Claude-skill-registry codex-cli-subagent

This skill lets you spawn a nested Codex process (a “subagent”) using

codex exec

so it can work with the same project and AGENTS.md instructions.

filename="$(openssl rand -hex 4)"
codex exec "count the total number of lines of code in this project" 2>>/tmp/${filename}.log

In non-interactive mode,

codex exec

runs without a TUI and follows the configured approval/sandbox policy (flags and/or

~/.codex/config.toml

). By default it runs in

read-only

mode, so it cannot edit files or run commands that require network access.

When you use this skill, follow these logging rules:

• Before running

  codex exec

  , generate a unique log file name with

  openssl rand -hex 4

  .
• By default, append

  2>>/tmp/${filename}.log

  to your

  codex exec

  command so only the subagent’s final message is visible to the caller, while stderr is captured for debugging.
• If a run fails, behaves unexpectedly, or the user explicitly asks to see what the inner agent is doing, read the log file (for example:

  cat /tmp/${filename}.log

  ).

Since the

codex exec

may run couple hours, set a generous timeout so long-running subagent work can complete:

• Use a timeout of at least 60 minutes for the

  Run

  (or

  Bash

  ) tool unless the user explicitly requests other limit.

Use this skill when you want a focused helper agent (for refactors, audits, scripted operations, or scans) while keeping your main session and context intact.

Use

codex exec --full-auto

to allow file edits. Use

codex exec --sandbox danger-full-access

to allow edits and networked commands, but only when the user clearly permits this level of access.

Practical gotchas

• Prefer stdin for long prompts to avoid shell quoting/globbing issues:

  codex exec - < /tmp/prompt.md

  .
• If you need built-in web search, pass

  --search

  (otherwise assume “no web tool” and use

  curl

  /repo sources).
• Don’t run nested subagents (avoid calling

  codex exec

  from inside a

  codex exec

  session).
• Consider

  --output-last-message /tmp/${filename}.out

  to capture the final answer without scraping logs.

Key Flags

• --full-auto

  : Unattended operation with workspace-write sandbox

• --dangerously-bypass-approvals-and-sandbox

  or

  --yolo

  : Complete hands-off mode (use carefully)

• --cd <path>

  : Set working directory

• --model <model>

  or

  -m

  : Specify model (e.g.,

  -m gpt-5.1-codex-max

  )

• --sandbox

  : Sandbox types:

  • read-only

    : No file edits, no networked commands

  • workspace-write

    : Can edit files in the workspace

  • danger-full-access

    : No sandboxing; full access (use with care)

• --search

  : Enable built-in web search tool

• --config

  : Pass config variables:

  • model_reasoning_effort

    : Model reasoning effort:

    low

    ,

    medium

    ,

    high

    ;

Examples

# As a subagent, perform automated refactoring
filename="$(openssl rand -hex 4)"
codex exec --full-auto "Update all README links to HTTPS" 2>>/tmp/${filename}.log

# Run in a specific project directory
filename="$(openssl rand -hex 4)"
codex exec --cd /path/to/project "Fix failing tests" 2>>/tmp/${filename}.log

# Use AGENTS.md context for a focused refactor
filename="$(openssl rand -hex 4)"
codex exec --cd /path/to/project "Using this repo's AGENTS.md instructions, refactor the test helpers for clarity and consistency" 2>>/tmp/${filename}.log

filename="$(openssl rand -hex 4)"
codex exec --model gpt-5.1-codex-max --sandbox workspace-write --config model_reasoning_effort=medium - < /tmp/some-big-prompt.md 2>>/tmp/${filename}.log

# For audits / deep analysis
filename="$(openssl rand -hex 4)"
codex exec --model gpt-5.2 --sandbox read-only --config model_reasoning_effort=high "Audit this repo for security issues" 2>>/tmp/${filename}.log

Input Methods

# Pipe prompt from file
filename="$(openssl rand -hex 4)"
codex exec - < prompt.txt 2>>/tmp/${filename}.log

# Standard input
filename="$(openssl rand -hex 4)"
echo "Review this code" | codex exec - 2>>/tmp/${filename}.log

When this skill is invoked, you should decide:

• Whether a subagent is appropriate (e.g., long-running refactor, scan, or analysis).
• Which flags, sandbox level, and model to use based on user intent and risk.
• For code changes: default to

  --model gpt-5.1-codex-max --config model_reasoning_effort=medium

  unless the user explicitly requests otherwise.
• For audits: default to

  --model gpt-5.2 --config model_reasoning_effort=high

  unless the user explicitly requests otherwise.
• Whether to keep logs suppressed (default:

  2>>/tmp/${filename}.log

  or read/tail

  /tmp/${filename}.log

  )