Codex Orchestrator

Spawn specialized Codex CLI subagents for focused development tasks. Each profile becomes process-local developer instructions; the repository's global, root, and nested AGENTS.md files remain untouched and continue to apply.

Architecture

Claude Code (orchestrator)
    ↓ invokes skill
codex-orchestrator scripts
    ↓ spawns via Bash
Codex CLI with process-local persona + project AGENTS.md chain
    ↓ executes
Specialized subagent task

Prerequisites

Verify Codex CLI is installed and configured:

~/.claude/skills/codex-orchestrator/scripts/codex-status.sh

Required:

• Codex CLI: npm install -g @openai/codex
• API Key: export OPENAI_API_KEY=sk-...

Auto-Update

The skill automatically checks for Codex CLI updates on each invocation and updates if needed. This prevents issues caused by outdated CLI versions.

To manually check/update:

# Check version only
~/.claude/skills/codex-orchestrator/scripts/codex-version-check.sh

# Check and auto-update if needed
~/.claude/skills/codex-orchestrator/scripts/codex-version-check.sh --auto-update

Available Profiles

Quick Execution

Execute a one-shot task with a specific profile:

~/.claude/skills/codex-orchestrator/scripts/codex-exec.sh <profile> "<prompt>"

Examples:

# Code review
~/.claude/skills/codex-orchestrator/scripts/codex-exec.sh reviewer "Review src/auth.ts for security issues"

# Debug investigation
~/.claude/skills/codex-orchestrator/scripts/codex-exec.sh debugger "Debug the login timeout on slow networks"

# Architecture design
~/.claude/skills/codex-orchestrator/scripts/codex-exec.sh architect "Design a caching layer for the API"

# Security audit
~/.claude/skills/codex-orchestrator/scripts/codex-exec.sh security "Audit the payment module for vulnerabilities"

# Write profiles auto-approve by default (uses -a never + --sandbox workspace-write)
~/.claude/skills/codex-orchestrator/scripts/codex-exec.sh reviewer "Fix all lint errors"

# Create execution plan for complex feature
~/.claude/skills/codex-orchestrator/scripts/codex-exec.sh planner "Create an ExecPlan for adding WebSocket support"

# Build new feature from spec
~/.claude/skills/codex-orchestrator/scripts/codex-exec.sh builder "Implement user authentication with JWT"

# Continue from previous builder session
~/.claude/skills/codex-orchestrator/scripts/codex-exec.sh builder "continue"

# Open-ended conversation (read-only, ephemeral, gpt-5.6-terra via subscription)
~/.claude/skills/codex-orchestrator/scripts/codex-exec.sh chat "What are the tradeoffs of event sourcing vs CRUD?"

# Conversation via API billing (gpt-5.6-sol, bypasses Codex subscription)
~/.claude/skills/codex-orchestrator/scripts/codex-exec.sh chat "Explain quantum error correction" --api --model gpt-5.6-sol

# Multi-turn API chat with session file
~/.claude/skills/codex-orchestrator/scripts/codex-exec.sh chat "What is the tallest mountain?" --api --session /tmp/chat.json
~/.claude/skills/codex-orchestrator/scripts/codex-exec.sh chat "And the deepest ocean?" --api --session /tmp/chat.json

# Streaming API response
~/.claude/skills/codex-orchestrator/scripts/codex-exec.sh chat "Tell me about CQRS" --api --model gpt-5.6-sol --stream

# Ask a question about the codebase (read-only, no file changes)
~/.claude/skills/codex-orchestrator/scripts/codex-exec.sh researcher "Explain the authentication flow in this project"

# Weigh rival hypotheses or ambiguous evidence (read-only, high reasoning)
~/.claude/skills/codex-orchestrator/scripts/codex-exec.sh adjudicator "With Linear B and Linear A inscriptions side-by-side, weigh candidate values for the missing sound and rank the hypotheses."

