fix(mistral): forward stopSequences as Mistral native stop parameter

[cristiandrei1234]

Background

@ai-sdk/mistral accepts stopSequences on the public generateText / streamText API but silently drops the value, emitting only a warning:

[{ "type": "unsupported", "feature": "stopSequences" }]

The Mistral REST API itself supports stop sequences natively via the stop field, which accepts a string or array of strings (Mistral API reference → stop: string | array<string>). The adapter's job is to translate stopSequences → stop in the outgoing request, but that translation is missing.

Result: a documented, standard generateText parameter does nothing for Mistral users, with no error surfaced beyond the warning array.

This is an isolated regression specific to @ai-sdk/mistral. Five sibling providers already implement the exact same translation:

Provider	Source
@ai-sdk/openai	packages/openai/src/chat/openai-chat-language-model.ts:185
@ai-sdk/openai-compatible	packages/openai-compatible/src/chat/openai-compatible-chat-language-model.ts:295
@ai-sdk/groq	packages/groq/src/groq-chat-language-model.ts:146
@ai-sdk/deepseek	packages/deepseek/src/chat/deepseek-chat-language-model.ts:176
@ai-sdk/alibaba	packages/alibaba/src/alibaba-chat-language-model.ts:128

All of them do literally stop: stopSequences, in their request baseArgs. Mistral is the odd one out.

Summary

packages/mistral/src/mistral-chat-language-model.ts:

• Removed the if (stopSequences != null) warnings.push({ type: 'unsupported', feature: 'stopSequences' }) branch.
• Added stop: stopSequences, to baseArgs, in the same position used by sibling providers (between top_p and random_seed).

The change is minimal and matches an established intra-repo pattern. Vercel's stopSequences?: string[] is a valid subtype of Mistral's documented stop: string | array<string>, so direct assignment is correct.

Manual Verification

Unit tests (deterministic, in-process)

Added a regression test in mistral-chat-language-model.test.ts that:

1. calls doGenerate with stopSequences: ['foo', 'bar']
2. asserts the outgoing request body equals { model, messages, stop: ['foo', 'bar'] }
3. asserts result.warnings does NOT contain { type: 'unsupported', feature: 'stopSequences' }

Validated against the unfixed branch (git stash the source change): the new test fails as expected, plus 7 inline-snapshot tests fail because the snapshots are updated to reflect the new stop field in baseArgs. With the fix applied: all 72 mistral package tests pass.

Live API against api.mistral.ai

Ran a small script that imports the local built @ai-sdk/mistral (with the fix applied) and calls doGenerate on mistral-small-latest with stopSequences: ['Pacific'], prompt asking for facts about the Pacific Ocean:

--- request.body (raw payload sent) ---
{
  model: 'mistral-small-latest',
  max_tokens: 200,
  stop: [ 'Pacific' ],
  messages: [...]
}

--- warnings ---
[]

--- finishReason ---
{ unified: 'stop', raw: 'stop' }

--- content (extracted text) ---
"**Fact 1.** The "

Mistral truncates the output mid-token immediately before "Pacific" and returns finish_reason: "stop", confirming both the field name and the end-to-end behavior. This matches the curl-only verification the reporter already did in #15194 against mistral-large-latest.

Checklist

• All commits are signed (PRs with unsigned commits cannot be merged)
• Tests have been added / updated (for bug fixes / features)
• Documentation has been added / updated (for bug fixes / features) — JSDoc unchanged; behavior now matches existing stopSequences documentation in content/docs/.../generate-text#stop-sequences
• A patch changeset for relevant packages has been added (for bug fixes / features - run pnpm changeset in the project root)
• I have reviewed this pull request (self-review)

Future Work

The reporter flagged that other adapters may emit { type: 'unsupported', feature: 'stopSequences' } while their underlying API actually supports it under a different field name. A repo-wide grep finds 7 other call sites with the same dropping pattern:

packages/xai/src/xai-chat-language-model.ts:123
packages/xai/src/responses/xai-responses-language-model.ts:113
packages/perplexity/src/perplexity-language-model.ts:92
packages/huggingface/src/responses/huggingface-responses-language-model.ts:108
packages/openai/src/responses/openai-responses-language-model.ts:177
packages/open-responses/src/responses/open-responses-language-model.ts:97
packages/prodia/src/prodia-language-model.ts:81  // image-gen, not applicable

Each requires individual verification against the respective provider's API spec — left out of scope here to keep this PR narrow and easy to review. Happy to follow up with separate PRs for providers whose REST APIs actually honor stop sequences (xAI and Perplexity look like obvious candidates given their OpenAI-compatible designs).

Related Issues

Fixes #15194.

[aayush-kapoor]

@cristiandrei1234 can you attach the repro script you used to verify the fix? ideally it produced the warning before and then worked after

[cristiandrei1234]

I'd already deleted the original ad-hoc script after the verification run, so I rewrote a fresh one against the local build of @ai-sdk/mistral and ran it in both states. Same prompt, same stopSequences: ['Pacific'], the only thing that changes between runs is whether the fix is applied to mistral-chat-language-model.ts (toggled via git checkout HEAD~1 -- ... + pnpm -F @ai-sdk/mistral build).

Gist: https://gist.github.com/cristiandrei1234/ebf0b79e757c3aa470492774a8a34505

Without the fix

--- request.body (raw payload sent to api.mistral.ai) ---
{
"model": "mistral-small-latest",
"max_tokens": 200,
"messages": [
{ "role": "user", "content": [{ "type": "text", "text": "List five facts
about the Pacific Ocean. Number them 1 through 5. Be concise." }] }
]
}

--- warnings ---
[ { "type": "unsupported", "feature": "stopSequences" } ]

--- finishReason ---
{ unified: 'stop', raw: 'stop' }

--- content text ---
"1. Largest Ocean: The Pacific Ocean is the largest and deepest of Earth's
five oceans, covering about 30% of the planet's surface.

2. Mariana Trench: It contains the Mariana Trench, the deepest known
   point on Earth (Challenger Deep, ~10,984 meters / 36,037 feet deep).

3. Ring of Fire: The Pacific is bordered by the Ring of Fire, a major
   area of volcanic activity and earthquakes.

4. Home to Many Species: It supports more than half of the world's marine
   species, including coral reefs and whales.

5. Oldest Water: Some of its deepest waters are over 1,000 years old due
   to slow circulation."

No stop field in the request, the warning is emitted, and the model runs to

5. Oldest Water: Some of its deepest waters are over 1,000 years old due to slow circulation."

No stop field in the request, the warning is emitted, and the model runs to completion — Pacific appears twice in the output, untruncated. The stop finishReason here is just the natural end-of-message, not a stop-sequence hit.

With the fix applied

--- request.body (raw payload sent to api.mistral.ai) ---
{
"model": "mistral-small-latest",
"max_tokens": 200,
"stop": [ "Pacific" ],
"messages": [
{ "role": "user", "content": [{ "type": "text", "text": "List five facts about the Pacific Ocean. Number them 1 through 5. Be concise." }] }
]
}

--- warnings ---
[]

--- finishReason ---
{ unified: 'stop', raw: 'stop' }

--- content text ---
"1. Largest Ocean: The "

stop: ['Pacific'] is on the wire, no warning, and Mistral truncates mid-token immediately before the stop word with finish_reason: "stop". Happy to fold this into the PR description if you'd like it on the body rather than buried in a comment.

[github-actions]

⚠️ Backport to release-v6.0 created but has conflicts: #15293