| name | codex |
| description | Delegate well-scoped coding, analysis, and review tasks to the OpenAI Codex CLI (gpt-5.5) as a cross-model collaborator. Use for implementation chunks you want a second model to write or cross-check, independent second-opinion review of a design or diff, and applying Codex-authored changes. Triggers: "ask codex", "have codex implement/review X", "second opinion from codex", "cross-check with codex".
|
Codex — cross-model delegation
codex (OpenAI Codex CLI, model gpt-5.5, approval_policy = never so it never blocks
on prompts) is a second coding agent that reads this repo and reasons over it precisely
(verified). Consult it for a different model's perspective — it catches what a single
model misses.
When to consult Codex
- A well-scoped implementation chunk you want a second model to write, or to cross-check
your own implementation against.
- An independent design/diff review — different model, different failure modes caught.
- Adversarial verification of a claim ("is X actually true in the shipped code?").
How — always non-interactive codex exec
| Goal | Command |
|---|
| Analyze / review (read-only, safe — the default) | codex exec --sandbox read-only "<tight prompt>" 2>/dev/null |
| Let Codex EDIT files | codex exec --sandbox workspace-write "<prompt>" 2>/dev/null |
| Independent review of the working tree | codex review |
| Apply Codex's last proposed diff | codex apply |
| Cheaper/deeper | add -c model_reasoning_effort=low|medium|high|xhigh |
| Capture output | append -o out.txt, or --json for structured events |
| Continue a session | codex exec resume --last "<follow-up>" |
Discipline (non-negotiable)
- Scope tightly + hand it concrete material — point Codex at exact
file:line, the
locked contract (tasks/v11-locked-contract.md), or the spec. It works best on
concrete input, not vague asks.
2>/dev/null to suppress thinking tokens unless you're debugging the run.- Verify every workspace-write run with the repo's own gates —
go -C cli build ./...,
go -C cli test ./..., c3local check. Codex's edits are a proposal, never ground
truth; treat its diff like an untrusted PR.
- For high-stakes work, run it twice across models — have Codex and a Claude subagent
do the same task independently, then reconcile the diff. Cross-model disagreement is
signal.
