Review
- Two reviewers read the same changeset, code-reviewer for correctness and ponytail for size, so a blind spot in one is caught by the other.
- /delegate owns the dispatch Prompt; /orchestrate owns judging what returns.
1. Read the changeset yourself first
Read every changed file against the Architecture before either reviewer is dispatched
Your own findings are the baseline the two returns are measured against, so a problem neither reviewer saw still surfaces. Never: a Review whose every finding arrived in a reviewer's report.
2. Dispatch both reviewers on one target
Use /delegate to send the code-reviewer and ponytail Agents in the same message, both given the same target and the same Verification.
Give both reviewers the identical target
Same diff, same surrounding code, same Prompt. A difference in scope makes the two returns incomparable.
Require a cause, never a symptom
The dispatch tells each reviewer to trace every finding to its cause before reporting it, and to report that cause. A reviewer that reports what it saw sends you one round per symptom of the same cause.
3. Triage the returns with /triage
Append each round to the one triage.md
Each round appends its own section to the run's one triage.md, so a recurring finding is visible and a clean round is checked against the sheet, never against memory.
4. Re-run both reviewers on the fixed changeset
Repeat steps 1 through 3 on the updated changeset.
Stop only on a clean round
The Review ends when one full round from both reviewers returns no breaking finding. A round that fixed something is never the last round.