plan-review
Invokes the merged plan-reviewer agent (agents/plan-reviewer.md) BLIND.
The agent carries the whole contract itself: honesty bounds, surface enumeration, grounding, the 7-point checklist, the elegance gate, bounded findings with a coverage ledger, and the revised handoff plan.
This skill only builds a clean blind prompt and handles the result.
The one rule that makes it work: the reviewer must not see the author's reasoning. A reviewer that recognizes the author's rationale rates the work higher, so the isolation must be structural.
Steps
- Resolve
$ARGUMENTSto the artifact path. If empty, ask which plan; never guess. - Record the artifact's content fingerprint. Reuse an existing durable report when that fingerprint and the review criteria are unchanged.
- Build the agent prompt with ONLY:
- the artifact path to read;
- the hard constraints the solution must satisfy, stated as bare facts (settled rulings, budget, deadlines);
- pointers to the criteria or rules files the review should judge against. NEVER include: why the approach was chosen, the planning conversation, prior review verdicts, or your own opinion of the artifact.
- Spawn one
plan-revieweragent with that prompt. It does not rerun an unchanged full suite. A focused check needs a named evidence gap. - Save the report with the content fingerprint. Write the revised handoff plan next to the original, present the findings and verdict, and surface every deferred decision.
- On REJECT there is no revised plan. Present the findings and stop.
