htekdev/copilot-home-assistant/agent-dispatch

Agent Dispatch Skill

The canonical decision framework for choosing between launching a fresh agent (task tool) vs steering an existing idle agent (write_agent). This applies to ALL orchestration decisions, not just cron (cron has its own absolute rule — see cron-dispatch skill).

The Core Question

Does this message CONTINUE an existing conversation, or START a new one?

Decision Flow

1. list_agents() → check for IDLE agents
2. Is there an idle agent in the SAME domain as this request?
   NO → Launch new (task tool)
3. Is this a FOLLOW-UP to that agent's prior work?
   (correcting, clarifying, continuing a discussion)
   NO → Launch new (task tool)
4. Does that agent have CONTEXT that would be lost by launching fresh?
   (names, decisions, partial work, conversation state)
   NO → Launch new (task tool)
5. Is this a CRON dispatch?
   YES → Launch new ALWAYS (see cron-dispatch skill — ABSOLUTE rule)
6. All conditions met → STEER via write_agent

Launch New Agent (task tool) — WHEN ANY is true

Steer Existing Agent (write_agent) — WHEN ALL are true

Examples of Valid Steering

# User said "add eggs to the list" — shopping agent already has the grocery context
write_agent(agent_id: "shopping-abc", message: "Also add eggs to the H-E-B cart")
 
# User corrected a misidentified card — finance agent has the card discussion
write_agent(agent_id: "finance-xyz", message: "No, the Savor is the subscription card, not the travel card")
 
# User wants to continue a meal plan conversation
write_agent(agent_id: "meal-abc", message: "What about Thursday dinner?")

Anti-Patterns (NEVER do these)

• ❌ Funneling every task through write_agent just because an agent is idle
• ❌ Steering cron jobs into existing agents (ABSOLUTE prohibition)
• ❌ Injecting unrelated work into an agent that's idle but in a different domain
• ❌ Assuming an agent's context is still relevant hours after its last turn
• ❌ Using write_agent for complex new requests that deserve clean context

Priority Guideline

When in doubt: launch new. The cost of a fresh agent is ~30-60s startup time. The cost of polluted context is degraded output quality, confused behavior, and debugging nightmares.

Relationship to Other Skills

• `cron-dispatch` — Contains the absolute rule for cron jobs (always fresh). This skill covers the general case.
• `quick-task-transition` — For "done/next" transitions, handle inline (no agent at all). This skill is for when you DO need an agent.