Codex Review

Codex Review

Use Codex as an independent reviewer. Always read-only, always scoped, always machine-parseable when a harness consumes the result — and never relay a finding you haven't checked against the code.

Command surface verified against codex-cli 0.142.5; on flag errors, recheck codex --help and codex exec --help — placement and availability shift between versions.

When to use

• Independent second-opinion review of a diff, branch, commit, or PR
• Review workflows that route to gpt-5.6 alongside Claude reviewers
• Pre-merge risk sweeps (correctness, security, data loss, concurrency, migrations)

Preflight

codex --version       # confirm CLI present; stop and report if missing
git fetch origin main # when reviewing against a base

Review targets (codex exec review)

The dedicated subcommand supports exactly three targets in 0.142.5 — no --range, no --pr:

# Uncommitted work (staged + unstaged + untracked)
codex -a never -s read-only exec review \
  --uncommitted \
  "<scope prompt>" </dev/null 2>/dev/null

# Branch diff vs base
git fetch origin main
codex -a never -s read-only exec review \
  --base origin/main \
  "<scope prompt>" </dev/null 2>/dev/null

# Single commit (optional --title for context)
codex -a never -s read-only exec review \
  --commit abc1234 --title "auth refresh concurrency fix" \
  "<scope prompt>" </dev/null 2>/dev/null

codex exec review also accepts -m, -c, --json, --output-last-message, --output-schema, --ephemeral. Global flags (-a, --search) go before exec.

The review subcommand accepts no -s/--sandbox flag of its own (verified via codex exec review --help), which is why every example places -s globally, before exec.

Commit ranges and PRs — pipe a diff to plain exec

# Commit range
git diff --find-renames origin/main...HEAD \
  | codex -a never -s read-only exec \
      -m gpt-5.6-sol -c model_reasoning_effort='"high"' \
      "<scope prompt>" \
  > /tmp/range-review.md 2>/dev/null

# GitHub PR — either checkout + --base, or pipe the patch
gh pr checkout 123 && git fetch origin main
codex -a never -s read-only exec review --base origin/main "<scope prompt>" </dev/null 2>/dev/null

gh pr diff 123 --patch --color never \
  | codex -a never -s read-only exec "<scope prompt>" > /tmp/pr-123-review.md 2>/dev/null

Piped stdin is appended to the prompt as a <stdin> block. When piping, do not add </dev/null.

Scope prompt template

Unscoped reviews produce nits. State target, scope, format, and reliability rules:

Review target: branch diff against origin/main.

Scope:
- Correctness, security, data loss, concurrency, and migration risks only.
- No formatting, naming, or preference comments unless they hide a real bug.
- Ignore generated files and snapshots unless they prove a behavioral regression.

Finding format:
- severity: blocker | high | medium | low
- file:line
- title, why this is a bug, minimal fix direction

Reliability rules:
- Every finding must cite a concrete file:line from the diff or surrounding code.
- Do not speculate. If evidence is insufficient, omit the finding.
- If there are no findings, output {"findings":[]}.

Final message must be valid JSON only:
{"findings":[{"severity":"","file":"","line":0,"title":"","rationale":"","fix":""}]}

Narrow the scope line to specialize: security-only, migration-only, test-coverage-only.

Machine-parseable findings

Prompt-only JSON formatting drifts. For strict parsing, combine --output-schema with --output-last-message:

schema=$(mktemp /tmp/review-schema.XXXXXX.json)
printf '%s\n' '{
  "type": "object", "additionalProperties": false, "required": ["findings"],
  "properties": { "findings": { "type": "array", "items": {
    "type": "object", "additionalProperties": false,
    "required": ["severity", "file", "line", "title", "rationale", "fix"],
    "properties": {
      "severity": { "enum": ["blocker", "high", "medium", "low"] },
      "file": { "type": "string" }, "line": { "type": "integer", "minimum": 1 },
      "title": { "type": "string" }, "rationale": { "type": "string" }, "fix": { "type": "string" }
    } } } }
}' > "$schema"

codex -a never -s read-only exec review \
  --base origin/main \
  --output-schema "$schema" \
  --output-last-message /tmp/review-findings.json \
  "Review correctness and security only. Final message must match the schema." \
  </dev/null 2>/dev/null

Parse /tmp/review-findings.json, not stdout.

Model and effort

Default -m gpt-5.6-sol (plain gpt-5.6 is not a valid slug; -terra/-luna are the balanced/fast variants). Effort by stakes: medium for routine diffs, high for security-sensitive or concurrency-heavy changes, xhigh for high-risk pre-release audits. Reviews at high/xhigh can run 10–60+ minutes with no intermediate output — budget the harness timeout accordingly.

Relaying findings — verification duty

Before reporting any Codex finding:

1. Open the cited file:line and decide whether the finding is real.
2. Separate confirmed issues from unverified Codex suggestions in the report.
3. If Codex found nothing, say so explicitly and name the review target it inspected.
4. If codex is missing or exits nonzero, report the error and offer to review directly instead.

Follow-up on a finding reuses the session:

codex exec resume --last "Expand on finding 2 — show the exact failure interleaving." </dev/null 2>/dev/null

Troubleshooting