review

Codex Review

Delegate a read-only code review to a Codex subagent running GPT 5.6 Sol (or another GPT model the user names). The subagent reads code and returns findings; it does not write files.

Choosing the model

The -m flag selects the model. Default to gpt-5.6-sol, but honor any specific model the user asks for:

• Use gpt-5.6-sol unless the user names a different model.
• If the user specifies a model — e.g. "review with gpt-5.6-terra", "use gpt-5.5-codex", "with the <name> model" — pass that exact string to -m instead. Don't validate or second-guess the name; Codex will error if it's unknown.
• If they typed /codex:review --model <name> <task> (or -m <name>), strip that flag from the review instructions and use <name> as the model.

Choosing the effort

Reasoning effort is set with -c model_reasoning_effort="<level>". Default to high, but honor any level the user asks for:

• Use high unless the user names a different level.
• If the user asks in prose — e.g. "low effort", "medium effort" — substitute that level.
• If they typed /codex:review --effort <level> [focus], strip that flag from the review instructions and use <level> as the effort.

The command below shows -m gpt-5.6-sol and high effort; substitute the chosen model and effort.

How to invoke

Codex has a dedicated review subcommand. Pick the target based on what the user wants reviewed:

User asks to review	Flag
Working-tree changes (staged + unstaged + untracked)	--uncommitted
Everything on this branch vs main (or another branch)	--base main
A specific commit	--commit <SHA>
Nothing specified (defaults to current branch vs its merge base)	(no flag)

Command shape:

codex exec review \
  -m gpt-5.6-sol \
  -c model_reasoning_effort="high" \
  --skip-git-repo-check \
  -C "$PWD" \
  <target-flag> \
  "<REVIEW-PROMPT>"

• -m gpt-5.6-sol — the model; default gpt-5.6-sol, or the model the user named (see Choosing the model).
• -c model_reasoning_effort="high" — reasoning effort; default high, or the level the user named (see Choosing the effort).
• codex exec review (not plain codex review) — the exec form is non-interactive and prints to stdout.
• No -s / -a needed: review mode is inherently read-only.
• Always set an explicit Bash timeout. Reviews are slow and the Bash default (120000 ms / 2 min) will cut Codex off before it finishes. Pass timeout: 600000 (10 min — the maximum the Bash tool allows) on every codex exec review call. For a very large diff that may exceed 10 minutes, run the Bash call with run_in_background: true and poll instead, since a foreground call cannot exceed the 600000 ms cap.
• If the user did not specify a target, ask once — or default to --uncommitted if there are uncommitted changes, otherwise --base main (or the repo's default branch).