full-council

Full Council

Scope

Get a multi-model analysis of a subject: code (files, a diff, a PR or a module) or a problem described in the conversation. A Moderator picks three categories for the subject, gives each to two Analysts, one Codex and one Claude, and turns their six reports into one report of at most three sentences. The main agent (you) only orchestrates: it gathers context, starts the Moderator and relays its report. It does not analyze the subject itself.

Each role's model, effort, responsibilities and output live in its reference file; use them as written unless the user overrides a setting:

• Moderator
• Analyst

Invocation authorizes reading code and context and starting the agents above. Every agent is read-only: no file edits, commits, pushes, PR comments or other publication. Honor narrower user constraints and pass them to every agent.

Workflow

Agents start with no parent history, so every brief must be self-contained: include the subject, constraints, relevant context and the absolute path of the role's reference file, and tell the agent to read it first. The Moderator also needs the absolute path of the Analyst's file to pass it down. In T3 Code, check orchestrator_capabilities once per thread and start agents with delegate_task using the provider, model and options from the reference file; follow each task's next without polling.

1. Identify the subject and the user's question. For code, collect the relevant paths, revision or diff, repository instructions and conventions. For a chat problem, restate it with its goals, constraints and known facts from the conversation. Resolve verifiable repository facts yourself; ask the user only if the subject or question is genuinely unclear.
2. Start the Moderator (role: "review", shared checkout) with that context and any categories the user requested. It picks the categories, runs the six Analysts in parallel according to its reference file and returns the report. Do not start the Analysts yourself.
3. Wait for the Moderator's report. If it returns a blocking question, relay it to the user and send the answer to the Moderator's thread.

Result

Relay the Moderator's report verbatim: at most three sentences that synthesize the six analyses rather than list them. Add one line naming the three categories and any agent that was unavailable. If the Moderator was unavailable or the workflow stopped, say which stage completed and what blocks the rest.