Codex Subagent Orchestrator

A lightweight MVP for running parallel subagents with configurable models and reasoning effort. SKILL.md defines the Codex workflow, mcp_server.py exposes an MCP tool, and orchestrator.py launches and collects subprocess results.

Configuration

Set the command template environment variable. It must include {model}, {effort}, and {prompt} placeholders:

$env:CODEX_SUBAGENT_COMMAND = 'codex exec --ephemeral --skip-git-repo-check --model {model} -c reasoning_effort="{effort}" "{prompt}"'

This template works with the current Codex CLI. Adjust it according to codex exec --help if your CLI version differs. Then run:

python .\orchestrator.py "Analyze this project's test gaps" --model gpt-6.1-sol --effort high --count 3

Results are printed in subtask order. The process exits with code 1 if any subtask fails.

Related MCP server: learn-mcp-agent-crew

Use with Codex

Use this directory as a local Skill or copy its contents to your Codex plugin directory. The natural-language entry point is /delegate; its parameters match the script.

Configure mcp_server.py as an stdio server in your MCP client. It exposes the delegate tool with task, model, effort, and count arguments.

Development

python -m py_compile orchestrator.py mcp_server.py
python -m unittest discover -s tests -v