Coldstart has not been released. The behavior on this page runs in the development build; there is no release you can install today.
Before you ask the coding agent anything, it already has a small set of labels describing the commands, routers, and agents it can use. Those labels are the resident floor: short descriptions present in every session so the right deeper guide can be chosen later.
The floor is a catalogue, not the books. The full procedure behind a command, router, or delegated agent stays on disk until selected.
The ordinary alternative
For a small project, one short instruction file may be clearer than a routing system. You keep the few rules that matter in one place, and the agent reads them at the start. The trouble begins when specialist methods accumulate. Putting all of them in the opening context makes an export task pay attention to guidance for work it is not doing.
Coldstart puts only discovery text on the floor. A description says what something is for and when it may apply; it does not try to carry the method itself. In the export session, the floor can point toward routing without preloading a long catalogue of unrelated techniques.
What is checked
The product reads the same description text for measurement and selection. That avoids one reader measuring a label while another routes on a different interpretation of it.
A single description has a size ceiling. If it grows large enough to behave like an instruction body, or cannot be read, the check fails instead of silently dropping it from discovery. Growth in the whole roster is measured against a baseline and reported for human review; it is not rejected merely because the product gained another legitimate surface. Both numbers — the ceiling on one description and the baseline for the whole roster — are values the development build carried at the revision reviewed here, and both are expected to change.
That distinction matters:
| Observation | What follows | What does not follow |
|---|---|---|
| One description exceeds its ceiling | The catalogue entry is carrying too much text | The underlying method is poor |
| A description cannot be read | Discovery and measurement would omit it | The router deliberately chose not to match |
| The total floor grew | Every session now carries more discovery text | The growth is automatically harmful |
The nearest limit
The floor can make guidance discoverable without making it good. A short, readable description can misstate its body or fail to fire for the situation it names. Measurements expose size and parsing; they do not prove relevance, honesty, or better outcomes.
If your project has only a handful of stable instructions, a single maintained file remains enough. The resident floor earns its extra structure when distinct methods need to stay dormant but still findable.
Next
That is what is always present. What happens when the session needs something that is not: routed capability.