Comparing changes

@ai-sdk-factory

…tered clients after authentication errors. (#21367) This is an automated backport of #21029 to the release-v6.0 branch. FYI @ai-sdk-factory[bot] This backport has conflicts that need to be resolved manually. ### `git cherry-pick` output ``` Auto-merging content/docs/03-ai-sdk-core/16-mcp-tools.mdx CONFLICT (content): Merge conflict in content/docs/03-ai-sdk-core/16-mcp-tools.mdx Auto-merging packages/mcp/src/tool/oauth.test.ts CONFLICT (content): Merge conflict in packages/mcp/src/tool/oauth.test.ts Auto-merging packages/mcp/src/tool/oauth.ts error: could not apply 7fc2bc1... fix: Prevent MCP OAuth from deleting or replacing pre-registered clients after authentication errors. (#21029) hint: After resolving the conflicts, mark them with hint: "git add/rm <pathspec>", then run hint: "git cherry-pick --continue". hint: You can instead skip this commit with "git cherry-pick --skip". hint: To abort and get back to the state before "git cherry-pick", hint: run "git cherry-pick --abort". hint: Disable this message with "git config set advice.mergeConflict false" ``` Fixes #21020 --------- Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: Aayush Kapoor <aayushkapoor34@gmail.com>

… UI messages (#21335) ## Background UI message validation returned original metadata and custom-data values after successful schema parsing, losing defaults, coercions, and transformations and causing runtime values to contradict their TypeScript types. ## Root Cause The shared validation implementation awaited validateTypes for metadata and custom data but ignored its parsed return values, as confirmed by the reproduction retaining strings and missing defaults despite successful validation. ## Summary Assigned parsed metadata and custom-data outputs back to the cloned validated messages and added a patch changeset for the ai package. ## Testing Added regression tests covering metadata defaults, async transformations, data coercion, safe validation output, runtime number usage, and caller input immutability. ## End-to-end Validation - `pnpm -C examples/ai-functions exec tsx src/reproduction/validate-ui-messages-schema-output.ts` via immutable replay exited successfully, preserved all parsed outputs, and emitted no reproduction signal. ## Related Issues Fixes #21122 Closes #21320 Backport of #21128 Co-authored-by: ai-sdk-factory[bot] <305873210+ai-sdk-factory[bot]@users.noreply.github.com> Co-authored-by: ai-sdk-factory <308175966+ai-sdk-factory@users.noreply.github.com> Co-authored-by: brennanbutler01 <64561607+brennanbutler01@users.noreply.github.com> Co-authored-by: Gregor Martynus <39992+gr2m@users.noreply.github.com>

…apping request finishes (#21340) ## Background Overlapping React useObject submissions allowed an older request to finish while a newer request was active, incorrectly disabling loading state and preventing stop() or clear() from aborting the newer stream. ## Root Cause The stream close handler unconditionally cleared isLoading and abortControllerRef. Regression tests and the original reproduction confirmed that an older stream therefore discarded the newer stream's controller. ## Summary The close handler now clears request state only when its abort controller is still current. Added a patch changeset for @ai-sdk/react. ## Testing Added overlapping-request regression cases covering both stop() and clear(), including loading state, stream cancellation, and final object state. The release-branch streaming-abort tests now wait for the relevant text or tool-call chunk to reach the output before triggering cancellation. This removes the stream scheduling assumptions that caused five snapshot failures on Node 24.21.0. Local validation: - All 374 `stream-text` tests passed on Node 20, 22, and 24.21.0, and in the edge runtime on Node 24.21.0. - All 18 React `useObject` tests passed. - `pnpm type-check:full` and `pnpm check` passed. ## End-to-end Validation - `pnpm -C examples/ai-functions exec tsx src/reproduction/use-object-overlapping-stop.ts` via immutable replay exited successfully; both newer requests were aborted and retained the expected object state. ## Related Issues Fixes #21124 Closes #21337 Backport of #21135 --------- Co-authored-by: ai-sdk-factory[bot] <305873210+ai-sdk-factory[bot]@users.noreply.github.com> Co-authored-by: ai-sdk-factory <308175966+ai-sdk-factory@users.noreply.github.com> Co-authored-by: brennanbutler01 <64561607+brennanbutler01@users.noreply.github.com> Co-authored-by: Nick Oates <58091943+n1ckoates@users.noreply.github.com> Co-authored-by: Gregor Martynus <39992+gr2m@users.noreply.github.com>

