Codex Orchestration

How OpenAI Codex CLI discovers skills, fires hooks, and loads instructions. Use this skill instead of web-searching the docs each time.

Source of truth: developers.openai.com/codex/{skills,hooks,config-advanced,config-reference,guides/agents-md}. Verify against live docs if behavior seems off - orchestration features evolve.

When to Activate

• Install a skill and need the correct target directory
• Wire a Codex hook and need the event name, matcher, or payload shape
• Debug why an implicit skill isn't triggering on a matching prompt
• Author or review agents/openai.yaml for a skill
• Explain to a teammate why format-on-edit hooks don't work in Codex
• Migrate hook or skill conventions from Claude Code to Codex
• Add a notify program for turn-complete events
• Write or update this repo's README/hooks documentation

Discovery: Skills

Skills are folders containing SKILL.md (required) plus optional scripts/, references/, assets/, and agents/openai.yaml. Codex scans these paths (earlier wins on activation; duplicate names do NOT merge - both appear in selector):

Priority	Path	Scope
1	$CWD/.agents/skills	Current folder
2	$CWD/../.agents/skills	Parent folder
3	$REPO_ROOT/.agents/skills	Repo root
4	$HOME/.agents/skills	User
5	/etc/codex/skills	Admin
6	Bundled	System

Common mistake: copying skills to ~/.codex/skills/ - that path does nothing. Use ~/.agents/skills/.

SKILL.md Format

---
name: kebab-case-name
description: What it does, when to use it, AND when NOT to use it. Codex matches this for implicit invocation.
---

The body loads only after the skill activates. Front-load triggers and scope boundaries in description.

agents/openai.yaml (Optional)

Put it at skills/<name>/agents/openai.yaml:

interface:
  display_name: "User-facing name"
  short_description: "User-facing description"
  icon_small: "./assets/small-logo.svg"
  icon_large: "./assets/large-logo.png"
  brand_color: "#3B82F6"
  default_prompt: "Optional surrounding prompt"

policy:
  allow_implicit_invocation: true

dependencies:
  tools:
    - type: "mcp"
      value: "serverName"
      description: "Required MCP server"

Set allow_implicit_invocation: false for skills that should only fire on explicit /skills or $skillname invocation.

Activation Methods

Method	How
Implicit	Codex auto-selects based on description matching the user prompt
Explicit (menu)	User runs /skills
Explicit (inline)	User types $skillname in a prompt

Hooks (Experimental)

Enable with:

# ~/.codex/config.toml
[features]
codex_hooks = true

Config lives at ~/.codex/hooks.json or <repo>/.codex/hooks.json.

Events

Event	Fires	Matcher matches
SessionStart	Session start or resume	"startup" or "resume"
PreToolUse	Before tool execution	Tool name - Bash only currently
PostToolUse	After tool completion	Tool name - Bash only currently
UserPromptSubmit	User submits prompt	N/A - use "*" or omit
Stop	Turn concludes	N/A - use "*" or omit

Critical limit: PreToolUse/PostToolUse match Bash only. File-edit hooks (format-on-save) cannot fire through Codex hooks. Use pre-commit, husky, or a Makefile target.

hooks.json Schema

{
  "hooks": {
    "PreToolUse": [
      {
        "matcher": "Bash",
        "hooks": [
          {
            "type": "command",
            "command": "hooks/scripts/guard_force_push.sh",
            "statusMessage": "Checking force-push",
            "timeout": 10
          }
        ]
      }
    ]
  }
}

Stdin JSON Payload

Hook scripts receive JSON on stdin, not as $1. Common fields:

Field	Meaning
session_id	UUID
transcript_path	Path to the session transcript
cwd	Working directory
hook_event_name	Matches the event
model	Model name
turn_id	Current turn UUID
tool_input	Tool-specific - for Bash, has command

Parse with jq:

payload="$(cat)"
cmd="$(printf '%s' "$payload" | jq -r '.tool_input.command // empty')"

Exit Codes

Code	Meaning
0	Allow / continue
2	Block the tool call (PreToolUse)
Other non-zero	Error surfaced to operator

For SessionStart, UserPromptSubmit, Stop, scripts can emit a JSON object on stdout with: continue (bool), stopReason (string), systemMessage (string), suppressOutput (bool).

notify (Stable, simpler)

Separate from hooks.json. Lives in config.toml:

notify = ["python3", "/path/to/notify.py"]

Fires on agent-turn-complete only. Receives JSON as argv[1] (single string arg), NOT stdin. Fields: type, thread-id, turn-id, cwd, input-messages, last-assistant-message.

