.