…pts (#21377) ## Background OpenAI gpt-audio Chat Completions returned content as null and the spoken words in message.audio.transcript, but @ai-sdk/openai produced empty generated text. ## Root Cause The chat response schema omitted message.audio, so Zod stripped the valid audio object before doGenerate processed it; the recorded response consequently produced no text while retaining other message fields. ## Summary Added audio transcript parsing to the OpenAI chat response schema and used the transcript as generated text when regular message content is empty, while preserving coexisting tool calls. ## Testing Added regression coverage for an audio transcript returned with null content and a coexisting tool call, and added a patch changeset for @ai-sdk/openai. ## End-to-end Validation - `replay_original_reproduction`: the immutable recorded response completed successfully with the transcript exposed as text; status was `no-longer-reproduces`. - `pnpm -C examples/ai-functions exec tsx -e "<live gpt-audio-1.5 validation>"`: OpenAI returned `content: null` and transcript `audio transcript validation`; generated text matched the transcript. ## Related Issues Fixes #21289 Closes #21356 Backport of #21338 Co-authored-by: ai-sdk-factory[bot] <305873210+ai-sdk-factory[bot]@users.noreply.github.com> Co-authored-by: ai-sdk-factory <308175966+ai-sdk-factory@users.noreply.github.com> Co-authored-by: remorses <31321188+remorses@users.noreply.github.com> Co-authored-by: Gregor Martynus <39992+gr2m@users.noreply.github.com>

…s v6 (#21285) Automated update of the gateway model settings files. Regenerated from the upstream gateway model list by [`update-model-settings.yml`](https://github.com/vercel/ai/blob/main/.github/workflows/update-model-settings.yml) ([this run](https://github.com/vercel/ai/actions/runs/35753288015)). Merging this PR is what surfaces newly-released gateway models as IDE autocomplete suggestions when users type a model string with `@ai-sdk/gateway`. See the workflow file for details on cadence and scope. Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: Zachary <zachsnkrs@gmail.com>

#21342

…s v6 (#21388) Automated update of the gateway model settings files. Regenerated from the upstream gateway model list by [`update-model-settings.yml`](https://github.com/vercel/ai/blob/main/.github/workflows/update-model-settings.yml) ([this run](https://github.com/vercel/ai/actions/runs/35888046520)). Merging this PR is what surfaces newly-released gateway models as IDE autocomplete suggestions when users type a model string with `@ai-sdk/gateway`. See the workflow file for details on cadence and scope. Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>

…from earlier messages (#21396) ## Background Tool approvals can remain open while later conversation turns are added, so decisions and resumed output must update the historical message owning the approval without truncating or reordering later messages. ## Root Cause AbstractChat updated approvals and initialized resumed stream state from only the final message, making historical invocations unavailable to result processing. Failed partial streams also restored an obsolete message ID. The immutable reproduction and historical, competing-approval, hydration, and retry regressions confirmed these defects. ## Summary Approval responses now update the owning historical assistant message while preserving metadata and later turns. Manual, automatic, hydrated, and retried continuations use the correct historical context, and failed retries retain the target's current message ID. The ai patch changeset remains, and the unrelated issue #20951 reproduction matches the target base. ## Testing Regression coverage verifies historical approval and rejection, resumed outputs, later-message and metadata preservation, competing approvals, hydrated and automatic continuation, and retry after streamed ID replacement. All 47 chat tests passed in Node and Edge; full type checking, package build, formatting, linting, and both reproductions passed. ## End-to-end Validation - `pnpm -C examples/ai-functions exec tsx src/reproduction/issue-21193-earlier-tool-approval.ts` via `replay_original_reproduction`: exited 0 and reported that earlier approvals and resumed results succeeded. - `pnpm -C examples/ai-functions exec tsx src/reproduction/issue-20951-generated-file-url.ts`: exited 0 and reported that the generated-file URL issue did not reproduce. ## Related Issues Fixes #21193 Closes #21391 Backport of #21203 --------- Co-authored-by: ai-sdk-factory[bot] <305873210+ai-sdk-factory[bot]@users.noreply.github.com> Co-authored-by: ai-sdk-factory <308175966+ai-sdk-factory@users.noreply.github.com>

…tic batches (#21397) ## Background Google embedMany supports one multimodal content entry per value, but calls exceeding the 100-value provider limit failed instead of automatically batching aligned values and content. ## Root Cause Core batched the values but reused the full providerOptions object for every batch, so Google validated the complete content array against each smaller value chunk. The original 101-value reproduction failed before any request and now completes successfully. ## Summary Added an experimental provider-options transformation capability in provider-utils, applied original batch index ranges in embedMany, preserved the capability through embedding-model wrappers, and made Google validate and slice per-value content for each batch. Added patch changesets for ai, Google, and provider-utils without dependency or configuration changes. ## Testing Coverage verifies index alignment through sequential, parallel, byte-limited, wrapped, retried, null-content, and middleware-normalized batches, plus mismatch validation. All 89 focused tests passed in both Node and Edge; formatting, linting, and full type checking also passed. ## End-to-end Validation - `pnpm -C examples/ai-functions exec tsx src/reproduction/google-embed-many-content-batching.ts` — exact original synthetic reproduction exited 0 and retained all 101 value/content pairings without the bug signal. ## Related Issues Fixes #21095 Closes #21392 Backport of #21143 --------- Co-authored-by: ai-sdk-factory[bot] <305873210+ai-sdk-factory[bot]@users.noreply.github.com> Co-authored-by: ai-sdk-factory <308175966+ai-sdk-factory@users.noreply.github.com> Co-authored-by: Gregor Martynus <39992+gr2m@users.noreply.github.com>

