Codex Discuss

You drive a structured, turn-based discussion with OpenAI Codex to converge on a plan before implementation. A second model that argues the other side is worth more than a second model that agrees with you — so the roles are deliberately adversarial.

Topic

$ARGUMENTS

Why two models

You review your own work with a positive bias and review others' work with a critical one. This skill exploits that: you propose, Codex challenges. Codex is instructed to find the weakest part of each proposal, not to validate it. Agreement only counts once the objections have actually been answered.

When to use this (and when not to)

Turn-based discussion is not the default — it costs several minutes per Codex turn plus the upkeep of a shared artifact. For most design questions a one-shot critique (a single codex exec over a brief, or /codex-review) delivers most of the value at a fraction of the cost. Reach for codex-discuss only when the decision has durable architecture impact, irreversible migration cost, or a live Blocker/Major disagreement a single round can't settle. Otherwise prefer the one-shot and stop there.

The shared file

One markdown file is the entire conversation. You are the only writer — Codex runs read-only, reads the file from disk, and returns its turn as text, which you append. This keeps a single source of truth and prevents write races.

Create it in the invoking repo so the reasoning trail can be committed (or ignored) as the user sees fit:

codex-discussions/<slug>.md      # slug = short kebab-case topic, e.g. caching-strategy

Template:

# Discussion: <Topic>

- **Started:** <date>
- **Participants:** Claude (proposer) · Codex (challenger)
- **Status:** IN_PROGRESS    <!-- IN_PROGRESS | CONVERGED | IMPASSE -->
- **Round:** 1 / <max>

## Goal
<the decision to be made, in one or two sentences>

## Constraints
- <hard requirements, non-negotiables, context>

## Repo context
- <relevant files you found and what each implies>
- <conventions / invariants the challenger must not break>

## Live state
<!-- The canonical current surface — refresh it every round so a stateless Codex
     reads the latest proposal and open objections without mining the transcript
     below. The `## Turn N` sections are the audit log, not the source of truth. -->
- **Current proposal:** <one-paragraph latest proposal>
- **Accepted constraints / risks:** <what both sides have accepted so far>

## Objections (ledger)
<!-- Stable IDs, appended as Codex raises them. Codex's closing turn must address
     every entry by ID. Format:
     - O1: <claim> — OPEN | ANSWERED | ACCEPTED_RISK | WONT_FIX — closure: <turn/file/check> -->
- <none yet>

---

## Turn 1 — Claude (proposer)
<the proposal and the reasoning behind it>

**Signal:** PROPOSE
**Questions for Codex:** <what you most want pressure-tested>

The loop (autonomous)

Run to a conclusion without pausing for the user. Each round:

1. Codex's turn. Capture it to a temp file, then append it verbatim under a ## Turn N — Codex (challenger) heading:

   codex exec --sandbox read-only --skip-git-repo-check --cd "$(pwd)" \
     --output-last-message /tmp/codex-turn.md \
     "You are the challenger in a design discussion. Read the discussion file at \
      codex-discussions/<slug>.md — start from its '## Live state' and any OPEN \
      '## Objections' — AND open the actual repo files it lists under \
      '## Repo context'; argue from the code itself, not from the proposer's \
      summary of it. Respond with ONLY your next turn as markdown: attack the \
      latest proposal — surface risks, edge cases, hidden costs, and a concrete \
      alternative where you have one. End with a line: \
      'Signal: AGREE | AGREE_WITH_CAVEATS | DISAGREE | NEEDS_INFO'. \
      On the first round you MUST raise at least one substantive objection — do not \
      agree yet. Do not modify any files."
   

   --cd "$(pwd)" runs Codex in the repo so it can read those files directly; --skip-git-repo-check avoids the trusted-directory refusal.

   Round 1 only — point the challenger at the code; don't summarize it for it. Codex runs read-only and starts cold. Write the ## Repo context section as an index of files to read — each path plus one line on why it matters and any invariant it must not break. This list is a pointer, not the evidence: the prompt above --cds Codex into the repo and tells it to open those files itself, so its objections are anchored to the actual code — including the negative space (an omitted invariant, a stale command, a sibling-skill mismatch) that a prose digest silently hides. Never let the digest stand in for the read.

   Once the ledger has entries, append to the prompt: For each OPEN objection in the ledger, reply ANSWERED / ACCEPTED_RISK / WONT_FIX with one line of closure evidence, or keep it OPEN and say what would close it. This is what makes convergence checkable instead of a prose vibe.

2. Your turn. Read Codex's turn. Address each objection honestly — concede what's right, defend what's wrong with reasons, revise the proposal. Update the ledger: give each new objection a stable ID, and set status + closure evidence on the ones you've answered. Append a ## Turn N+1 — Claude (proposer) heading ending in a **Signal:** line. Then refresh the ## Live state block (current proposal + accepted constraints + open O#s) so the next, stateless Codex turn reads the canonical surface rather than re-deriving it from — and relitigating — the whole transcript.

3. Check convergence (see below). If not converged and under the round cap, re-run Codex.

Convergence

• Maintain the ledger every round. Each new objection gets a stable ID (O1, O2, …); when addressed, set its status (ANSWERED | ACCEPTED_RISK | WONT_FIX) with one line of closure evidence — a turn, a file, or a passing check. An AGREED_FOLLOWUP (design settled, only implementation left) counts as closed for the plan when both sides name its acceptance test.
• CONVERGED only when the two most recent turns both signal AGREE or AGREE_WITH_CAVEATS and Codex's final turn has addressed every ledger entry by ID with no OPEN left. Convergence rests on the ledger, never on a prose vibe — and the proposer doesn't declare it alone: Codex must sign off on each closure.
• IMPASSE when the round cap is hit (default 5 round-trips) without convergence, or the same disagreement repeats two rounds running.

When the loop ends, fill in an ## Outcome section:

## Outcome
- **Result:** CONVERGED | IMPASSE
- **Decision:** <the agreed approach, or "no agreement">
- **Agreed plan:**
  1. <concrete, ordered steps both sides accept>
- **Caveats / risks accepted:** <from AGREE_WITH_CAVEATS>
- **Unresolved (impasse only):** <the specific points still in dispute, each
  side's position, and what would settle it>

Variations

• Codex proposes. When you want Codex's design instead, write the goal and ask Codex for Turn 1; you take the challenger role. Same loop, roles swapped.
• Deeper context. Add --search to let Codex consult external sources, or --model <model> to pick the reviewer model.

Reporting

After the loop, tell the user: the result (converged or impasse), the agreed plan (or the open disagreements with both positions), the number of rounds, and the path to the discussion file. Do not start implementing unless the user asks — converging on the plan is the deliverable.