Agent Delegation Skill

Purpose

Use the local codex-agent-delegator MCP server to hand bounded, mechanical coding work to an installed agent while keeping architecture and review in the current Codex task.

Good delegation candidates

• mechanical refactors and API migrations;
• focused unit-test additions;
• documentation updates;
• repetitive style or UI changes;
• repository searches that benefit from a persisted report.

Do not delegate secrets, authentication/payment/cryptography changes, irreversible data migrations, major architecture decisions, or a task the user explicitly asked the current agent to implement personally.

Required task card

Before delegating, define:

• the exact expected outcome;
• narrow allowedFiles and appropriate forbiddenFiles;
• acceptance commands in testCommands;
• the backend: agy, codex, or claude;
• the least privilege needed.

Use permissionMode: "workspace-write" and useWorktree: true for edits. Do not set full-access or allowUnsafe unless the user has authorized the specific risk. A custom backend always requires that explicit trust decision.

Workflow

1. Call list_agent_backends if backend availability is unknown.
2. Call delegate_to_agent with a narrow task card.
3. For the default asynchronous mode, poll get_agent_run_report using the returned runId.
4. Follow the review skill. Fetch logs or patches only when the compact report does not provide enough evidence.
5. Call apply_agent_run only after review and with confirm: true.
6. Call cleanup_agent_run when the run and worktree are no longer needed.

Example

{
  "repoPath": "/workspace/my-app",
  "task": "Convert model interfaces to exported type aliases without changing runtime code.",
  "agent": "codex",
  "allowedFiles": ["src/models/*.ts"],
  "forbiddenFiles": ["src/models/legacy/**", ".env"],
  "testCommands": ["npm run typecheck"],
  "permissionMode": "workspace-write",
  "useWorktree": true
}