0002. Mistral-native provider for pi

• Status: Trial
• Date: 2026-10-30

Context

pi’s built-in Mistral paths fall short in four documented ways: the generic OpenAI-completions adapter cannot parse the block-array thinking content that Mistral’s chat-completions API streams, the Le Chat conversations API runs on stateful beta conversation state, the model list is static and hand-maintained, and none of Mistral’s non-chat surface (knowledge/RAG, OCR, connectors, server-side tools) is reachable. The interim fix in pi’s ecosystem was a hand-patch applied into node_modules — fragile and lost on every pi update. pi’s provider-extension API offers a supported native path: a registered provider with a custom api string and its own streamSimple wire handler.

Decision

Ship this package as a Mistral-native provider extension for pi: a Mistral-specific, optimized, full-feature-surface integration. The extension registers a native pi-ai provider (buildMistralNativeProvider, api mistral-chat) implementing the block-array chat-completions wire format against the vendored OpenAPI spec (0003), with dynamic model-catalog refresh and the whole non-chat surface exposed as pi tools. Distribute it as a pi npm package (pi install npm:pi-mistral-native).

Consequences

• Provider behavior survives pi updates (the extension is loaded, not patched in), and is versioned, reviewed and tested like any code.
• The wire format is our own implementation: it must be kept current with Mistral’s spec ourselves and covered with tests.
• streamSimple owns request conversion, streaming events, usage and cost accounting, cancellation and error normalization per pi’s custom-provider contract — more surface than delegating to a built-in API implementation, but no built-in implementation speaks the block-array chat-completions format with thinking content.