Expand description
Running a task: the call, repairs with the exact error, votes, escalation, records.
One run is: resolve the instructions, build the request from the task’s messages
and schema and its kind’s profile, reserve the call, send it, parse and check the
answer. A structurally wrong answer goes back to the same model with the error,
up to the profile’s repairs. Votes run concurrently and are compared with the
task’s own agree; a vote without a strict majority escalates, asks or fails, as
the profile says. Every call leaves a record in the scope.
Structs§
- Record
Policy - What a turn keeps of each call beyond the parsed answer.
- Task
Call - Where a task call sits: its identifier, its parent, and how deep its chain already is.
- Task
Engine - Runs model tasks. Cheap to clone; shared by every turn.
- Task
Engine Builder - Builds a
TaskEngine. - Task
Scope - The spending and the records of one phase of one turn.
Enums§
- Task
Failure - Why a task produced no answer to use.
- Task
Outcome - What a task run produced, and how deep its chain ended.
Constants§
- TASK_
LABEL - The request metadata key carrying the task id a call runs as:
u2/extract#repair1. - TURN_
LABEL - The request metadata key carrying the turn a call belongs to, when the scope names it.