Council

Resolve the council members before review. Use the member list, providers, or count the user requests. By default, select one strong model from each of three different providers at high reasoning effort, such as Claude Opus, GPT Sol, and Gemini Pro.

In T3 Code, resolve models with orchestrator_capabilities. Elsewhere, check the harness's available subagent tools. Report unavailable members and keep their requested identities. If the user leaves the member list open and does not require cross-provider coverage, a harness without all three default providers may run one read-only round with available native agents. Label it as native-only and report the missing provider coverage. If no delegate can run, report the limitation and stop.

After determining the review mode, inspect outstanding PR comments and run relevant read-only checks. Fix known issues before review only when the request authorizes implementation and the run is outside the native-only fallback. Otherwise include the known issues in the brief. Skip fixes to work the user wants to discard.

Before dispatch, verify that the harness enforces a read-only boundary for each delegate's filesystem access and mutation-capable tools. A read-only brief, review role, approval-required mode, or writable worktree is not an enforced boundary. If the boundary is unavailable, do not dispatch that delegate; report its missing coverage. Do not invent a read-only runtime mode or widen permissions to obtain coverage.

After preflight, start the selected eligible delegates in parallel. Give them the request and relevant context. Every delegate stays read-only even when the coordinator is authorized to implement fixes. Collect their responses and summarize the findings, disagreements, and next steps.

Set a ten-minute deadline when dispatching each round. Cap it by any remaining time budget in the request. Use bounded waits or completion notifications with deadline support. At the deadline, stop waiting and request cancellation for this round's outstanding tasks where supported. Report missing coverage with the findings already received and distinguish unconfirmed cancellation. A timeout by itself does not start another round.

For a code review, ask each delegate for a score out of 10. Count each full review pass as one round. A fresh review after fixes starts a new round.

If the current request authorizes editing the PR description, place the results before its model and harness line. Use this format in the report.

## Council review

Rounds: 2

| Delegate    | Round 1 | Round 2 |
| ----------- | ------- | ------- |
| Claude Opus | 7/10    | 9/10    |
| GPT Sol     | 8/10    | 10/10   |
| Gemini Pro  | 7/10    | 9/10    |

If a delegate gives no score, write no score. Do not invent one. If there is no PR or the request leaves the description outside scope, put the results in the final reply.

A review-only request or native-only fallback ends with the first round's findings. For other runs with implementation authorization, apply agreed, in-scope fixes and request a fresh review while a round remains. Stop when no warranted fix remains or three review rounds have completed. Report any remaining findings, disagreements, or gaps in verification. Keep every round within the original scope.

<!-- CORE-OVERLAY-START -->

This repository (Asymmetric-al/core)

On conflict, follow the current user request, root AGENTS.md, the nearest nested AGENTS.md, and the matching docs/ai/rules/* rulebook. Provenance and refresh steps live in references/upstream.md.

Triggers

• Use when the user explicitly asks for a council review or invokes /council (Claude Code), $council (Codex), or the standalone council command. A mention of a council in other content does not invoke the skill.
• Honor an explicit council request with one member. Route other single-reviewer requests to the usual review workflow.

Workflow

1. Resolve council members from the current request first. An explicit member list, provider list, or count (for example "use only Claude and Codex") overrides the upstream three-provider default for that run. Otherwise use the upstream default, subject only to the labeled native-only fallback above. Report actual members and missing coverage in the results. Never replace an explicitly requested provider or model.
2. In T3 Code, call orchestrator_capabilities once per thread, and again only after an error or a user-named provider change. Verify the enforced read-only boundary above before starting one delegate_task per eligible member in parallel. Set target.providerInstanceId and target.model and use target.options for advertised reasoning settings. Use a distinct clientRequestId per member and round, reusing it on retries of that dispatch. Start a fresh task for each new round. Use mode: "wait" with timeoutMs within the remaining deadline when waiting for results, or follow next with a separate deadline mechanism. A wait timeout leaves the task running. Use task_cancel on outstanding task IDs from the current round when it expires.
3. Brief each delegate with the request, the diff or files in scope, and the applicable Core rules: root AGENTS.md "Code review rules", the nearest nested AGENTS.md, and the matching docs/ai/rules/* rulebook. Keep review delegates read-only. Never put secrets, .env values, tokens, or credentials in a brief.
4. Weight findings by Core's review priorities: data loss, tenant/auth bypass, money/payment/webhook errors, migration hazards, crashes, and real regressions, each with concrete file:line evidence. On later rounds, accept only new blocking or important findings; do not run fix-then-nit loops.
5. Apply authorized fixes under Core rules (TDD for behavior changes, smallest relevant checks first, protected-path procedures) within the round limit above. Push, merge, or edit a PR description only when the current request authorizes that action.

Checklist

• Members came from the request, the upstream default, or the permitted labeled native-only fallback; unavailable coverage was disclosed and explicitly requested providers were not substituted
• Review delegates had verified enforced read-only boundaries and received no secrets; unavailable boundaries were reported as missing coverage
• Each result is attributed to the model that actually responded; a missing score reads no score
• Source edits were authorized and the report states any unresolved disagreements or missing coverage
• Scope stayed limited to the original request and the review round cap
• Skill edits change docs/ai/skills/council/ (never a generated mirror) and update references/upstream.md before running bun run skills:sync and the non-mutating bun run skills:verify

<!-- CORE-OVERLAY-END -->