Delegation

Delegation Skill

Delegate tasks to specialized subagents for autonomous execution.

When to Use

Use this skill when:

• You need to delegate tasks to run autonomously (especially when you have a session plan)
• You have multiple independent tasks that can run in parallel

When NOT to Use

• Simple tasks that you can complete directly
• Tasks that require immediate user interaction
• Read-only operations (use ripgrep or glob/read tools instead)
• Quick file edits (use edit tool directly)
• Analysis, planning, or debugging (e.g. performing code review)

Available Functions

subagent(task, fromPlan, relevantFiles, relevantSkills)

Launch a subagent to handle a task synchronously. Blocks until the subagent completes and returns the result.

Parameters:

• task (str, required): Task ID from session plan (e.g., "T003") OR brief task description
• fromPlan (bool, default False): If True, subagent reads full task context from .local/session_plan.md
• relevantFiles (list[str], optional): File paths the subagent should access
• relevantSkills (list[str], optional): Paths to all implementation skills you've read. Use the full path from the skills view. Pass every skill with integration details (auth, storage, payments) — not orchestration skills meant for you (design, delegation, react-vite).
• specialization (str, default "GENERAL"): "GENERAL" or "SMALL_TASK" for quick tasks

Returns: Dict with task results

{
    "success": true,
    "message": "Task summary",
    "subagentAlias": "subagent_1",
    "result": "Full task output..."
}

Usage Patterns:

1. From Plan (Recommended when you have a session plan):

// Reference tasks by ID - subagent reads full context from .local/session_plan.md
const result = await subagent({ task: "T003", fromPlan: true });
console.log(result);  // Always print the result

2. Direct Task (for ad-hoc tasks without a plan):

const result = await subagent({
    task: "Fix the auth bug in src/auth.ts",
    relevantFiles: ["src/auth.ts"],
    relevantSkills: [".local/skills/clerk-auth/SKILL.md", ".local/skills/database/SKILL.md"]
});
console.log(result);  // Always print the result

startAsyncSubagent(task, fromPlan, relevantFiles, relevantSkills)

Launch a subagent to handle a task asynchronously in the background. Returns immediately without waiting for completion. Use the wait_for_background_tasks tool to collect results later.

Parameters:

• task (str, required): Task ID from session plan (e.g., "T003") OR brief task description
• fromPlan (bool, default False): If True, subagent reads full task context from .local/session_plan.md
• relevantFiles (list[str], optional): File paths the subagent should access
• relevantSkills (list[str], optional): Paths to all implementation skills you've read. Use the full path from the skills view. Pass every skill with integration details (auth, storage, payments) — not orchestration skills meant for you (design, delegation, react-vite).
• specialization (str, default "GENERAL"): "GENERAL" or "SMALL_TASK" for quick tasks

Returns: Immediately with acknowledgment. Use the wait_for_background_tasks tool to collect results.

Usage Patterns:

1. From Plan (Recommended when you have a session plan):

// Reference tasks by ID - no print needed, result comes via wait_for_background_tasks
await startAsyncSubagent({ task: "T003", fromPlan: true });

2. Direct Task (for ad-hoc tasks without a plan):

await startAsyncSubagent({
    task: "Fix the auth bug in src/auth.ts",
    relevantFiles: ["src/auth.ts"],
    relevantSkills: [".local/skills/clerk-auth/SKILL.md", ".local/skills/database/SKILL.md"]
});

Parallel Execution:

// Launch independent tasks simultaneously
console.log(await startAsyncSubagent({ task: "T002", fromPlan: true }));
console.log(await startAsyncSubagent({ task: "T003", fromPlan: true }));
console.log(await startAsyncSubagent({ task: "T004", fromPlan: true }));

Giving Subagents Access to Skills:

// Pass skill paths from the skills view
await startAsyncSubagent({
    task: "T005",
    fromPlan: true,
    relevantSkills: [".local/skills/database/SKILL.md", ".local/skills/clerk-auth/SKILL.md"]
});

messageSubagent(subagentId, message)

Send a follow-up message to an existing subagent asynchronously. Returns immediately after the message is delivered. Use this when the subagent should continue in the background.

Parameters:

• subagentId (str, required): Alias returned when the subagent was started
• message (str, required): Follow-up instruction or clarification for the subagent

Returns: Acknowledgment that the message was sent. The subagent keeps running in the background.

Example:

// Async follow-up: continue background work
await messageSubagent({
    subagentId: "subagent-happy-tiger",
    message: "After the fix, add regression tests for the auth edge case."
});

Use the wait_for_background_tasks tool to collect results later.

messageSubagentAndGetResponse(subagentId, message, timeoutSeconds)

Send a follow-up message to an existing subagent synchronously. Blocks until the subagent completes, times out, or is interrupted.

Parameters:

• subagentId (str, required): Alias returned when the subagent was started
• message (str, required): Follow-up instruction or clarification for the subagent
• timeoutSeconds (float, default 300.0): Max time to wait before returning a timeout status

Returns: Dict with success, message, result, and exitReason.

Example:

// Sync follow-up: wait for the answer now
const result = await messageSubagentAndGetResponse({
    subagentId: "subagent-happy-tiger",
    message: "Summarize root cause and include exact changed files.",
    timeoutSeconds: 180.0
});
console.log(result.result);

When to choose which:

• Use messageSubagent when you want fire-and-forget behavior
• Use messageSubagentAndGetResponse when you need the response immediately

Best Practices

1. Use session plans: For 2+ tasks, create a .local/session_plan.md and use fromPlan=True
2. Launch in parallel: Independent tasks can run simultaneously with startAsyncSubagent
3. Use subagent() when you need the result immediately: For tasks where you need to act on the output, use the synchronous subagent() and print the result.
4. Use startAsyncSubagent() for independent tasks that can run in the background: The tasks will be performed in parallel.
5. Use messageSubagent() for async follow-ups: Message running subagents without blocking, then use the wait_for_background_tasks tool to collect results
6. Use messageSubagentAndGetResponse() for sync follow-ups: Use this when you need the subagent's output before continuing
7. Trust the results: Subagent outputs should generally be trusted
8. Pass all implementation skills via relevantSkills: Include every implementation skill you've read — use the full path from the skills view for each one

Subagent Capabilities

The subagent has access to:

• File operations (read, write, edit, glob, grep)
• Bash commands
• LSP diagnostics
• Depending on specialization, additional tools like media generation or canvas tools may be available
• Skills (loaded via relevantSkills parameter)

The subagent does NOT:

• Run or restart workflows
• Preview/test the app (that's your job as main agent)