Agent-Skills.md

Agent Skills: Codex Skill Guide

Use when the user asks to run Codex CLI (codex exec, codex resume) or references OpenAI Codex for code analysis, refactoring, or automated editing. Triggers on "run codex", "codex analyze", "use codex for", "codex refactor", "ask codex", or when user wants OpenAI's Codex model for code tasks.

UncategorizedID: buvis/home/use-codex

Repository

buvis/home
buvis
2

Install this agent skill to your local

pnpm dlx add-skill https://github.com/buvis/home/tree/HEAD/.claude/skills/use-codex

Skill Files

Browse the full folder contents for use-codex.

Download Skill

Loading file tree…

.claude/skills/use-codex/SKILL.md

Skill Metadata

Name
use-codex
Description
Use when running OpenAI Codex via the codex CLI (copilot CLI fallback) for code analysis, refactoring, or editing. Triggers on "run codex", "codex analyze", "ask codex", "copilot with codex".

Codex Skill Guide

Always run Codex through the codex-run.sh helper. The helper auto-detects its backend: it uses the native codex CLI when installed, and falls back to the copilot CLI otherwise. Always invoke the helper - never call codex or copilot directly - so backend detection and flag mapping stay correct.

  • codex backend (preferred): OpenAI ChatGPT subscription, no per-request billing multiplier. Uses codex's own configured default model unless -m is passed.
  • copilot backend (fallback): GitHub Copilot, billed with a per-request multiplier - see below.

Multiplier Policy (copilot backend only)

This applies only when the helper falls back to copilot. GitHub Copilot bills models with a per-request multiplier. The CLI does not expose multipliers locally, so the helper hardcodes a curated 1x default rather than auto-picking the highest version (which silently lands on premium tiers - gpt-5.5 once burned 25% of a monthly quota in one run).

  • copilot default: gpt-5.4 (1x). Always use unless the user has explicitly approved a different model.
  • Before passing -m to opt into a higher-multiplier model (e.g. gpt-5.5 for harder reasoning), confirm with the user via AskUserQuestion. State the model name and that it carries a higher Copilot multiplier.
  • Verify current multipliers in the GitHub Copilot dashboard if unsure - they change.
  • The codex backend has no multiplier, so -m on codex needs no quota confirmation.

Running a Task

Every run is non-interactive. By default each call is a fresh, one-shot Codex session; --resume-thread continues a prior codex session (codex backend only).

  1. Select the permission mode required for the task; default to no special flags (read-only) unless edits are necessary.
  2. Assemble the codex-run.sh command with appropriate helper flags:
    • -m, --model MODEL to override the model; on the copilot backend, ask the user first if the override is a higher-multiplier model
    • -f, --file FILE to pass the prompt (preferred - avoids shell escaping)
    • -a, --allow-tools to auto-approve tool use (codex: workspace-write sandbox)
    • -y, --yolo for full permissions (codex: bypass approvals and sandbox)
    • -d, --dir DIR to allow access to a specific directory (repeatable)
    • -o, --output FILE to capture output to a file
    • -s, --silent for clean scripting output (copilot backend only; ignored on codex)
    • --emit-thread-id FILE to capture the codex thread id from the JSON path (codex backend only; requires -o)
    • --resume-thread VALUE to resume a codex thread (VALUE is the thread id, or a file whose first line is the id; codex backend only; requires -o)
  3. Run the command, capture output, and summarize the outcome for the user.

Quick Reference (helper flags)

| Use case | Key flags | | --- | --- | | Read-only analysis | -f prompt.txt | | Auto-approve tools | -a -f prompt.txt | | Full auto (edits + tools) | -y -f prompt.txt | | Allow specific directory | -d <DIR> -f prompt.txt | | Scripting (clean output) | -s -f prompt.txt |

Background Dispatch and Waiting

A codex-run.sh call can run for many minutes. When you need to do other work while it runs, or you are inside an autopilot run:

  1. Dispatch the helper script with run_in_background: true. The dispatch tool result returns the task's output file path.
  2. Wait with the TaskOutput tool: TaskOutput(task_id, block=true, timeout=600000) (600000 ms = 10 min is the max per call). It returns when the task completes or at the deadline. It is the watchdog.
  3. On completion, Read the output file. On a timeout return, treat it as an infrastructure hang (see Error Handling); do not silently re-dispatch.

Never hand-roll a polling loop. Do not pass a while/if/wc -c stability loop to Monitor or Bash to detect completion. Such commands contain shell control flow that Warden cannot statically analyze, so they prompt for approval, which stalls an unattended autopilot run. The harness already notifies you when a background task finishes; TaskOutput is the only wait primitive you need.

Following Up

  • After every codex-run.sh command, use AskUserQuestion to confirm next steps.
  • By default each run is independent - a follow-up task is a new run with a new prompt, so restate the relevant context. The exception is --resume-thread (codex backend only), which continues a prior codex session so its context carries over.
  • Restate the permission mode when proposing follow-up actions.

Error Handling

  • Stop and report failures whenever a codex-run.sh command exits non-zero; request direction before retrying.
  • Before using high-impact flags (--allow-all, --yolo, --allow-all-paths) ask user permission via AskUserQuestion unless already given.
  • When output includes warnings or partial results, summarize them and ask how to adjust.

Helper Script

IMPORTANT: Always use -f with a temp file for prompts to avoid shell escaping issues.

# Write prompt to temp file, then run
echo 'Your prompt here (can contain "quotes", parens(), etc.)' > /tmp/codex-prompt.txt
~/.claude/skills/use-codex/scripts/codex-run.sh -f /tmp/codex-prompt.txt

# With auto-approve tools
~/.claude/skills/use-codex/scripts/codex-run.sh -a -f /tmp/codex-prompt.txt

# Override model (on copilot backend, only after user approval - higher multiplier may apply)
~/.claude/skills/use-codex/scripts/codex-run.sh -m gpt-5.5 -f /tmp/codex-prompt.txt

# Full permissions
~/.claude/skills/use-codex/scripts/codex-run.sh -y -f /tmp/codex-prompt.txt

# Capture output to file
~/.claude/skills/use-codex/scripts/codex-run.sh -a -o /tmp/result.txt -f /tmp/codex-prompt.txt

# Silent mode for scripting
~/.claude/skills/use-codex/scripts/codex-run.sh -s -f /tmp/codex-prompt.txt

Run ~/.claude/skills/use-codex/scripts/codex-run.sh --help for all options.