router · diff
git:20260817.0cc07d9 to git:20260817.f04083f
6 added, 36 removed. Audit A to A.
---
name: router
description: Spawn agents and verify their output — route work, never do it. Use when the work splits into pieces agents can carry and the session should route and check, not execute.
---
- ## Extract
-
- From the conversation/context/user:
-
- - The **work** on the table — the goal as given.
- - Its **pieces** — the parts it splits into, each one an agent could carry alone.
-
- ## Gate
-
- Proceed only when: the work splits into pieces an agent can carry alone, and each returned piece can be checked for less than the cost of doing it.
-
- Anything else, say which in one plain line — never route anyway:
-
- - The work is one piece — a single pass, or a plain answer → do it yourself, plainly, and proceed with what it produces; there is nothing to split.
- - A piece only this session can do → stop.
- - Verifying a piece's return would mean redoing the piece → stop.
- - The work will run without a human watching, optimizing toward something, and how it's judged is undefined before it runs → use **target** skill.
-
- ## Hold
-
For as long as the loop below runs, you are a router, not a worker:
- - Anything that touches files, shell, or the web is work — it goes to an agent, never to you.
- - Your own turns go to spawning, verifying, and talking to the user — never doing.
+ - Anything that touches files, shell, or the web is work — it goes to an agent.
+ - Your own turns go to spawning, verifying, and talking to the user.
- Prefer a turn spent checking an agent's work over a turn spent doing the work yourself.
- An explicit order from the user to do something yourself overrides this.
- ## Route
-
- 1. Dispatch the pieces. Each agent starts blind: write its prompt to carry the piece's goal, the constraints, why it matters, and what is out of scope. Independent pieces dispatch together, never one at a time.
- 2. Verify each return against what its prompt asked. An agent's summary is what it intended, not what it did — read what it produced, never what it reported. Nothing reaches the user unverified.
+ 1. Dispatch the pieces. Each agent starts blind: write its prompt to carry the piece's goal, the constraints, why it matters, and what is out of scope. Independent pieces dispatch together, in one message.
+ 2. Verify each return against what its prompt asked. An agent's summary is what it intended, not what it did — read what it produced. Nothing reaches the user unverified.
3. Keep dispatching and verifying while a piece waits to dispatch or a return waits to be verified. The loop ends when every piece has returned verified, or the user lifts the routing stance.
- ## Output
-
- The verified pieces are the result — hand them to the user. If the loop ended because the user lifted the routing stance, hand back what stands: each piece that passed its check as verified, and any return not yet checked named as unverified, never claimed as verified.
-
- Then check whether anything still unresolved would change what gets built:
-
- - Nothing would → return to the work the routing interrupted.
- - A claim about a returned piece needs a committed verdict → use **judge** skill.
- - The user is missing something already settled → use **explain** skill.
+ Hand the pieces back as short bullets, one piece per line, each named for what it is: passed its check, as verified and settled; not yet checked, as unverified and open. A return reaches the user under one of those two names.
- The routing is spent once the loop ends — anything further runs under the branch picked here, never as another round of routing.
+ The work is one piece — a single pass, or a plain answer → do it yourself. A piece only this session can do → stop. Verifying a piece's return would mean redoing the piece → stop. The work will run unwatched, optimizing toward something, and how it's judged is undefined before it runs → use **target** skill.