Goal

Use this skill to delegate a focused, bounded task to delegate_subagent_task. The child agent executes the task once in an isolated context, returns its result to the master agent, and terminates. It does NOT create an interactive session with the end user or switch roles in the current chat.

Use when

• The user asks to delegate, split, investigate, or run an isolated subtask with a separate agent.
• The task benefits from a separate chat/context or a shorter, focused prompt.
• The user explicitly wants background execution or the work is long-running and does not block the next master step.

Sequential Workflow

1. Call delegate_subagent_task with a self-contained prompt.
2. Omit parallel or set it to false.
3. Use the returned tool result as the subagent answer.
4. Continue only after integrating the returned result into the master response or next step.

Parallel Workflow

1. Call delegate_subagent_task with parallel: true.
2. Read the immediate tool result and keep both identifiers:
   • background_process_id
   • initiating_message_key
3. Continue with master work that does not depend on the subagent result.
4. Use inspect_background_process with background_process_id to check status and stream events.
5. Use get_subagent_execution_trace_by_initiating_message with initiating_message_key to inspect the subagent trace and final answer.
6. If the result is not ready, inspect again later with the newest last_event_counter.
7. Use cancel_background_process only when the user wants the background run stopped or when the run is clearly obsolete.

Verification Rules

• For sequential runs, the tool response is the verification surface.
• For parallel runs, check the background process first. A completed process can still have trace details that should be read through get_subagent_execution_trace_by_initiating_message.
• If get_subagent_execution_trace_by_initiating_message returns no trace yet, the subagent may still be running or may not have persisted a trace. Use inspect_background_process before treating that as failure.
• If the background_process_id is lost, use search_background_processes with process_type: "subagent_execution" and process_subject: "delegate_subagent_task".

Prompt Rules

• Make the prompt self-contained: goal, constraints, expected output, and what not to change.
• Mention whether the subagent should only inspect, produce a proposal, or implement changes.
• Include relevant keys or names when targeting another chat/context.
• Keep the prompt narrow enough that the master can evaluate the result.

Targeting

• chat_key selects an exact target chat and takes priority over context/name targeting.
• context_name selects a target context when no chat_key is given.
• chat_name selects or creates a chat inside the target context.
• history_length controls how much target-chat history the subagent sees. Use 0 for an isolated run.

Examples

Sequential:

{
  "prompt": "Inspect the current context and summarize the risks in the proposed scheduler change. Return concrete findings only.",
  "history_length": 0,
  "parallel": false
}

Parallel:

{
  "prompt": "Investigate the generic-agent-interface tests for brittle assumptions around subagent tool schemas. Return a concise report with file references.",
  "history_length": 0,
  "parallel": true
}