# Research with Exa web search (appends Exa guide to this process's persona)
~/.claude/skills/codex-orchestrator/scripts/codex-exec.sh researcher "What are the latest React Server Component patterns?" --web-search

# Research with native Codex web search (model-level tool, works in all sandboxes)
~/.claude/skills/codex-orchestrator/scripts/codex-exec.sh researcher "What are the latest React patterns?" --search

# Review a screenshot (vision input)
~/.claude/skills/codex-orchestrator/scripts/codex-exec.sh reviewer "Review this mockup for UX issues" --image screenshot.png

# Resume previous builder session
~/.claude/skills/codex-orchestrator/scripts/codex-exec.sh builder "continue" --resume

# JSONL output for structured capture
~/.claude/skills/codex-orchestrator/scripts/codex-exec.sh researcher "What is 2+2?" --json | head -5

GPT-6-Astra Subagents

Use codex-astra.sh to run any existing persona on GPT-6-Astra. The launcher defaults to medium reasoning on the standard tier and validates Astra-specific choices.

~/.claude/skills/codex-orchestrator/scripts/codex-astra.sh list
~/.claude/skills/codex-orchestrator/scripts/codex-astra.sh architect "Design a fault-tolerant queue" --reasoning max --service-tier priority

Astra exposes 12 effort/tier permutations per persona. Use priority when latency materially matters and ultra only for independent parallel workstreams. See references/codex-models.md for the matrix and CLI mapping.

Session Management

For more control, use the Python session manager:

# List available profiles
python3 ~/.claude/skills/codex-orchestrator/scripts/codex-session.py list

# Start non-interactive session
python3 ~/.claude/skills/codex-orchestrator/scripts/codex-session.py start debugger "Trace the null pointer in UserService"

# Start interactive session
python3 ~/.claude/skills/codex-orchestrator/scripts/codex-session.py interactive architect

# Show profile details
python3 ~/.claude/skills/codex-orchestrator/scripts/codex-session.py info security

Profile Selection Guide

Review Tasks

• reviewer for general code quality and bugs
• security for vulnerability-focused review
• refactor for cleanup opportunities

Investigation Tasks

• debugger for bug investigation
• architect for understanding system behavior
• researcher for questions and analysis (read-only, no changes)
• adjudicator for weighing rival hypotheses, ambiguous evidence, and corpus-level comparisons (read-only, high reasoning)
• chat for open-ended conversation and brainstorming (read-only, ephemeral)

Creation Tasks

• architect for design decisions
• builder for new feature implementation
• docs for documentation
• refactor for implementation improvements
• planner for multi-hour implementation plans
• goal for autonomous /goal objective specifications

Chaining Patterns

Review → Debug → Fix

# 1. Identify issues
~/.claude/skills/codex-orchestrator/scripts/codex-exec.sh reviewer "Review src/api/ for bugs"

# 2. Investigate specific bug
~/.claude/skills/codex-orchestrator/scripts/codex-exec.sh debugger "Debug the race condition found in cache.ts"

Planner → Architect → Builder

# 1. Create comprehensive ExecPlan
~/.claude/skills/codex-orchestrator/scripts/codex-exec.sh planner "Create ExecPlan for new authentication system"

# 2. Validate architecture
~/.claude/skills/codex-orchestrator/scripts/codex-exec.sh architect "Review the auth system ExecPlan for design issues"

# 3. Build (plan guides implementation)
~/.claude/skills/codex-orchestrator/scripts/codex-exec.sh builder "Implement milestone 1 from the auth ExecPlan"

Architect → Builder → Reviewer

# 1. Design approach
~/.claude/skills/codex-orchestrator/scripts/codex-exec.sh architect "Design a caching layer for the API"

# 2. Build the feature
~/.claude/skills/codex-orchestrator/scripts/codex-exec.sh builder "Implement the caching layer from architect's design"

# 3. Review the implementation
~/.claude/skills/codex-orchestrator/scripts/codex-exec.sh reviewer "Review the new caching implementation"

Architect → Review → Refactor

# 1. Design approach
~/.claude/skills/codex-orchestrator/scripts/codex-exec.sh architect "Design repository layer extraction"

# 2. Validate design
~/.claude/skills/codex-orchestrator/scripts/codex-exec.sh reviewer "Review the proposed repository pattern"

# 3. Implement
~/.claude/skills/codex-orchestrator/scripts/codex-exec.sh refactor "Extract repository pattern from services"

