Router
Router lays out which skills a task needs and in what order, as far as that can be known now, and stops where the next step depends on a result that does not exist yet. It belongs to the Sensemaking domain.
When to use it
- The next skill is not obvious.
- The work spans several skills.
- Part of the request may already be done, and you want only the steps still needed.
When to use something else
When one skill obviously fits, use it directly; Router stays out of the way.
How to start it
The suite marks Router as model-invoked: the agent starts it when a task matches, and it is not meant to be started by name. Ask for the outcome in plain words. For example (illustrative):
We have to move our job queue from Redis to Postgres without losing an in-flight job, and I cannot tell whether that starts with a decision, a plan, or a ticket. Work out which skills I need, and in what order.
What happens
Router reads the current state, drops work that is already done, and lists only the skills still needed. Each step is a concrete skill, never a domain. A route that stops early is a correct answer, not a partial one. Router never runs a routed skill, plans, edits source, commits, or opens a pull request.
What you get
A sealed route.md in the ignored folder .first-four-hours/routes/<slug>/. Each step names a skill, why it is needed now, its input, its expected result, and a ready-to-copy prompt. When the current state already meets the target, the route is a single None: line instead. Router ends by offering step 1's prompt, or no next step for a None: route.
When it stops
Router stops at the first step whose need depends on an unmade decision, an unknown result, or a fact not yet produced. Running out of steps returns control to you; it does not mean the whole target is done, and a later state may need a fresh route.
Why it's built this way
The failure it prevents. A plan written before the first result arrives is mostly guesses dressed as steps. When the next skill is unclear, an agent either picks one and drifts, or lays out a long plan that is wrong as soon as the first result comes in. Router shows the path as far as it can be known now: the longest stretch with no open question in it, not just the next step and not the whole journey.
What it refuses, and why. It never guesses past a step whose need depends on an unmade decision, an unknown result, or a fact not yet produced; a route that stops there is a correct answer. It never invents a step to fill the path, never runs a routed skill, and keeps no loops or history, so changed circumstances get a fresh route. Its validation can reject an unsafe or impossible route, but never adds, removes, or reorders steps, or decides where the route stops.
Floor and ceiling. A weaker agent drops work that is already done and gives each remaining step a reason, an input, an expected result, and a ready prompt. A stronger agent stays out of the way when one skill obviously fits, and when a short sequence is already fully determined, it shows all of it.
Related skills
Only your later yes starts a step. You run the prompt for a user-invoked skill, such as Diagnose or Ticket, yourself. See How skills hand work to each other.