agent-board-routing

Agent Board Routing

Trigger

Load this skill when Dan explicitly delegates project work to a named coding agent, for example:

• “Have Claude implement…”
• “Send this to Codex”
• “Use Copilot to fix…”
• “Have Grok work on…”
• “Put this on the Agent Board and run it”

Do not route general coding questions, comparisons, hypotheticals, quoted examples, or requests that explicitly say not to start work.

Procedure

1. Call agent_board_list_projects when the project is not already identified by a canonical Board project ID
2. Resolve only an exact ID, exact project name, or unique project alias. Never silently use the Default project
3. Call agent_board_list_agents and confirm the explicitly requested agent is available. Never silently substitute another agent
4. Build a bounded execution contract containing:
   • objective
   • requirements and non-goals
   • acceptance criteria
   • relevant files or context
   • verification commands or expected checks
   • safety boundaries
5. Call agent_board_route_task with auto_start=true unless Dan said to queue it without starting
6. Return the project, agent, card ID, direct Board link, branch, and current status
7. Follow-ups for an active card should use agent_board_send_message, not create a duplicate card
8. Continuation of stopped/completed work must call agent_board_route_task with existing_task so the durable card ID is reused. New work that is connected but distinct must use related_item; never encode relationships with group IDs
9. Resolve projects and item references only by exact ID, exact name/title, or a unique alias. If the Board reports ambiguity, stop and ask rather than guessing
10. Use agent_board_get_task to check status when Dan asks for progress
11. Use timeout_minutes for work expected to exceed the 60-minute Board default. If a task times out, call agent_board_retry_task on the existing card with a larger limit; never bypass tracking by launching the named coding agent directly