Syseng → Architect → Planner

# 1. Assess infrastructure needs
~/.claude/skills/codex-orchestrator/scripts/codex-exec.sh syseng "Evaluate current deployment for scaling to 10x traffic"

# 2. Design architecture changes
~/.claude/skills/codex-orchestrator/scripts/codex-exec.sh architect "Design infrastructure to support 10x scale"

# 3. Create implementation plan
~/.claude/skills/codex-orchestrator/scripts/codex-exec.sh planner "Create ExecPlan for infrastructure scaling"

Security → Syseng

# 1. Security audit
~/.claude/skills/codex-orchestrator/scripts/codex-exec.sh security "Audit the Kubernetes cluster configuration"

# 2. Infrastructure hardening
~/.claude/skills/codex-orchestrator/scripts/codex-exec.sh syseng "Implement security recommendations from audit"

Researcher → Architect → Builder

# 1. Understand the problem space
~/.claude/skills/codex-orchestrator/scripts/codex-exec.sh researcher "How does the current caching work? What are its limitations?"

# 2. Design the solution
~/.claude/skills/codex-orchestrator/scripts/codex-exec.sh architect "Design a new caching layer addressing the limitations"

# 3. Implement
~/.claude/skills/codex-orchestrator/scripts/codex-exec.sh builder "Implement the caching layer from architect's design"

Backgrounding & Parallel Execution

Codex CLI v0.124.0+ requires a controlling TTY. When run with shell & or Claude Code's run_in_background: true, the TTY is detached and Codex silently produces empty output (openai/codex#19945). Longer prompts increase failure likelihood (empirically observed) — the researcher profile is especially vulnerable.

codex-exec.sh and codex-goal.sh automatically detect non-TTY contexts and wrap codex exec with script(1) to re-attach a pseudo-TTY. No user action is required.

Inside Codex's own exec_command tool the command already has a PTY, so the wrapper stays out of the way; requesting tty:true there changes nothing and is not a fix for missing output.

For direct codex exec calls (not through codex-exec.sh), wrap manually:

# macOS
script -q /dev/null codex exec --skip-git-repo-check -- "prompt" </dev/null

# Linux
script -qfc 'codex exec --skip-git-repo-check -- "prompt" </dev/null' /dev/null

Argv rules for direct codex exec calls

• End options with -- before the prompt. -i/--image <FILE>... is variadic: placed right before the prompt it consumes the prompt as another image path, Codex then reads stdin, finds nothing, and exits 1 with No prompt provided via stdin. --image also splits its value on commas.
• With resume, pass -i after the subcommand: codex exec <opts> resume --last -i shot.png -- "prompt". A parent-level -i before resume swallows the word resume.
• Keep </dev/null. When stdin is not a terminal and a prompt is present, Codex prints Reading additional input from stdin... and appends whatever it reads as a <stdin> block; an immediate EOF makes that a harmless one-liner, while an open pipe hangs forever (openai/codex#27019).

Personas are passed with -c developer_instructions=...; multiple codex-exec.sh instances can run in the same directory without changing or serializing around AGENTS.md. Still serialize agents that write overlapping project files, or give them separate worktrees.

--no-cleanup flag (codex-exec.sh only): When set, the output temp file is preserved after exit and its path is printed to stderr (OUTPUT_FILE=<path>). Use when the caller needs to retrieve the output file asynchronously.

Goal Runs

The /goal command sets a persistent objective that Codex works toward autonomously for hours. Goals are TUI-only (not available in codex exec mode), so goal runs use a two-phase workflow.

Two-Phase Workflow

Phase 1 — Draft: Generate a structured goal specification file using codex exec with the goal profile.

Phase 2 — Run: Launch the interactive Codex TUI with /goal set from the spec file.