AGENTS.md Precedence

Codex builds an instruction chain from AGENTS.md files:

1. ~/.codex/AGENTS.override.md (if present, else ~/.codex/AGENTS.md)
2. Repo root AGENTS.md
3. Subdirectory AGENTS.md files closer to $CWD

@file references inside AGENTS.md resolve and inline, so AGENTS.md can pull in skill bodies directly:

## Domain Guidance
- Security audits: @.agents/skills/security/SKILL.md

BAD / GOOD: Installing a Skill

BAD

cp -r my-skill ~/.codex/skills/     # wrong directory; Codex ignores this

GOOD

mkdir -p ~/.agents/skills
cp -r my-skill ~/.agents/skills/    # user-scope discovery
# or
mkdir -p ./.agents/skills
cp -r my-skill ./.agents/skills/    # repo-scope discovery

BAD / GOOD: Writing a PreToolUse Hook

BAD

#!/usr/bin/env bash
# Assumes the command arrives as $1 - it doesn't.
if [[ "$1" == *"rm -rf"* ]]; then exit 2; fi

GOOD

#!/usr/bin/env bash
set -euo pipefail
payload="$(cat)"
cmd="$(printf '%s' "$payload" | jq -r '.tool_input.command // empty')"
if [[ "$cmd" == *"rm -rf /"* ]]; then
  echo "Blocking dangerous rm -rf /" >&2
  exit 2
fi
exit 0

BAD / GOOD: Format-on-edit

BAD

{
  "hooks": {
    "PostToolUse": [
      { "matcher": "Edit", "hooks": [{ "type": "command", "command": "prettier --write" }] }
    ]
  }
}

Edit is not a matchable tool name for PostToolUse today. The hook never fires.

GOOD

Move it to pre-commit:

repos:
  - repo: https://github.com/pre-commit/mirrors-prettier
    rev: v3.1.0
    hooks: [{ id: prettier }]

Or expose it via make fmt and document it in AGENTS.md under ## Environment Setup.

Codex vs Claude Code Differences

Concern	Codex	Claude Code
Skill install dir	.agents/skills/	.claude/skills/ or plugin packs
Skill frontmatter	name, description	name, description, allowed-tools, etc.
Hooks config	hooks.json (experimental)	settings.json hooks block (GA)
Pre/PostToolUse matchers	Bash only	All tools including Edit, Write, etc.
Hook input	stdin JSON	stdin JSON
Turn-complete hook	notify in config.toml	Stop hook
Instructions file	AGENTS.md	CLAUDE.md

Debugging Checklist

When a skill or hook isn't firing:

• Confirm the skill folder lives under one of the six .agents/skills paths. Duplicate names surface in the selector, they don't merge.
• Check description for triggering vocabulary. A vague description will miss implicit activation.
• For explicit invocation, run /skills and verify the skill is listed.
• For hooks, verify [features] codex_hooks = true is set and hooks.json parses as valid JSON.
• For PreToolUse/PostToolUse, confirm the matcher is Bash - Edit/Write tools won't fire it.
• Add a trace line to the hook script: echo "fired: $(date)" >> /tmp/codex-hook.log. If nothing appears, Codex is not invoking the hook.
• Tail the transcript path from the hook payload if you need to see what Codex saw.

Checklist

• Skills install under .agents/skills/ (not .codex/skills/)
• Every SKILL.md description declares scope boundary ("do NOT use for ...")
• agents/openai.yaml added when a skill needs implicit-invocation policy or MCP dependencies
• hooks.json lives at ~/.codex/hooks.json or <repo>/.codex/hooks.json
• Hook scripts parse JSON from stdin, not $1
• PreToolUse hooks use matcher "Bash" (no other tools supported yet)
• Format-on-edit is routed through pre-commit or make, not hooks.json
• [features] codex_hooks = true set in config.toml before expecting hooks to fire
• notify is used only for agent-turn-complete; payload arrives via argv[1]
• AGENTS.md uses @file imports to pull skill bodies into long-lived context when needed