OptionalcallOptionaleffortOptionalmaxOptionalmodelPer-purpose model overrides. Keys are purpose names (e.g., "fkPruning"). Each override can specify a different model, temperature, maxTokens, and effortLevel. If a purpose is not listed, the default model/settings from the parent AIConfig are used.
OptionalpricingOptionalrateRate limiting configuration for API calls
OptionalrequestsOptionalretryRetry configuration for failed API calls (429, network errors)
Optionaltemperature
Wall-clock ceiling, in milliseconds, on EACH model call.
0disables the bound. Default: 120000 (two minutes).Every LLM call in this package was previously unbounded. The retry machinery below cannot help: it only engages once a promise settles, and a provider that accepts the socket and then stops sending never settles one.
maxTokensbounds the RESPONSE, not the wait.The bound is applied two ways at once, mirroring how
AIPromptRunnerbounds its own calls: the request is aborted throughChatParams.cancellationToken(which the provider drivers forward to their SDKs, so the socket is torn down rather than abandoned) AND the awaited promise is rejected, so a driver that ignores the signal still cannot park a caller forever.