Skip to content

Models & tools

Partial

Typed reasoning stages, reservation and admission exist. A general dynamic tool registry is not implemented.

Code and tests

The implemented reasoning interface has typed research, synthesis and assessment requests. Product-worker validates their outputs and controls execution. A general dynamic tool registry with arbitrary effect classes and child agents remains a target.

A generated plan is data from an external provider. capabilities/planner.go decodes and validates it before the worker issues retrieval requests. model.go defines answer and assessment shapes; evidence.go checks cited sources. Tests alongside those files exercise malformed plans, missing evidence and invalid assessments.

Provider routing and paid-call uncertainty are documented once in Model gateway. The gateway’s named structured-output tool is a provider response mechanism, not a permission to execute arbitrary system tools.

Credit reservation lives in services/product-worker/internal/capabilities/accounting.go and apps/kesita-app/platform/db/aix-accounting.sql. Reserving before dispatch prevents concurrent stages from each spending the same remaining task budget. accounting_test.go and the AIX lifecycle integration tests cover settlement and reservation behavior.

Provider admission is a separate limit, implemented in capabilities/admission.go. It controls concurrent calls by provider/lane across replicas; it does not increase a task’s credit allowance. TestAIXProviderQueueCancellation verifies cancellation while waiting.

The broader proposal for child tasks sharing protected completion allowances is not fully implemented. It should not be inferred from the presence of a credit ledger. Exact prices and bounds belong in modelgateway/pricing.go, the SQL/accounting implementation and their tests.