Codex Skill Guide

Configuration

Read model and sandbox settings from ~/.claude/skills/codex/codex-config.json:

{
  "model": "gpt-5.3-codex",
  "sandbox": "danger-full-access",
  "suppress_stderr": true
}

User can edit this file anytime to switch models. Do NOT ask the user to choose model or sandbox.

Dynamic Effort Routing

Priority: User explicitly specifies effort → use that. Otherwise → auto-select.

Source tag convention: The effort argument includes a source tag:

• •effort=low(user) → User explicitly requested this effort level. Say "用户指定 effort=low".
• •effort=low(auto) → Claude Code auto-selected based on task complexity. Say "自动选择 effort=low（任务较简单）".
• •No effort in args → Auto-select here and mark as auto in output.

CRITICAL: Never say "用户指定" when the source tag is (auto). This is a context boundary — the args were written by pre-skill Claude Code, not the user.

Running a Task

1. •Read config from codex-config.json for model and sandbox.
2. •Auto-select reasoning effort based on task complexity (see table above), unless user specified one.
3. •Assemble the command:
   • •-m <MODEL> (from config)
   • •--config model_reasoning_effort="<effort>" (auto-selected or user-specified)
   • •-s <SANDBOX> (from config)
   • •-C <DIR> (current working directory)
   • •--skip-git-repo-check
   • •CRITICAL: Do NOT use --full-auto with -s danger-full-access. --full-auto forces --sandbox workspace-write, overriding explicit --sandbox. If the user has approval_policy = "never" in config.toml, auto-approval works without --full-auto. Otherwise, users may need to approve actions manually.
4. •Append 2>/dev/null to suppress thinking tokens (stderr), unless user requests to see them.
5. •Run the command, capture output, and summarize the outcome in plain language.
6. •After Codex completes, inform the user: "You can resume this Codex session at any time by saying 'codex resume'."

Command Template

Note: examples here use bash-style shell syntax. Claude Code on Windows runs commands through its built-in shell, which supports this syntax.

codex exec -m <MODEL> --config model_reasoning_effort="<EFFORT>" \
  -s <SANDBOX> -C "<WORKING_DIR>" \
  --skip-git-repo-check "<PROMPT>" 2>/dev/null

Quick Reference

Resuming Sessions

• •Resume syntax: echo "your prompt here" | codex exec --skip-git-repo-check resume --last 2>/dev/null
• •All flags go between exec and resume.
• •Don't use config flags when resuming unless user explicitly requests them.
• •Resumed sessions inherit original model, effort, and sandbox.

Following Up

• •After Codex completes, summarize results in plain language for the user.
• •If follow-up work is needed, ask the user before resuming.
• •When resuming, pipe the new prompt via stdin.

Critical Evaluation of Codex Output

Codex is powered by OpenAI models. Treat Codex as a colleague, not an authority.

Guidelines

• •Trust your own knowledge when confident. Push back if Codex is wrong.
• •Research disagreements using WebSearch before accepting Codex's claims.
• •Remember knowledge cutoffs - Codex may not know about recent changes.
• •Don't defer blindly - evaluate suggestions critically.

When Codex is Wrong

1. •State disagreement clearly to the user
2. •Provide evidence (your knowledge, web search, docs)
3. •Optionally resume to discuss:
   bash

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

4. •Let the user decide if there's genuine ambiguity

Error Handling

• •Stop and report failures when codex exec exits non-zero.
• •-s danger-full-access is the default and does NOT need user confirmation.
• •Summarize warnings and ask how to adjust.