pi-neuralwatt: v0.17.0
Minor Changes
- 53c4e26: Add
provider.apiBaseUrlconfig. It sets the base URL for the extension’s catalog and quota requests and is the fallback for models that carry nobaseUrlof their own — models.json overrides, custom providers, and proxy routing all take precedence. Defaults to the direct upstream. - bbca292: Migrate to pi 1.0.0. The provider extension now injects a per-request
options.fetchwrapper (pi-ai >= 0.83) instead of swappingglobalThis.fetch, so 429 rate-limit headers are still captured and successful SSE bodies are still teed for live quota comments without touching the global fetch. Stream signatures useTranscriptContext, the model catalog narrowsProviderModelConfigto its chat member (ProviderChatModelConfig), and the peer floor moves to@earendil-works/pi-ai/@earendil-works/pi-coding-agent>=0.86.0. No user-facing behavior change. - 99b9ee8: Support Neuralwatt decision models as Pi classifiers. Key-scoped
/v1/modelscatalogs now compiletask: "decision"entries (e.g.clef-flash) into classifier models stamped with thetypesafe-system-oneclassifier API, served through a new System One module that posts{ model, state, questions }to/v1/systemoneand returns pi-aiClassifierResultanswers with per-choice probabilities. Classifiers appear ingetAllModels()(andmodels.classify()/models.getAvailableOfType("classifier")) but never in the chat-onlygetModels(), so/modelis unaffected. The offline fallback catalog stays chat-only. Requires pi 1.0.0: the peer floor for@earendil-works/pi-aiand@earendil-works/pi-coding-agentmoves to>=1.0.0.
Patch Changes
- 3ccee9c: Refetch the model catalog when
provider.apiBaseUrlchanges. The cached catalog was keyed only by auth scope, so after switching to a gateway pi kept the upstream catalog for up to four hours and missed models only the gateway serves, such as theclef-flashclassifier. - a8f8f15: Fix provider loading on pi 1.1.0 by replacing the deep pi-ai import with a minimal Neuralwatt System One client. Restore classifier registration.