Prompt caching for Mistral is silently disabled: @ai-sdk/mistral 3.x drops promptCacheKey, and 4.x requires a spec-v4 core

Summary

Mistral's API supports explicit prompt caching via a prompt_cache_key body parameter (cached tokens billed at 10% of the input price, hits reported in usage.prompt_tokens_details.cached_tokens). opencode already does the right thing on its side — packages/opencode/src/provider/transform.ts (~line 1321) sets result["promptCacheKey"] = input.sessionID for @ai-sdk/mistral — but the parameter never reaches the wire, and today no provider combination can enable caching for Mistral reasoning models.

This affects mistral-large-latest, mistral-medium-*, magistral-*, and the relayed GLM models (zai-glm-5, zai-glm-5-2, zai-glm-5-3, zai-glm-latest) — i.e. any workload where long conversation histories are resent every turn, which is exactly the workload prompt caching is meant for.

The blocker chain (verified)

1. @ai-sdk/mistral 3.x silently drops promptCacheKey. The chat language model parses per-call provider options strictly against mistralLanguageModelOptions (safePrompt, structuredOutputs, strictJsonSchema, documentImageLimit, documentPageLimit, parallelToolCalls, reasoningEffort...). promptCacheKey is not in that schema, so it is stripped by parseProviderOptions and prompt_cache_key is never included in the request body. opencode pins @ai-sdk/[email protected] (bundled 3.0.51 in the current desktop app; fix(mistral): bump @ai-sdk/mistral to 3.0.64 to unbreak streamed tool calls #49689 bumps to 3.0.64 — same limitation).

2. @ai-sdk/mistral 4.x fixes it but cannot be loaded. As of @ai-sdk/[email protected], mistralLanguageModelOptions includes promptCacheKey: z.string().optional() and getArgs maps it to prompt_cache_key in the body. However, 4.x declares specificationVersion = "v4", and opencode's [email protected] core rejects anything other than v2/v3 in resolveLanguageModel() with UnsupportedModelVersionError (the v2-compat shim only handles "v2" and "v3"). So pinning "npm": "@ai-sdk/[email protected]" via config fails at model resolution.

3. @ai-sdk/openai-compatible can inject prompt_cache_key but breaks on Mistral reasoning models. The unknown-key passthrough (providerOptions[providerOptionsName] entries not in openaiCompatibleLanguageModelChatOptions.shape go verbatim into the request body) makes "options": { "prompt_cache_key": ... } work at the config level. But for reasoning models:

   • History: assistant reasoning parts are serialized as a flat reasoning_content field on the message. Mistral's API rejects it with HTTP 422: {"detail":[{"type":"extra_forbidden","loc":["body","messages",1,"assistant","reasoning_content"],"msg":"Extra inputs are not permitted"}]}. This 422s every multi-turn session that contains reasoning. By contrast, @ai-sdk/mistral serializes history reasoning as native thinking content blocks ({"type":"thinking","thinking":[{"type":"text",...}],"closed":true} inside the content array), which Mistral accepts.
   • Stream chunks: Mistral reasons GLM/Magistral through native thinking content blocks in streamed chunk content arrays; @ai-sdk/[email protected] declares content: z.string() and throws AI_TypeValidationError: expected string, received array. Fixed upstream in 2.0.75+ (chunk content is now union(string, array<looseObject>) with thinking-block → reasoning conversion), see Bump bundled @ai-sdk/openai-compatible 2.0.41 -> 2.0.75 (fixes fragmented reasoning streams) #50290.

Evidence (reproduced against https://api.eu.mistral.ai/v1, model zai-glm-5-3)

# prompt_cache_key works: two identical requests
# 1st: usage.prompt_tokens_details.cached_tokens = 0
# 2nd: usage.prompt_tokens_details.cached_tokens = 640 (of 751 prompt tokens)

# history reasoning as reasoning_content -> 422
{"detail":[{"type":"extra_forbidden","loc":["body","messages",1,"assistant","reasoning_content"],"msg":"Extra inputs are not permitted"}]}

# history reasoning as native thinking content blocks -> 200

Ask

Upgrade the ai core to a spec-v4-capable release (AI SDK line that accepts LanguageModelV4) and bump @ai-sdk/mistral to the 4.x line (4.0.48 verified: sends prompt_cache_key = session ID, i.e. the per-conversation key Mistral's docs recommend). That would turn the existing promptCacheKey = input.sessionID in transform.ts into working, per-session prompt caching for all Mistral models, with no config changes required by users.

Related: #49692 / #49689 (GLM streamed tool-call fix, @ai-sdk/mistral 3.0.64), #50290 (bundled @ai-sdk/openai-compatible lag), #50338 (v2/v3 provider handling).