Skip to main content

Completing the ai-sdk removal

Status: done. All 8 items landed. Item 7 — deleting GenerationHandler — was the last, in 47717a665. ai and @ai-sdk/* are gone from package.json and pnpm-lock.yaml, and scripts/check-banned-deps.ts keeps them out.

What is left, how each piece is solved, and the order forced by their dependencies. Every claim here was checked against the source or the installed package, not inferred.

The rule that governs all of it​

Run test:providers-mocked on every step, not just the live matrix. The live matrix has 40 cells across six providers and it passed a change that broke ten of them, because every provider reachable from this machine supports streaming and the mocked gate is the only thing that serves a non-streaming body. The live matrix proves behaviour; the mocked gate proves the wire.

1. Anthropic native generate​

createAnthropicLoopAdapter hardcodes stream: true when it calls messages.create (loopAdapter.ts:120), which is why routing generate through it changed the wire.

Solution is the shape already proven for the OpenAI-compatible family: loop over the provider's own delegating-model doGenerate, which issues a non-streaming messages.create. Wrap each step in withProviderRetry, funnel failures through handleProviderError, run the turn inside runGenerateWithModelFallback (now protected), and call fireGenerateOnFinish.

Carry over the two fixes found the first time. A schema arriving with no tools must declare final_result as the turn's only tool and pin tool_choice to it, because appendFinalResultTool declines an empty tool list. And onFinish must be fired explicitly.

2. SageMaker​

SageMakerLanguageModel.doGenerate makes one invokeEndpoint call and already returns toolCalls; no streaming is involved, so the wire hazard does not apply. Same loop shape as above.

This machine has no SageMaker endpoint or credentials. Its single-step behaviour is identical by construction because it is the same doGenerate call; the multi-step branch is the new code and needs a live endpoint before it is trusted. Say so in the commit rather than implying coverage.

3. Guardrails filter and video-analysis formatting​

Both want a single no-tool turn. generateOnceNative did this cleanly before the revert took it with everything else: it calls doGenerate directly and reads the text out of the v3 content array. Re-apply unchanged.

4. wrapLanguageModel​

Upstream is about fifty lines: reverse the middleware array, reduce it, and return an object that keeps specificationVersion, provider, modelId and supportedUrls while routing doGenerate and doStream through transformParams plus the optional wrapGenerate / wrapStream hooks. One consumer, middleware/factory.ts. Reimplement directly.

Record while doing it that wrapStream never runs today, because every streaming path is native and bypasses the wrapped model.

5. The tool and schema type algebra​

This is the one that must move as a unit, and the reason the first attempt failed: tool() was replaced while Tool still came from ai, so the replacement had to satisfy a type it no longer matched.

The algebra is small and fully specified in the installed package:

  • Schema<T> is { [schemaSymbol]: true; _type: T; validate?; jsonSchema }.
  • FlexibleSchema<S> is the union of Schema, LazySchema, ZodSchema and StandardSchema.
  • InferSchema<S> is a conditional chain over those four.
  • Tool<INPUT, OUTPUT> carries inputSchema: FlexibleSchema<INPUT>, which is what ties execute's first parameter to the schema.

Declare all of it in src/lib/types/, repoint types/tools.ts at the local declarations, and only then implement tool, jsonSchema and stepCountIs in utils/tool.ts. tool() is identity upstream and stepCountIs(n) is ({steps}) => steps.length === n; the work is entirely in the types.

jsonSchema() should keep stamping Symbol.for("vercel.ai.schema"). Nothing in this repo reads it — convertZodToJsonSchema looks for the plain jsonSchema property — but keeping it costs nothing and preserves recognition by anything that does.

6. The rest of the public type surface​

Three files hand-declare structurally: conversation.ts (the message and part types), providers.ts (the model, usage and finish-reason types) and middleware.ts (the middleware contract plus the LanguageModelV3 protocol types from @ai-sdk/provider).

The non-obvious one is dist/files/fileTools.d.ts, which embeds inferred import("ai").Tool<...> references. No source edit removes those; they disappear only once Tool itself is local. Verify by hiding the package and re-running a consumer typecheck, which is how the leak was confirmed in the first place.

7. GenerationHandler​

1409 lines, and its executeGeneration is unreachable once Anthropic and SageMaker are native. Four result-formatting helpers on it are still called from BaseProvider, but they take the generateText result shape, so they die with it. Confirm by trapping the seam and running the full matrix plus the mocked gate, then delete. Output and stepCountIs lose their last consumers here.

8. Browser bundle​

src/browser/entry.ts still re-exports generateText, streamText, generateObject and streamObject through the seam. No consumer was found for them. Either drop them or map them onto NeuroLink's own equivalents; this is a public-subpath decision, not a technical one.

Order​

Type algebra last among the type work but before the package drop; everything else is independent.

  1. Anthropic, SageMaker, guardrails and video-analysis — the last runtime callers of generateText.
  2. wrapLanguageModel.
  3. Delete the dead GenerationHandler path.
  4. The tool/schema algebra together with types/tools.ts and utils/tool.ts.
  5. The remaining public types.
  6. Browser re-exports.
  7. Drop ai and @ai-sdk/provider, extend BANNED_PACKAGES, and rewrite the documentation snippets that still tell users to import from ai directly.