Usage

# Draft a goal specification
~/.claude/skills/codex-orchestrator/scripts/codex-goal.sh draft "Add authentication with JWT to the API"

# Draft with custom output path
~/.claude/skills/codex-orchestrator/scripts/codex-goal.sh draft "Migrate to PostgreSQL" --output goals/pg-migration.md

# List existing goals
~/.claude/skills/codex-orchestrator/scripts/codex-goal.sh list

# Launch Codex TUI with a goal
~/.claude/skills/codex-orchestrator/scripts/codex-goal.sh run goals/goal-20260518-143000.md

# Run with custom model
~/.claude/skills/codex-orchestrator/scripts/codex-goal.sh run goals/goal-01.md --model gpt-5.6-sol

Goal File Format

Goal specifications follow a structured format with YAML frontmatter:

---
objective: "One-line summary"
status: draft
created: 2026-05-18
project: /path/to/project
---

# Goal: <objective>

## Objective / Stopping Condition / Context / Constraints / Validation / Checkpoints / Progress Log

Goals under 3,500 characters are inlined directly into the /goal command. Longer goals are referenced by file path.

Chaining: Planner → Goal → Run → Reviewer

# 1. Plan the feature
~/.claude/skills/codex-orchestrator/scripts/codex-exec.sh planner "Create ExecPlan for caching layer"

# 2. Draft a goal from the plan
~/.claude/skills/codex-orchestrator/scripts/codex-goal.sh draft "Implement caching layer per ExecPlan"

# 3. Launch autonomous goal run
~/.claude/skills/codex-orchestrator/scripts/codex-goal.sh run goals/goal-20260518-143000.md

# 4. Review what Codex built
~/.claude/skills/codex-orchestrator/scripts/codex-exec.sh reviewer "Review the caching implementation"

Caveats

• /goal is TUI-only — the run subcommand launches an interactive Codex session (stdin is not redirected)
• Requires features.goals = true in ~/.codex/config.toml (or --enable goals flag, which the script passes automatically)
• One goal per thread — must /goal clear before setting a different goal
• 4,000 character limit on /goal content (the draft phase targets under 3,500 to leave overhead)

See references/goal-command.md for the full /goal command reference.

Script Options

codex-exec.sh Options

Model & Reasoning Defaults

Each profile has a default model and reasoning effort. User flags override these.

General reasoning effort levels: none < minimal < low < medium < high < xhigh < max < ultra. Model support varies; Astra accepts low through ultra.

# Uses profile defaults (builder → gpt-5.6-sol + high)
~/.claude/skills/codex-orchestrator/scripts/codex-exec.sh builder "Implement auth module"

# Uses profile defaults (planner → gpt-5.6-sol + high)
~/.claude/skills/codex-orchestrator/scripts/codex-exec.sh planner "Create ExecPlan for caching"

# Override model for quick tasks
~/.claude/skills/codex-orchestrator/scripts/codex-exec.sh reviewer "Style check" --model gpt-5.6-luna

# Override reasoning only
~/.claude/skills/codex-orchestrator/scripts/codex-exec.sh builder "Quick lint fix" --reasoning medium

# Override both
~/.claude/skills/codex-orchestrator/scripts/codex-exec.sh planner "Design distributed cache" --model gpt-5.6-sol --reasoning max

# Astra at maximum depth on the priority tier
~/.claude/skills/codex-orchestrator/scripts/codex-exec.sh planner "Design distributed cache" --astra --reasoning max --service-tier priority

API Mode (Direct OpenAI API)

The --api flag bypasses Codex CLI entirely and calls the OpenAI API directly via gpt-api-chat.py. Billing goes to your OPENAI_API_KEY, not the Codex subscription.

When to use API mode:

• Access models not in your Codex subscription (e.g. gpt-5.6-sol)
• Multi-turn conversations with session persistence
• Streaming responses
• Custom system prompts

# One-shot API query
~/.claude/skills/codex-orchestrator/scripts/codex-exec.sh chat "Explain X" --api --model gpt-5.6-sol

# Multi-turn with session file (context carries across calls)
~/.claude/skills/codex-orchestrator/scripts/codex-exec.sh chat "What is X?" --api --session /tmp/session.json
~/.claude/skills/codex-orchestrator/scripts/codex-exec.sh chat "Tell me more" --api --session /tmp/session.json

# Streaming output
~/.claude/skills/codex-orchestrator/scripts/codex-exec.sh chat "Describe Y" --api --model gpt-5.6-sol --stream

