hmem-dispatch

name	hmem-dispatch
description	Dispatch an isolated sub-agent that returns only [RESULT]...[/RESULT]. Use for any search, lookup, 'does X exist?', 'find Y', calculation, or isolated writing — never do these inline when dispatch fits.

hmem-dispatch

TRIGGER

Always dispatch for:

• Searching — "does a spec exist?", "find the plan for X", "search the codebase for Y", "check if skill Z covers this", "read the docs for W", "is there already a file that does X?" — any exploration or lookup task
• Calculation or lookup — deterministic, context-free
• Writing an isolated section or document — no conversation history needed
• Writing or modifying code — never write code directly in the main context. Pass the project ID and relevant node IDs (e.g. .2 Codebase) in the task so the sub-agent can sync them after completing the work. Always include in the task: "Invoke superpowers:coding-discipline before writing code."

Never dispatch for tasks that require knowing the conversation context.

The key rule for searching: if you're about to run grep, find, read, or search_memory to answer a question — dispatch instead. The sub-agent explores, you synthesize the result.

STEP 1: Define the task

Write out before dispatching:

• INPUT: exactly what the sub-agent needs — no more, no less
• TASK: what to do with the input
• OUTPUT FORMAT: what to return
• VERIFY (optional): a concrete shell command that proves completion (e.g. npx tsc --noEmit, grep "pattern" file.ts). Use for tasks that produce code or file changes.

STEP 2: Dispatch

Send the sub-agent ONLY this prompt — no conversation history, no project context unless the task explicitly requires it:

---

Task: Input: [If VERIFY set:] After completing the task, run: <VERIFY_CMD> Return your answer in exactly this format: [RESULT] [/RESULT] [If VERIFY set:] [VERIFY_RESULT] pass | fail: [/VERIFY_RESULT] Max 200 words. Use the hmem-subagent skill.

STEP 3: Inject result

Take ONLY the content between [RESULT] and [/RESULT]. If [VERIFY_RESULT] is present: check pass/fail. On fail, report the verification error instead of declaring the task complete. Discard all sub-agent reasoning and preamble. Use the result directly in the main conversation.