Codex Council 1.0 — review, forge, decide

Codex Council

One Codex pass gives you one train of thought. Forge shapes a proposal from creative specialists; Council pressure-tests it before you ship.

Install the plugin See a real run
  • Up to 6 reviewers
  • 5 Forge creators
  • Local estimate first
  • Browser evidence
  • No outside model APIs

The idea

Make Codex argue with itself, on purpose.

Each reviewer works from a fixed brief: architecture, reliability, security, product, red team, performance. Their answers are anonymized and ranked, a Chairman writes the verdict, and you see a local token estimate before anything runs. It's still one model wearing different hats, not rival labs. But the disagreement is real, and it catches what a single pass talks itself past.

New in 1.0: decision intelligence Sessions now keep a compact evidence trail you can inspect later. The experimental Decision Runtime adds a shadow view without changing the original verdict.
6 core review lenses
2 reviewer lanes in standard mode
1 browser evidence runner (Bob)
0 outside model APIs to wire up

How a run works

Four stages in a full Council run.

01

Preflight

Before anything runs you get the mode, the agent count, and a local heuristic estimate — never billing telemetry.

02

First opinions

Up to six named reviewers take the problem on independently, each guarding a different concern.

03

Review

Answers lose their authorship, get scored against a rubric, and are checked for gaps by dedicated reviewer lanes.

04

Synthesis

The Chairman writes the final call from what was actually saved. No invented test results, no made-up token counts.

The difference

A normal prompt answers. The council interrogates.

Normal prompt

Fine when you just need one solid plan.

Review this proposal: add shareable session reports
with automatic redaction, HTML/JSON export,
and team handoff links.

It gave a sensible local-export design and a couple of redaction tests. Useful, and roughly right.

Council prompt

Worth it when the plan has to survive privacy, reliability, and red-team scrutiny first.

Standard Council: review the same proposal.
Find blockers, dissent, verification gaps,
and the safest v1 scope.

Same prompt, sharper outcome: it cut v1 down, killed the unsafe sharing claims, demanded fail-closed redaction, and added performance gates.