# Custom system prompt
~/.claude/skills/codex-orchestrator/scripts/codex-exec.sh chat "Analyze this" --api --system "You are a security analyst"

# Any profile works with --api (sends raw prompt to OpenAI API, no Codex CLI)
~/.claude/skills/codex-orchestrator/scripts/codex-exec.sh researcher "What are the latest React patterns?" --api --model gpt-5.6-sol

Supported API-mode models: gpt-5.6-sol, gpt-5.6-terra, gpt-5.6-luna, gpt-5.5, gpt-5.5-pro, gpt-5.4, gpt-5-mini. Use codex-astra.sh for Astra subagents.

Three-way mode (Claude + GPT): Within a Claude Code session, invoke gpt-api-chat.py via Bash, read GPT's response, synthesize perspectives, and steer the conversation. Use --session to maintain GPT's context across turns.

Requires: OPENAI_API_KEY in environment (loaded from ~/.config/env/secrets.env)

Testing

Run the test suite to verify installation:

~/.claude/skills/codex-orchestrator/scripts/test-codex.sh         # Full tests
~/.claude/skills/codex-orchestrator/scripts/test-codex.sh --quick # Skip API test

Reference Documentation

For detailed information:

• references/codex-cli.md - Complete CLI command reference
• references/codex-models.md - Model, reasoning-effort, and service-tier matrix
• references/agents-md-format.md - AGENTS.md syntax and best practices
• references/subagent-patterns.md - Delegation patterns and examples
• references/goal-command.md - /goal command reference and best practices

Troubleshooting

"Codex CLI not found"

npm install -g @openai/codex

"Authentication error"

export OPENAI_API_KEY=sk-...
# or
codex login

"Model not supported with ChatGPT account"

Older model names (codex-mini, o3, o4-mini) have been deprecated. Current models include gpt-6-astra, gpt-5.6-sol, gpt-5.6-terra, gpt-5.6-luna, gpt-5.5, gpt-5.5-pro, and gpt-5-mini. Previous generation (gpt-5.4, gpt-5.4-pro, gpt-5.3-codex, gpt-5.2) still works but is superseded. Set an API key instead of using codex login:

export OPENAI_API_KEY=sk-...

"Profile not found"

Available profiles: reviewer, debugger, architect, security, refactor, docs, planner, syseng, builder, researcher, adjudicator, chat, goal

Check profile exists:

ls ~/.claude/skills/codex-orchestrator/agents/

"Codex produced no output"

The researcher/adjudicator/chat profiles capture output to a temp file. If Codex exits without writing to it, the script prints the last diagnostic lines from Codex (stderr, PTY-merged log lines, and JSONL error events), warns, and exits 1. Read those lines first; they name the cause. Common causes:

• Prompt never reached Codex: stderr shows Reading prompt from stdin... then No prompt provided via stdin. This was the 2026-09-19 failure: --image was emitted directly before the prompt and the variadic -i consumed it. The launcher now places -i after resume and terminates options with --; if you see this from a hand-written codex exec, apply the argv rules above. Codex CLI 0.155.x did not change PTY handling or prompt forwarding, so do not reach for TTY workarounds on this symptom.
• Auth or model errors: stderr shows the Codex banner followed by an API error. Retry or fix credentials.
• TTY detachment: Codex CLI v0.124.0+ silently crashes when backgrounded without a TTY. codex-exec.sh auto-wraps with script(1), so this only applies to hand-written background launches.
• Codex session too short to produce a response
• Model returned empty response (retry)
• Profile instructions were rejected by an outdated Codex CLI (update Codex and retry)

Upgrade from AGENTS.md-injection releases

An interrupted run from an older release may have left an AGENTS.md symlink into codex-orchestrator/agents/ plus .AGENTS.md.codex-backup.<pid>. Inspect those paths and restore the matching backup once before parallel use. Keep the backup until the repository instructions are verified; current launchers deliberately do not guess among legacy backups.

Poor Results

• Narrow the task scope
• Provide more context in the prompt
• Try a different profile
• Use --model gpt-5.6-sol --reasoning max for complex tasks