Take Agent Discovery Turn

One conversational turn of the kickoff. THIS ONE COSTS MONEY.

The only route on this surface that makes an LLM call, which is why it is tiered billable and why the app rate-limits it at 20 requests a minute per organisation - far above a human conversation, far below anything that shows up on a bill. Over that, the app answers 429 with Retry-After.

The model replies AND states what it learned in one forced tool call, so a message that answers four questions records four fields. Everything it claims is validated before it is written: a key outside the framework is dropped, and a green with no supporting quote is DOWNGRADED to amber rather than trusted.

The turn writes only to the discovery record. It provisions nothing, calls no tools, and sends nothing.

There is no history parameter and that is deliberate - see the module docstring.