fix(mistral): bump @ai-sdk/mistral to 3.0.64 to unbreak streamed tool calls - #49689

Issue for this PR

Closes #49692

Type of change

• Bug fix
• New feature
• Refactor / code improvement
• Documentation

What does this PR do?

Bumps @ai-sdk/mistral from 3.0.51 to 3.0.64 so that streamed tool calls work on the GLM models Mistral relays (zai-glm-5, zai-glm-5-2, zai-glm-5-3, zai-glm-latest). Today they abort on the first tool call.

The pinned 3.0.51 declares id: z.string() for every entry in a tool_calls delta. That is stricter than the wire format: Mistral sends id and function.name only in the first delta for a given index, and continuation chunks carry arguments fragments alone — clients are meant to accumulate by index. So the chunk that continues a tool call fails validation, the arguments never get assembled, and the tool is invoked with an empty object.

This only surfaces on the relayed GLM models because Mistral's own models don't fragment tool-call arguments — they return the whole tool call in one chunk, id included, so the continuation path is never taken. That is also why the bump is safe for them: their behaviour doesn't change.

3.0.59 fixed this upstream by relaxing the schema to id: z.string().nullish() and routing deltas through StreamingToolCallTracker, which falls back to toolCallsByIndex when the id is absent. I picked 3.0.64 as the latest release on the 3.x line this repo already tracks; 3.0.59 is the minimum that works.

The lockfile change is the one resolution entry plus its dependency pins (@ai-sdk/[email protected], which is where the tracker lives).

How did you verify your code works?

Reproduced the failure on 1.18.29, then pinned the new version locally via a config override ("npm": "@ai-sdk/[email protected]"), which exercises the same code path as this diff:

$ opencode run --model mistral/zai-glm-5-3 "read README.md, CONTEXT.md and list packages/, then summarize"
> build · zai-glm-5-3
→ Read packages
→ Read README.md
→ Read CONTEXT.md
BMT is a platform for behavioral-economics studies: …

Several consecutive tool calls, clean finish. Same check on zai-glm-5-2. I also confirmed against the raw SSE stream that mistral-large-latest, mistral-medium-latest, magistral-medium-latest and codestral-latest emit tool calls in a single chunk with id set, so they never hit the changed path.

Screenshots / recordings

n/a — not a UI change.

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR