deps(ai): migrate the Vercel AI SDK from 5 to 6

[mrsibe]

What

Upgrade the AI SDK from 5 to 6, keeping the runtime behaviour identical.

Package	Before	After
ai	5.0.265	6.0.297
@ai-sdk/provider	2.0.5	3.0.18
@ai-sdk/openai	2.0.129	3.0.121
@ai-sdk/anthropic	2.0.105	3.0.125
@ai-sdk/google	2.0.99	3.0.129
@ai-sdk/openai-compatible	1.0.56	2.0.80

This is the second step of the two-stage 5 → 6 → 7 plan: it lands the provider-spec V2 → V3 migration on its own, so the semantics change in 7 does not hide inside a double jump.

How

Ran the official codemod npx @ai-sdk/codemod@3.0.8 v6, then fixed the three things it cannot know:

1. The test fixtures, not just the types. The codemod renamed MockLanguageModelV2 → MockLanguageModelV3 but left the scripted provider chunks as V2 stream parts, so a V3 mock received a string finishReason and flat usage numbers. The suite caught this — 10 tests failed — and the fixtures now build the V3 shape:

   finishReason: { unified, raw }
   usage: { inputTokens: { total, noCache, cacheRead, cacheWrite },
            outputTokens: { total, text, reasoning } }

   This is the part of the migration a type-only change would have shipped silently: the as never[] cast on those fixtures is why it typechecked while being wrong.

2. Embedding entry points — textEmbedding → embedding, textEmbeddingModel → embeddingModel.

3. A misleading codemod rename. The local convertToCoreMessages in ModelClient is a plain mapper, not the SDK function the codemod named it after. Renamed to toModelMessages with the stale CoreMessage comment corrected.

The one real behaviour change, and how it is prevented

AI SDK 6 flips OpenAI's strictJsonSchema default from false to true. QuizService, AnkiCardService and MindMapService build schemas with .optional() fields, and OpenAI strict mode requires every property to be listed in required.

I verified both halves of this at the request-body level rather than assuming:

# the zod schema conversion leaves optional fields out of `required`
required at root: [ 'title', 'items' ]      # 'tags', 'meta', 'hint' absent

# the same call, with and without the fix
responses default                -> text.format.strict = true
responses strictJsonSchema:false -> text.format.strict = false

So without a fix, openai-responses would start rejecting the generated quiz / Anki / mind-map schemas. ProtocolAdapter and ModelClient now carry structuredOutputProviderOptions(), the openai-responses adapter answers it, and the three services pass it to streamObject.

Only openai-responses needs it. openai-completions goes through @ai-sdk/openai-compatible, which sends response_format: { type: 'json_object' } unless structuredOutputs is enabled, so its strictJsonSchema never applies — I checked the provider source, not the guide.

Verification (Linux x64, WSL2)

Command	Result
npm run typecheck	clean
npm test	470/470 pass (469 + the strictness regression test)
npm run build:unpack && npm run smoke:packaged	29 checks pass

The packaged smoke matters here because the AI SDK is a devDependency that rollup bundles into out/main, so a bad upgrade can break the bundle, not just the types.

Deliberately not changed

• streamObject is deprecated in 6 (in favour of streamText + Output.object) but still works. Rewriting three services does not belong in a version bump; it is the natural content of the 6 → 7 PR.
• The 'openai-compatible' providerOptions key is deprecated in favour of 'openaiCompatible'. It still works, including [Feat] Reasoning models hit the output ceiling: make truncation recoverable #179's reasoning-effort path, but now logs a warning. Flagged rather than renamed so this PR stays behaviour-identical.

Rollback

Revert the branch. No schema, storage format, or persisted data changes; the provider-spec V2 → V3 change is internal to the SDK.

Stacking

The 6 → 7 PR will stack on this branch (base deps/ai-sdk-6), because the v7 codemod can only run on v6 code.