codex

name	codex
description	Use when the user asks to run Codex CLI (codex exec, codex resume) or references OpenAI Codex for code analysis, refactoring, or automated editing
context	fork
agent	general-purpose

Codex Skill Guide

Running a Task

1. Ask the user (via AskUserQuestion) for reasoning effort (xhigh, high, medium, or low).
2. Select sandbox mode for the task; default to --sandbox read-only unless edits or network access needed.
3. Assemble the command:
   • -m, --model <MODEL>
   • --config model_reasoning_effort="<xhigh|high|medium|low>"
   • --sandbox <read-only|workspace-write|danger-full-access>
   • --full-auto
   • -C, --cd <DIR>
   • --skip-git-repo-check
4. Always use --skip-git-repo-check.
5. When continuing a previous session: echo "your prompt" | codex exec --skip-git-repo-check resume --last 2>/dev/null
6. Always append 2>/dev/null to suppress thinking tokens unless user asks for them.
7. Run command, capture output.

Quick Reference

Use case	Sandbox	Key flags
Read-only review	read-only	--sandbox read-only 2>/dev/null
Apply local edits	workspace-write	--sandbox workspace-write --full-auto 2>/dev/null
Network/broad access	danger-full-access	--sandbox danger-full-access --full-auto 2>/dev/null
Resume session	Inherited	echo "prompt" | codex exec --skip-git-repo-check resume --last 2>/dev/null
Run from other dir	Match task	-C <DIR> plus other flags 2>/dev/null

Return Output to Main Agent

After codex completes, collect and return a structured summary:

git diff --stat        # changed files
git diff --name-only   # file list

Return format:

## Codex Result

**Status**: success | partial | failed
**Changed files**:
- path/to/file (added|modified|deleted)

**Summary**: [1-3 sentences of what codex did]

**Output** (truncated if long):
[relevant stdout]

Following Up

• After every codex command, use AskUserQuestion to confirm next steps or decide whether to resume.
• When resuming: echo "new prompt" | codex exec resume --last 2>/dev/null

Tapping Into a Previous Session

When the user wants to read what codex explored/decided in a prior session, extract the transcript:

# Find latest session (or nth-latest with tail -N | head -1)
LATEST=$(ls -t ~/.codex/sessions/*/*/*.jsonl 2>/dev/null | head -1)
echo "Session: $LATEST"
python3 - "$LATEST" <<'EOF'
import json, sys, textwrap

path = sys.argv[1]
with open(path) as f:
    items = [json.loads(l) for l in f]

for item in items:
    t = item.get("type")
    p = item.get("payload", item)

    if t == "response_item":
        kind = p.get("type")
        role = p.get("role")

        if kind == "message" and role == "assistant":
            for c in p.get("content", []):
                if c.get("type") == "output_text" and c.get("text"):
                    print(f"\n[assistant]\n{textwrap.fill(c['text'], 100)}")

        elif kind == "function_call":
            args = p.get("arguments", "{}")
            try:
                cmd = json.loads(args).get("cmd", args)
            except Exception:
                cmd = args
            print(f"\n[exec] {cmd[:200]}")

        elif kind == "function_call_output":
            for out in p.get("output", []):
                if out.get("type") == "text":
                    print(f"[out] {out['text'][:300]}")
EOF

To narrow to a specific session by date: ls ~/.codex/sessions/2026/02/26/*.jsonl

After reading, use Read on the session file directly for more depth, or feed a summary to codex with:

echo "Context from last session: <summary>. Now: <new task>" | codex exec --skip-git-repo-check resume --last 2>/dev/null

Critical Evaluation of Codex Output

Codex is powered by OpenAI models with their own knowledge cutoffs. Treat as a colleague, not an authority.

• Trust your own knowledge when confident — push back if Codex is wrong.
• Research disagreements via WebSearch before accepting Codex's claims.
• Knowledge cutoffs — Codex may not know recent releases or API changes.
• When disagreeing, identify yourself as Claude and your current model name:

  echo "This is Claude (<model>) following up. I disagree with [X] because [evidence]." | codex exec --skip-git-repo-check resume --last 2>/dev/null
  

Error Handling

• Stop and report on non-zero exit from codex --version or codex exec.
• Ask permission before high-impact flags (--full-auto, --sandbox danger-full-access) unless already given.
• Summarize warnings/partial results and ask how to adjust.

Model Override

Always use codex-mini-latest. Do not ask the user which model to use.