Council
Run independent read-only reviewers on different models, then synthesize one verdict. Council is opt-in and expensive; never make it the default verification step.
When to use
- Architecture or design choice with real trade-offs.
- Risky migration, security decision, or data-correctness question.
- A review where a single model may share the author's blind spot.
Don't use for routine implementation, one-file changes, or anything the
normal reviewer gate already covers.
Seats
Dispatch three reviewer tasks in one task batch, each with a different
model and the same question plus enough context to answer cold:
| Seat | Model (example) | Bias |
|---|---|---|
| alpha | commandcode/gpt-5.6-sol |
correctness and failure modes |
| beta | commandcode/zai-org/GLM-5.3 |
architecture and maintainability |
| gamma | commandcode/moonshotai/Kimi-K3 |
edge cases and operability |
Rules:
- One batch, so the seats run concurrently.
- Every seat is read-only: it may read the repo and run read-only checks; it never edits.
- No seat sees another seat's output. Independence is the point; do not ask one seat to critique another.
- If a seat fails or returns empty, continue with the remaining seats and say which one is missing. Never silently replace it with the same model twice.
- Every seat prompt names the exact files, the question, and the required output. The child inherits no conversation.
- If the seats are still running, use
wait; do not poll.
Seat prompt shape:
Review the decision below as an independent council seat.
Question: <one paragraph>
Scope: <exact files and behavior to inspect>
Your bias: <correctness | architecture | edge cases>
Output: findings with file:line, severity, and a one-line recommendation.
Do not edit anything and do not ask follow-up questions.
Synthesis
Read each seat at agent://<name> (or the delivered task result). Synthesize
on the N0 when the N0 model is at least as strong as the seats. Otherwise
dispatch one more reviewer with the strongest model, passing the question
and the three reports, and use its synthesis.
The final report MUST contain:
- Verdict - the recommended answer, in one paragraph.
- Confidence - high / medium / low, and why.
- Per-seat findings - what each seat found, compressed to the points that changed or confirmed the verdict.
- Dissent - disagreements and the evidence behind them. Never average away a dissenting finding.
Cost
Three strong-model reviews cost roughly three times a single review. Use it when a wrong decision is more expensive than the review, not as a ritual.