orchestrator

Orchestrator — Codex + AutoAgent Integration

You have two complementary external analysis tools available. Each has distinct strengths. This skill guides when and how to use them.

Tool Comparison

Dimension	Codex CLI	AutoAgent
Invocation	codex exec or /codex	auto_oneshot.py or /auto
Model	gpt-5.4 (default)	COMPLETION_MODEL (default: claude-3-5-sonnet)
Execution	Single-pass, non-interactive	Multi-turn, agent orchestration
Environment	Read-only sandbox (default)	Docker sandbox or local
Best for	Static analysis, code review	Dynamic tasks, research, coding
Speed	Fast (single LLM call)	Slower (multi-agent, imports)
Risk	Low (read-only default)	Moderate (can execute code)

When to Use Codex

Codex excels at focused, single-pass analytical tasks:

• Code review — security audits, bug hunting, quality checks
• Planning — implementation plans, edge case analysis
• Static analysis — finding patterns, reviewing logic

Invoke via:

• /codex <task> — explicit user command
• codex-agent subagent — Claude's autonomous delegation
• /dispatch <task> — automatic routing

Prompt best practices for Codex:

1. Be specific about scope: "Review ONLY the auth flow"
2. Add constraints: "ONLY security issues, no style suggestions"
3. Request format: "Format as file:line - issue"
4. Skip fluff: "Skip preambles. Lead with findings."

When to Use AutoAgent

AutoAgent excels at multi-step tasks requiring execution or browsing:

• Coding with execution — write, run, iterate on code in a sandbox
• Web research — browse websites, gather current data
• File analysis — process PDFs, documents, spreadsheets
• Complex orchestration — tasks needing multiple agent types

Invoke via:

• /auto <task> — explicit user command
• auto-agent subagent — Claude's autonomous delegation
• /dispatch <task> — automatic routing

Agent functions:

Function	Use Case
get_system_triage_agent	Complex tasks — auto-routes internally
get_coding_agent	Write and execute code
get_websurfer_agent	Browse the web
get_filesurfer_agent	Read/analyze files

The /dispatch Command

When unsure which tool to use, /dispatch <task> analyzes the task and routes automatically:

• Code review/analysis tasks go to Codex
• Execution/research/file tasks go to AutoAgent
• Simple tasks are handled directly by Claude

Combining Both Tools

For complex tasks, use both tools together:

1. Codex for review, AutoAgent for implementation: Plan with Codex, then hand the plan to AutoAgent for execution
2. AutoAgent for research, Codex for review: Research with AutoAgent's web surfer, then have Codex review the resulting code
3. Parallel analysis: Spawn both as subagents for independent perspectives

The router-agent

The router-agent subagent can be spawned when Claude wants to automatically pick the right tool. It analyzes task signals and dispatches to either codex-agent or auto-agent.

Prerequisites

• Codex CLI: Must be installed (brew install codex) and authenticated (codex login)
• AutoAgent: Must be installed (cd ~/Control/autoagent && pip install -e .)
• Docker (optional): Required for AutoAgent's full sandboxed execution. Use --local flag without Docker.

Troubleshooting

Issue	Solution
codex: command not found	brew install codex then codex login
AutoAgent import errors	cd ~/Control/autoagent && pip install -e .
Docker not available	Use --local flag with /auto
Timeout on AutoAgent	Break task into smaller pieces
Codex API key issues	Run codex login to re-authenticate