delegate

name	delegate
description	Delegate implementation to OpenAI Codex via the codex-delegate-mcp MCP delegate tool. Use when the user says delegate to Codex, hand off to Codex, offload this to Codex, use Codex for coding, plan before building with Codex, review with Codex, or resume a Codex delegation. Do not shell out to codex — use the delegate MCP tool.

Delegate to Codex

You orchestrate; Codex implements. Use the codex-delegate-mcp MCP server — never run codex from the shell for these tasks.

For the full input and result field semantics, the mode rules, and how concurrency and timeouts behave, read reference.md in this skill directory.

When to delegate

• Trivial (one-liner, rename, typo): do it yourself.
• Medium (multi-file feature or refactor): one delegate call.
• Large or risky: mode: "plan" first; implement after approval.
• Advisory questions → mode: "ask". Code review of a diff/commit → mode: "review".

Workflow

1. Build the brief inline in spec:
   • Goal — the outcome, precisely.
   • Scope — which files/directories are in play.
   • Decisions already made — quote the user's exact values verbatim.
   • Done when — verifiable acceptance criteria. Point at files to read; don't paste large code blocks.
2. Call delegate on codex-delegate-mcp.
3. Review — check status before trusting result: a run that spawns and then fails returns normally rather than raising, so a caller that only catches errors reads a failure as an empty success. Then read warnings, then filesReportedByEditTools, then the git diff; run tests/lint. The field lists only what Codex's edit tool reported — files it wrote through a shell command are missing, and anything it edited outside the workspace is listed as an absolute path — so the diff is the better record, not a complete one.
   • A warnings entry always means something real; empty warnings is not a clean bill of health — the bridge sees only failures Codex reports as failed or declined tool calls, not ones it narrates in result. Such a warning reports Codex's status, not a verdict; it carries its own reading of that status.
   • result is the authoritative final message only on status: "completed"; every other status carries an empty result and a reason. result-unavailable means the run finished but its final message could not be read or parsed — inspect the diff and resume the thread for a concise final answer.
   • If criteria fail: resume the same thread with resumeThreadId and a specific fix brief (pass the same workspace).
   • If a resume returns resumed: false with a new threadId, Codex minted a fresh thread and prior context did not carry over.
   • After 2 failed resumes, start a fresh thread with a rewritten brief.
4. Report — summarize what changed and whether acceptance criteria are met.

Running several at once

Delegations run concurrently. Worth doing for the same question to different models (one call per model, then compare) and for independent work in independent directories (one worker per workspace). Not worth doing when the tasks touch the same files — two agents writing one tree overwrite each other and the diff cannot say which did what. Split by directory, or serialize.

To cancel one of several, pass its delegationId. See reference.md.

Leave model, reasoningEffort and fast at their defaults unless the user asks. Other models (gpt-5.6-sol, gpt-5.6-terra) are available — pass model only on request. workspace is required on every call, resumes included — pass the one the thread started in, or the resumed turn runs somewhere else.

Plan mode

1. delegate(spec, mode="plan") → save threadId, read plan.
2. Present the plan; wait for approval.
3. delegate("implement the approved plan", mode="agent", resumeThreadId=<threadId>, workspace=<same workspace>).

Review mode

Pass exactly one reviewTarget:

• { "kind": "uncommitted" }
• { "kind": "base", "branch": "main" }
• { "kind": "commit", "sha": "..." }

Review cannot be resumed. Put focus instructions in spec.

Timeouts

A quiet run is not a stuck run: Codex emits nothing while a shell command runs or the model reasons, so long silent stretches are normal and are not timed out. Guards are a 60s spawn-to-first-output deadline and a 1h hard cap (timeoutMs). Raise timeoutMs for work that legitimately runs longer.