majiayu000 New space
Contract
| Field | Bound contract |
|---|---|
| Trigger | User starts a new work session and asks to split human decisions from agent execution. |
| Authority | Write only under run/<session-id>/ and execute only the approved spawn target. Human sign-off remains external to this skill. |
| Side effect | Five artifacts under run/<session-id>/: brief, spawn evidence, work record, review artifact, and human-gate record. |
| Done | All five artifacts exist under run/<session-id>/: the brief is approved, spawn output is captured with exit status, the work record reflects actual execution, the review artifact states conformance and verification items, and the human-gate record carries the decision, timestamp, and conditions. |
Inputs
- Approved brief (required): human-authored statement of work goal and constraints. State what is optional and what must be supplied.
- Session id (required): caller-supplied, filesystem-safe, unique under
run/. Reject if missing, not filesystem-safe, or collides with an existing directory. - Spawn command or agent target (required): the command or agent to launch, what it may read, what outputs it must produce, and what it must not change.
- Review criteria (required): the standard the work is checked against.
- Context (optional): current directory, project layout, open issues, or prior session artifacts.
Procedure
- Validate the session id and approved brief. Confirm the session id is filesystem-safe and unique under
run/. Confirm the user approves the brief. Halt if the intent is ambiguous, the brief is unapproved, or the session id is missing or colliding. Done when: the session id is validated and the brief is approved. - Write the brief to
run/<session-id>/brief.md. Include the work goal, scope, constraints, and done criteria. Done when:brief.mdexists at the stated path. - Execute the spawn target. Capture stdout, stderr, and the exit code. Write the spawn evidence to
run/<session-id>/spawn.md. Done when: spawn output is captured with exit status recorded inspawn.md. - Record the work phase. Document the agent's actual execution path, outputs, and any deviations from the brief in
run/<session-id>/work.md. Done when: the work record reflects the actual execution path and deviations. - Record the review artifact. Write what was produced, whether it meets the brief, and what the human must verify to
run/<session-id>/review.md. Present it to the user. Await explicit sign-off or revision request. Done when: the review artifact states production status, brief conformance, and human verification items, and the user signs off or requests revision. - Collect and record the human gate. On sign-off, write
run/<session-id>/human-gate.mdwith the decision, timestamp, and any conditions. Done when: the human-gate record is written with decision, timestamp, and conditions.
Failure and recovery
- Missing or colliding session id: stop. Ask the user for a filesystem-safe, unique session id before proceeding.
- Unapproved brief: stop. Ask the user for a concrete work goal. Do not invent scope.
- Spawn failure: attempt one retry with a corrected spawn definition. Record the failure and retry count in
spawn.md. Stop if the second attempt fails. - Review incomplete: present the divergence to the user. Do not proceed until the user resolves or approves.
- Human gate withheld: the session is not done. Record the last state and stop.
- Partial-result rule: if any phase halts midstream, all durable artifacts written up to that point remain. No automatic rollback.
Output
One workspace directory (run/<session-id>/) containing brief.md, spawn.md, work.md, review.md, and human-gate.md in that order. The session is executed, not merely prepared: spawn evidence and the final human-gate state are both present.