Codex Consult

The automatic dual-model layer. Fires on its own at every major decision point — there is no command to invoke, and no way to forget.

Overview

The Codex page explains that a second model is consulted at every decision point, and why. This skill is how.

A command would put the burden on you at exactly the moments you're least able to carry it — mid-design, mid-diagnosis, head down in the problem. So there is no command.

Step 0 — Should this consult happen?

It fires at eight points and nowhere else:

Not at Q0/Q1 — those are questions for you, not for a model — nor on mode selection, reads, renames, formatting, trivial diffs, or anything in scratch/.

Step 1 — Announce, before dispatching

◆ Consulting Codex — Gate 2 (design): "Propose the strongest alternative
  to the two-stage encoder and name the failure mode this design doesn't
  address."

One line: the decision point, and the actual question being sent. You should be able to tell from this line alone whether it was a good question. It is not optional and it is not retroactive — if you only learn Codex was involved when the findings appear, the announcement has failed.

Step 2 — The brief

Five to fifteen lines. Facts first, position second, question last.

CONTEXT
  Project:   <one line>
  Decision:  <what is being decided right now>
  Material:  <the design / diff / diagnosis, inlined>

CLAUDE'S POSITION
  <2-4 lines — omitted when Claude wrote the thing being reviewed>

QUESTION
  <one sharp, falsifiable question>

CONSTRAINTS
  Under 400 words. Cite file:line for every code claim. If you have no
  additional finding, say "no additional findings" rather than restating
  mine.

Brief-writing rules

• Don't bias the question. "Does this look right?" gets agreement. "Name the strongest counter-approach" gets a second opinion.
• Omit Claude's position on self-authored material — a Gate 3 diff, or a design Claude proposed. Anchoring Codex to the author's reading of their own work throws away most of the value.
• Inline the material. Codex doesn't share Claude's context, working memory, registry, or CLAUDE.md. A brief referencing "the design we discussed" produces a hallucinated review of a design Codex has never seen.
• Redact. Keys, tokens, unpublished data. If the brief can't be written without a secret, skip the consult and say why.

Step 3 — Dispatch through codex-bridge

Never run the CLI from the main context. The bridge subagent runs it, checks each claim against the repo, throws away the transcript, and returns a VERIFIED / DISPUTED / UNVERIFIED table.

The reason is context hygiene, and it's load-bearing: Codex's raw reply is long, confident, and partly wrong. In the main conversation it becomes background the model reasons next to for the rest of the session.

Step 4 — Judgment on top of verification

The bridge does the mechanical verification. Claude does the judgment.

For every VERIFIED finding: is it real? A confirmed line number is not a confirmed bug — Codex will correctly quote a line and draw the wrong conclusion from it. For every DISPUTED finding: was Codex wrong, or right about something in the wrong place?

Step 5 — The card

┌─ ◆ Codex consult — Gate 2 (design) ─────────────────┐
│ Asked: Propose the strongest alternative and name a failure  │
│        mode this design doesn't address.                     │
│                                                              │
│ Both models agree:                                           │
│   • [both] The encoder must be frozen before stage 2         │
│            — weak evidence, two models agreeing is cheap     │
│                                                              │
│ They disagree:                                               │
│   • Codebook init. [claude] k-means per §3.2. [codex]        │
│     uniform, argues k-means leaks val statistics.            │
│     Claude's read: codex is right about the leak, wrong      │
│     that §3.2 requires uniform. This is a real decision.     │
│                                                              │
│ New from Codex:                                              │
│   • [codex] EMA decay applied per-step, paper applies it     │
│     per-epoch — vq.py:88 (verified)      severity: high      │
│                                                              │
│ Codex claimed, Claude could not confirm:                     │
│   • [codex · unverified] "loss.py:210 detaches the codebook" │
│     — loss.py has 140 lines; no detach found anywhere.       │
│                                                              │
│ Your decision:                                               │
│   The codebook init question is the one that changes         │
│   results. Uniform, or k-means on a train-only sample?       │
└─────────────────────────────────────────────────────┘

Then the gate proceeds as normal. The card is input to the gate, not a replacement for it. Codex is not a fifth reviewer whose approval advances the pipeline — you still answer the gate question.

Controls

If you ask for a second opinion mid-turn — not at a gate — Propel just runs the consult. Same procedure; the schedule is a floor, not a ceiling.

See also: Codex — Two Models, codex-bridge, think-deeply (the internal anti-sycophancy pass), and c-review, which folds this consult into a merged Gate 3 card.