Codex

/codex

Invokes the Codex CLI as an independent reviewer. Codex sees the same files but not this session's prior context — that disconnect is the point. Use when you want a second opinion uncontaminated by the framing already in this conversation.

The skill itself runs in Claude Code (which is why cli_support: [claude-code] is single-entry). Codex is invoked as a subprocess; its output is parsed and surfaced back here.

When to use

• Pre-/ship outside-voice review on a non-trivial diff
• /investigate produced a confident root cause and you want it independently challenged
• A design decision was made in conversation and you want a fresh-eyes critique
• /plan-eng-review cleared with caveats — use Codex to triangulate the caveats

When NOT to use

• The work is trivial — Codex tokens cost real money
• You've already disagreed with Codex twice on this same diff — escalate to a human reviewer instead
• Codex CLI isn't installed — surface that as a setup gap, do not silently fall back to "Claude reviewing itself"

Inputs

• Required: target (one of --diff, --plan <file>, --code <file>, --hypothesis "<text>")
• Optional --prompt-style <strict|exploratory> — strict = "find what's wrong", exploratory = "what would you do differently"
• Optional --budget <tokens> — cap Codex call (default: 50k tokens)

Workflow

1. Preflight. Verify codex is on PATH. If not: report missing setup + exit. No fallback.
2. Construct prompt. Inject the target (diff, plan body, code excerpt, hypothesis statement). Wrap in a directive: "You are reviewing this independently. The author has their own reasoning; surface what you'd push back on."
3. Run Codex. codex exec --quiet --output json --prompt-file <tmp> (or equivalent — adjust to current Codex CLI flags). Capture stdout, stderr, exit code.
4. Parse output. Codex returns structured findings (severity, location, claim, evidence). Normalize to Lintel's P1/P2/P3 severity.
5. Compare to local reasoning. If invoked mid-/investigate or post-/review: explicitly diff Codex's findings against what was already concluded. Surface AGREEMENT and DISAGREEMENT separately.
6. Persist via bin/li-review-log with skill: codex so downstream /ship can read.
7. Report.

Report format

Codex Review: <target>

Codex CLI: codex@1.4.2 (exec mode, exploratory style, budget 50k)
Tokens used: 31,400 / 50,000

## Findings from Codex (3)

[P1] (Codex confidence: 9/10) src/lib/dlxClient.ts:87
   "The retry logic creates an AbortController but never resets between attempts. Second retry will abort itself."
   ← AGREES with /investigate root cause from this session.

[P2] (Codex confidence: 7/10) src/lib/dlxClient.ts:42
   "Error handling swallows the original stack trace. Hard to debug downstream."
   ← NEW finding, not raised in local review.

[P3] (Codex confidence: 6/10) src/lib/dlxClient.ts:120
   "Magic number 30000 should be a named constant."
   ← Style nit, agree.

## Synthesis
- 1 finding agrees with local conclusion (strengthens confidence)
- 1 new finding (P2) — recommend addressing before /ship
- 1 style nit (P3) — defer or include in same diff

Compliance integration

• Codex sees code. Per Layer 2: code is not customer-data, so no auth gate. But: if the target includes a fixture path containing customer-data patterns, BLOCK the Codex call and surface (Codex would receive that data).
• Token spend logged to .claude/runtime/audit/codex-spend.jsonl.
• The active pack may route Codex calls via a configured gateway per ~/.lintel/config.yaml (none by default).

Failure modes

• Codex CLI missing: report + exit. No fallback to Claude self-review (that defeats the purpose).
• Codex returns malformed output: capture raw stdout to .claude/runtime/audit/codex-raw-<ts>.txt, report parse failure, exit.
• Budget exceeded mid-call: Codex's own truncation kicks in. Report partial findings + budget overflow.
• Codex output contains a finding referencing a file we never sent: flag as hallucination, suppress from main report, log to debug appendix.

Examples

Outside opinion on a diff:

> /codex --diff
Codex found 2 P2 findings the local review missed. Recommend addressing before /ship.

Challenge a hypothesis:

> /codex --hypothesis "retry abort controller bug at dlxClient:87"
Codex agrees (9/10 confidence). Strengthens the root cause.

Style-only review:

> /codex --code src/components/case/AgentModeChip.tsx --prompt-style exploratory
3 stylistic suggestions, no correctness issues.

See also

• /review — local diff review (use /codex as a follow-up for important diffs)
• /investigate — --with-codex flag chains automatically
• /plan-eng-review — uses /codex internally as the outside-voice step
• /ship — reads codex review-log entries as part of clearance check