codex

/codex — Codex CLI Integration

Delegate tasks to OpenAI Codex CLI from within Claude Code.

Quick Reference

User says	Resolves to
/codex review	codex review --uncommitted
/codex review --base main	codex review --base main
/codex review with o3 effort: high	codex review --uncommitted -m o3 -c model_reasoning_effort="high"
/codex security	codex review --uncommitted -c model_reasoning_effort="high" "Focus on security: injection, auth bypass, data exposure, OWASP Top 10"
/codex perf --base main	codex review --base main -c model_reasoning_effort="high" "Focus on performance: N+1 queries, unnecessary allocations, missing indexes, expensive loops"
/codex quick explain this error	codex exec --full-auto -c model_reasoning_effort="low" "explain this error"
/codex thorough refactor auth	codex exec --full-auto -c model_reasoning_effort="xhigh" "refactor auth"
/codex fix	Capture git diff + git diff --cached → codex exec --full-auto "Review and fix issues in this diff: ..."
/codex fix the broken test	codex exec --full-auto "fix the broken test"
/codex second-opinion	Rerun last user prompt through Codex, present both answers
/codex <anything else>	codex exec --full-auto "<prompt>"

Routing

digraph routing {
  "Parse input" [shape=ellipse];
  "Preset?" [shape=diamond];
  "second-opinion?" [shape=diamond];
  "Bare fix?" [shape=diamond];
  "Review keyword?" [shape=diamond];
  "Expand preset" [shape=box];
  "Second opinion flow" [shape=box];
  "Diff-scoped fix" [shape=box];
  "codex review" [shape=box];
  "codex exec" [shape=box];

  "Parse input" -> "Preset?";
  "Preset?" -> "Expand preset" [label="yes"];
  "Preset?" -> "second-opinion?" [label="no"];
  "second-opinion?" -> "Second opinion flow" [label="yes"];
  "second-opinion?" -> "Bare fix?" [label="no"];
  "Bare fix?" -> "Diff-scoped fix" [label="yes"];
  "Bare fix?" -> "Review keyword?" [label="no"];
  "Review keyword?" -> "codex review" [label="yes"];
  "Review keyword?" -> "codex exec" [label="no"];
}

Review keywords: review, code review, check changes, audit, inspect, critique Default: Everything else routes to codex exec --full-auto

After routing, for all commands: extract model/effort flags → inject project context (review + presets only) → execute via Bash → summarize results.

Presets

Detect as first token after /codex. Additional text appends to the prompt. User can override model/effort.

Preset	Mode	Effort	Focus prompt
security	review	high	Security: injection, auth bypass, data exposure, OWASP Top 10
perf	review	high	Performance: N+1 queries, unnecessary allocations, missing indexes, expensive loops
quick	exec	low	(user's remaining prompt)
thorough	exec	xhigh	(user's remaining prompt)

Second Opinion

Trigger: /codex second-opinion or /codex second opinion

1. Find the user's last substantive prompt in conversation (skip meta-commands)
2. Run: codex exec --full-auto "<original prompt>"
3. Present: Claude's answer (summary) → Codex's answer → Key differences (bullets)

Append custom focus: /codex second-opinion focus on error handling

Diff-Scoped Fix

Trigger: /codex fix with NO additional prompt text

1. Capture git diff + git diff --cached — if empty, stop with "No uncommitted changes"
2. Run: codex exec --full-auto "Review the following diff and fix any bugs, logic errors, or code quality issues:\n\n<diff>"
3. Summarize what Codex changed

/codex fix <something> (with a prompt) routes to normal exec mode instead.

Model & Reasoning Effort

Extract model/effort from user input before building the command. Strip from the prompt.

Input pattern	Flag produced
-m o3 / --model o3 / with o3 / using o3	-m o3
--effort high / --thinking high / effort: high	-c model_reasoning_effort="high"

Valid levels: low, medium, high, xhigh Default: Inherit from ~/.codex/config.toml

Project Context Injection

Applies to: review commands and preset commands only (not plain exec).

Read CLAUDE.md from project root. Extract key conventions (~200 words max) and prepend:

Project conventions: <framework>, <DB>, <patterns>.
<extracted rules>
---
<original prompt>

Focus on rules Codex would otherwise violate: architecture layers, DB conventions, code style.

Rules

• Quote all prompt arguments (double quotes only)
• --uncommitted and --base are mutually exclusive — default to --uncommitted
• Pass through all explicit CLI flags (-m, --base, -c, etc.) verbatim
• Only inject project context for review + presets, not generic exec
• Second opinion requires a prior substantive prompt in conversation