## Background The v6 browser bundle reached 620.47 KB after the historical tool-approval fix (#21396), then 620.72 KB after #21397. Both exceeded the 620 KB budget and failed CI: - https://github.com/vercel/ai/actions/runs/35898714869 - https://github.com/vercel/ai/actions/runs/35903439574 ## Summary Raise the v6 bundle budget from 620 to 625 KB, following this release branch's existing 5 KB increments (#19298). Current bundles measure 618.12 KB for Node and 620.72 KB for browsers. This accommodates the recent fixes while retaining a bounded size check. Correct the failure message to point to the `LIMIT` constant in this script. The previously suggested GitHub Actions variable is not read by the check. Only the CI measurement script changes; published runtime code is unaffected, so no changeset is needed. ## Validation - Built the `ai` workspace dependencies with `pnpm exec turbo build --filter='ai^...'`. - `pnpm -C packages/ai check-bundle-size` passes for Node and browser bundles. - `pnpm type-check:full` passes. - `pnpm check` passes. - `git diff --check` passes. ## Checklist - [x] 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) - [ ] A _patch_ changeset for relevant packages has been added (for bug fixes / features - run `pnpm changeset` in the project root) - [x] I have reviewed this pull request (self-review)

… across model and UI continuations (#21408) ## Problem Manual tool approvals rejected valid inputs after schema transforms. For example, a schema that converts `{ count: "3" }` to `{ count: 3 }` produced an approval that failed when the SDK validated the transformed number against the original string schema. ## Changes - Preserve original schema input when it differs from the approved input, including persisted ModelMessage and UIMessage continuations. - Reconstruct schema output and input refinement before execution, require it to match the approved input, and retain the approved call unchanged. Missing metadata cannot cause a signed or unsigned input to be transformed into a different operation. - Check that transformed UI tool inputs match their reconstructed schema output, including configured input refinement. Preserve failed history as dynamic tool parts when its input cannot be validated. - Share the repeated UI approval schemas to reduce duplication and keep bundle sizes within the existing limits. - Document persistence requirements and the fact that original input, including fields removed by transforms, is included in approval metadata. ## Validation - Focused Node tests: 970 passed, 3 skipped. - Focused Edge tests: 970 passed, 3 skipped. - Regression coverage includes signed non-idempotent transforms with preserved or omitted metadata, unchanged approved values, mismatched UI values and types across approval states, and persisted UI continuation. - `pnpm type-check:full` passed. - `pnpm check` passed. - Bundle size checks passed without raising limits. Fixes #21091 Closes #21093 ## Related Issues Closes #21407 Backport of #21130 --------- Co-authored-by: ai-sdk-factory[bot] <305873210+ai-sdk-factory[bot]@users.noreply.github.com> Co-authored-by: ai-sdk-factory <308175966+ai-sdk-factory@users.noreply.github.com>

…has provider option (#21409)

…m errors (#21348) ## Background OpenAI insufficient-quota errors received before streamed output were retried three times, delaying failure and surfacing RetryError instead of one non-retryable APICallError. ## Root Cause createOpenAIStreamError assigned status 429 but omitted the existing stream-error retry classification, causing APICallError to infer that quota errors were retryable. The recorded insufficient_quota fixture confirmed isRetryable was incorrectly true. ## Summary Propagated stream-error retryability into APICallError, updated the insufficient-quota regression assertion, added an OpenAI patch changeset, and removed reproduction-only artifacts. ## Testing Updated the existing Responses API stream-error fixture test to require isRetryable: false for insufficient_quota errors. ## End-to-end Validation - `replay_original_reproduction`: passed with one fetch and a non-retryable APICallError in 77 ms; classified as no-longer-reproduces. ## Related Issues Fixes #21345 Closes #21347 --------- Co-authored-by: ai-sdk-factory <308175966+ai-sdk-factory@users.noreply.github.com> Co-authored-by: dor-sr <7157779+dor-sr@users.noreply.github.com>

Navigation Menu

Repository navigation

Comparing changes

Open a pull request

Commits on Sep 23, 2026

This comparison is taking too long to generate.

Uh oh!