codex-specialized-subagents

Artifact-first sub-agent delegation for Codex CLI (MCP server).

This repo provides a local (stdio) MCP server that exposes:

• delegate_autopilot — decide whether to delegate and, if yes, orchestrate one or more codex exec sub-agent runs
• delegate_run — run a single specialist sub-agent via codex exec
• delegate_resume — resume a prior sub-agent thread via codex exec resume

Each tool call writes a run directory under ${CODEX_HOME:-$HOME/.codex}/delegator/runs/<run_id>/ containing the prompt, selected skills, event stream, and structured results (artifact-first debugging).

When to use

• You want parallelism and specialization for multi-step / cross-cutting work.
• You want durable artifacts (logs + outputs) to debug and review what happened.

Requirements

• Node.js >=20 (see package.json#engines)
• npm
• mise (recommended): installs the pinned runtime from mise.toml
• codex CLI on your PATH and authenticated (required for real delegation runs)

Optional:

• Python 3 (only for helper scripts under .agent/)

Install & quickstart (from source)

From the repo root (installs deps + builds dist/):

# Recommended: install the pinned runtime (see mise.toml)
mise install

# Drift check (lockfiles + pins)
./toolchain-check.sh

# Install deps from lockfile
npm ci

# Build
npm run build

Configure Codex (recommended, prevents timeouts)

Delegated runs can take minutes. Set this server’s MCP tool timeout to 1200 seconds (20 minutes) in your Codex config ($HOME/.codex/config.toml):

mkdir -p "$HOME/.codex"
cat >> "$HOME/.codex/config.toml" <<'EOF'

[mcp_servers.codex-specialized-subagents]
tool_timeout_sec = 1200
EOF

If you already have a [mcp_servers.codex-specialized-subagents] section, edit the existing tool_timeout_sec instead of appending a duplicate.

Common gotcha: tool_timeout_sec is not an env var. If you put it under mcp_servers.codex-specialized-subagents.env.*, Codex will error with “expected a string” (env values must be strings).

If you see a TOML parse error about a “duplicate key” for [mcp_servers.codex-specialized-subagents], you have that table declared twice — keep only one header and put tool_timeout_sec = 1200 inside it.

Register with Codex (recommended defaults)

From the repo root (includes per-job reasoning-effort overrides for delegate_autopilot):

codex mcp add codex-specialized-subagents \
  --env CODEX_AUTOPILOT_REASONING_EFFORT_LOW=low \
  --env CODEX_AUTOPILOT_REASONING_EFFORT_MEDIUM=medium \
  --env CODEX_AUTOPILOT_REASONING_EFFORT_HIGH=high \
  -- node "$(pwd)/dist/cli.js"

Verify:

codex mcp get codex-specialized-subagents

Remove:

codex mcp remove codex-specialized-subagents

Usage

Interactive autopilot (recommended)

In Codex interactive mode, delegate_autopilot can split a request into jobs (scan / implement / verify) and run specialist sub-agents.

To make delegation feel automatic in interactive mode, install the included delegation-autopilot skill globally:

mkdir -p "${CODEX_HOME:-$HOME/.codex}/skills/delegation-autopilot"
cp .codex/skills/delegation-autopilot/SKILL.md \
  "${CODEX_HOME:-$HOME/.codex}/skills/delegation-autopilot/SKILL.md"

Then try prompts like:

• “Refactor the MCP server and update tests + README.”
• “Audit the repo docs and propose improvements.”

Optional: delegate_autopilot assigns each job a thinking_level (low | medium | high). You can set CODEX_AUTOPILOT_REASONING_EFFORT_LOW|MEDIUM|HIGH on the MCP server process to override Codex model_reasoning_effort per job (see docs/usage.md). (Legacy/advanced: CODEX_AUTOPILOT_MODEL_LOW|MEDIUM|HIGH overrides model name.)

Manual tool calls

If you prefer explicit tool usage, tell Codex to call one of:

• delegate_autopilot (multi-agent orchestration)
• delegate_run (single sub-agent run)
• delegate_resume (resume a prior sub-agent thread)

Optional (advanced): for delegate_run / delegate_resume, you can pass reasoning_effort (maps to codex exec -c model_reasoning_effort="...") or raw config_overrides (maps to codex exec -c <override>). If you want a default for manual runs, set CODEX_DELEGATE_REASONING_EFFORT on the MCP server process (only applies when reasoning_effort is omitted and config_overrides does not already set model_reasoning_effort).

Skills

Sub-agent runs can load Codex skills from:

• repo-local .codex/skills (nearest ancestor of the delegated cwd)
• global ${CODEX_HOME:-$HOME/.codex}/skills

Note: this repo’s delegation-autopilot skill is marked delegator_exclude: true (parent-only) to prevent delegation recursion.

Artifacts (run directories)

Each tool call writes a run directory under ${CODEX_HOME:-$HOME/.codex}/delegator/runs/<run_id>/.

Documentation

Start with docs/README.md (index), then:

• docs/usage.md — how to use the tools effectively
• docs/troubleshooting.md — common failure modes (timeouts, missing codex, etc.)
• docs/development.md — local development and test matrix
• docs/reference/tools.md — full tool schemas (inputs/outputs)
• docs/reference/run-directories.md — run directory layout and artifact meaning

Development

npm test
npm run lint
npm run dev

Integration tests (requires Codex CLI + auth):

RUN_CODEX_INTEGRATION_TESTS=1 npm test

Contributing: CONTRIBUTING.md.

Security

• Don’t commit secrets (.env is gitignored; use .env.example as a template).
• Run directories can contain sensitive prompts/output; treat ${CODEX_HOME:-$HOME/.codex}/delegator/runs as sensitive.

Reporting: SECURITY.md.

License

MIT