Skip to main content

AI configuration, inventory, and cost

Availability: provider configuration is descriptor-driven. The browser supplies no key, provider response, or direct model invocation.

Configure an OpenAI-compatible provider​

  1. Register an opaque credential locator in Access & security, for example env:CASEWEAVER_OPENROUTER_KEY when that name is available to the server secret backend. It is a reference, not an API key value.
  2. In AI configuration, create a provider draft from the registered OpenAI-compatible descriptor. Use an HTTPS API base endpoint and choose its immutable protocol mode (embeddings, chat completions, or responses) deliberately.
  3. Review and activate the draft. Activation makes a safe provider identity eligible for the next server-owned step; it does not prove every model is available.
  4. Choose Refresh models available from provider. The server calls the provider and persists a bounded, workspace-scoped inventory tied to that exact provider version.
  5. Create an immutable binding from a model in that inventory, select its CaseWeaver role, then configure price/budget policy and use the guarded capability test.

The trusted LiteLLM catalog is a separate pricing/capability source. Refreshing it does not make a global catalog model available from a provider. An exact price match enriches the provider inventory; an unmatched model has unknown price. Unknown price is never zero and cannot satisfy a hard-budget test without a deliberate, complete pricing override for that same provider-owned model.

Use an embeddings instance for embedding work and a distinct chat/responses instance for generation when the provider requires it. The capability test is server-owned, confirmation-bound, rate-limited, budget-gated, and executed only through @caseweaver/ai-execution.

Binding, budget, and history​

A binding selects a provider model for a role; a role default chooses an active binding; a collection/profile/recipe pins a binding version for work. Price components must cover the role's units: input tokens for embeddings/reranking, input and output tokens for generation, and image units for vision. Costs and unknown-price states are queryable, but provider credentials and raw account/model data are not.

The optional Copilot SDK BYOK repository-agent descriptor remains a separate, policy-bounded repository runtime. It is not a substitute for an embedding or chat binding and is never called directly by a feature package.