delegation-autopilot

This skill is meant for the parent Codex agent in interactive mode.

Prerequisites (avoid common first-run failures)

Delegated runs can take minutes. If tool calls fail around ~60s with an error like timed out awaiting tools/call after 60s, increase the MCP tool timeout for this server in ${CODEX_HOME:-~/.codex}/config.toml:

[mcp_servers.codex-specialized-subagents]
tool_timeout_sec = 600

When to call delegate_autopilot

• The user request is multi-step ("and", "then", "also", "plus") or touches multiple areas (code + tests + docs).
• The change is cross-cutting (multiple files/modules) or likely needs specialist attention (security/perf/research).
• The user explicitly asks to delegate / use subagents.

Default rules:

• If the request includes code changes and (tests or docs), call delegate_autopilot before running shell commands or editing files.
• If the request is multi-step research-only and you want consistent “scan/implement/verify” decomposition or per-job thinking_level behavior, call delegate_autopilot with sandbox="read-only".

When not to call

• The user is asking a simple question or wants a short explanation.
• The change is tiny and local (single file, trivial edit) and you can do it directly.

How to call (minimal)

Call MCP tool delegate_autopilot with:

• task: the user request (verbatim)
• cwd: current workspace directory (optional; defaults are OK)

Optional knobs:

• sandbox: use "workspace-write" when you want subagents to edit files; "read-only" for analysis-only delegation; avoid "danger-full-access" unless absolutely necessary.
• max_agents: default 3 (use 1 when you want a single subagent run).
• skills_mode: default "auto" (use "none" to disable skill selection).

After calling

• If structuredContent.decision.should_delegate is false: continue normally without delegation.
• Otherwise: use structuredContent.aggregate as the consolidated result, and inspect structuredContent.run_dir for artifacts.

If you choose delegate_run anyway

• If you need higher/lower reasoning for a one-off run, pass reasoning_effort (or a model_reasoning_effort=... entry in config_overrides).
• If you want a default for manual runs, set CODEX_DELEGATE_REASONING_EFFORT on the MCP server process.

Safety

• Never paste secrets from ${CODEX_HOME:-~/.codex} or any local config.
• Avoid pasting full codex mcp list output or entire config files; redact anything that looks like a token/key.
• Do not recurse: delegated subagents must not call any delegate_* tools.