…1628)

## Background

Pruning a converted multi-step Anthropic conversation can remove a code
execution source while retaining a dependent tool call with
`caller.toolId`. Anthropic rejects the next user turn with `source tool
... not found for tool use block ...`.

Backports #21623 to `release-v6.0` (`@ai-sdk/anthropic` 3.x). The v6
provider has the same affected caller serialization. The v5 provider
does not serialize this metadata, so this fix does not apply to v5.

## Summary

- Normalize dangling programmatic caller references entirely inside the
Anthropic converter. Keep retained calls, results, ordering, and cache
controls, and emit one warning per affected call.
- Only normalize history before a subsequent user message. Preserve
direct callers, valid references, and active tool continuations; pruned
source history is not reconstructed.
- Adapt the source PR to `convertToAnthropicMessagesPrompt` and V3
provider types. Include the same regressions, runnable reproduction,
provider documentation, and patch changeset.
- Preserve Daniel Lantz (@dflantz) as a commit co-author.

## Manual Verification

Ran `convertToModelMessages` → `pruneMessages` → `generateText` against
the live Anthropic API with `claude-opus-4-5`. Both
`code_execution_20250825` and `code_execution_20260120` succeeded,
returned the retained AAPL price ($185.42), and emitted the expected
omission warning. The reproduction asserts that pruning still removes
the source and retains the dependent tool call.

```sh
pnpm -C examples/ai-functions exec tsx src/generate-text/anthropic/pruned-programmatic-tool-calling.ts
pnpm -C examples/ai-functions exec tsx src/generate-text/anthropic/pruned-programmatic-tool-calling.ts code_execution_20260120
```

## Validation

- Confirmed the new normalization regressions fail on the unmodified v6
converter.
- Anthropic Node and Edge suites: 479 tests passed in each environment,
including generation and streaming request/warning checks.
- `pnpm type-check:full`
- `pnpm check`
- `git diff --check`

## Checklist

- [x] Tests have been added / updated (for bug fixes / features)
- [x] Documentation has been added / updated (for bug fixes / features)
- [x] 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)

## Related Issues

Backport of #21623 (source commit
f8f6985).
Addresses #12504 on AI SDK 6.

Co-authored-by: Daniel Lantz <6571678+dflantz@users.noreply.github.com>

…ter a disconnect (#21634)

## Background

The published ai package failed to resume an interrupted UI message
stream when reconnection began with a text-delta for a text part opened
before the disconnect.

## Root Cause

After a network disconnect, AbstractChat discarded the streaming state
containing active part ID mappings and created fresh state for
resumeStream. The reproduction confirmed that the resumed text-delta
could not find text-1 despite its existing partial message.

## Summary

AbstractChat now retains streaming state across network disconnects and
reuses it for reconnection, while clearing it after completion,
cancellation, non-network errors, or new requests. Added a patch
changeset.

## Testing

Added a regression test covering disconnection after text-start and
successful resumption beginning with a same-ID text-delta.

## End-to-end Validation

- `pnpm -C packages/ai build` followed by immutable replay `pnpm -C
examples/ai-functions exec tsx
src/reproduction/issue-13160-resume-text-delta.ts`: build succeeded and
the reproduction exited successfully without the original failure
signal.

## Related Issues

Fixes #13160

Closes #21503

Backport of #21509

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: KaranJayakumar <117019183+KaranJayakumar@users.noreply.github.com>
Co-authored-by: Gregor Martynus <39992+gr2m@users.noreply.github.com>

…21632)

## Background

Backport of #21631 to `release-v6.0`.

Claude Sonnet 5.5 (`claude-sonnet-5-5`) rejects `thinking: { type:
'disabled' }`, budget-based thinking, forced `tool_choice` (`required`
or a named tool, including the `jsonTool` structured output mode), and
the older `computer_*` tool types. It adds a new lowest thinking
setting, `thinking: { type: 'between_tools' }`, which is only accepted
at `low`, `medium`, and `high` effort.

## Summary

- Add the `claude-sonnet-5-5` model ID to `@ai-sdk/anthropic` and
`@ai-sdk/google-vertex`, `anthropic.claude-sonnet-5-5` /
`us.anthropic.claude-sonnet-5-5` to `@ai-sdk/amazon-bedrock`, and
`anthropic/claude-sonnet-5.5` to `@ai-sdk/gateway`.
- Add a `supportsBetweenToolsThinking` model capability and the
`thinking: { type: 'between_tools' }` provider option. Combining it with
`xhigh` or `max` effort lowers the effort to `high` with a warning.
- For `claude-sonnet-5-5`: `thinking: { type: 'disabled' }` is replaced
with `between_tools` (with a warning) and budget-based thinking becomes
adaptive. The existing `rejectsForcedToolUse` handling downgrades
`required` / named tool choices to `auto` and `jsonTool` to
`outputFormat`, each with a warning.
- Docs: model tables, effort guidance, a "Between-Tools Thinking"
section, and Sonnet 5.5 notes for thinking updates, block binding, and
the computer toolset.
- New examples:
`examples/ai-functions/src/generate-text/anthropic/claude-sonnet-5-5.ts`,
`generate-text/anthropic/claude-sonnet-5-5-between-tools.ts`,
`stream-text/anthropic/claude-sonnet-5-5-between-tools.ts`,
`stream-text/anthropic/claude-sonnet-5-5-thinking-updates.ts`.

Unlike `main`, this branch has no top-level `reasoning` option, so there
is no `reasoning: 'none'` mapping. Docs and examples that use
`claude-sonnet-5` generically are not bulk-updated, matching the Opus
5.5 backports.

## End-to-End Verification

The same changes were verified on `main` (#21631) against the Anthropic
API with `claude-sonnet-5-5`. On this branch, the Anthropic, Amazon
Bedrock, Gateway, and Google Vertex unit tests, `pnpm type-check:full`,
and `pnpm check` pass.

Co-authored-by: Claude <noreply@anthropic.com>

This PR was opened by the [Changesets
release](https://github.com/changesets/action) GitHub action. When
you're ready to do a release, you can merge this and the packages will
be published to npm automatically. If you're not ready to do a release
yet, that's fine, whenever you add more changesets to release-v6.0, this
PR will be updated.


# Releases
## ai@6.0.295

### Patch Changes

- 0b38b6b: fix(ai): continue active UI message parts when resuming after
a disconnect
- Updated dependencies [358683e]
  - @ai-sdk/gateway@3.0.205
## @ai-sdk/amazon-bedrock@4.0.186

### Patch Changes

- 358683e: feat(anthropic): add Claude Sonnet 5.5 support

- add the `claude-sonnet-5-5` model ID to `@ai-sdk/anthropic` and
`@ai-sdk/google-vertex`, `anthropic.claude-sonnet-5-5` and
`us.anthropic.claude-sonnet-5-5` to `@ai-sdk/amazon-bedrock`, and
`anthropic/claude-sonnet-5.5` to `@ai-sdk/gateway`
- add the `between_tools` thinking type (`thinking: { type:
'between_tools' }`), the lowest thinking setting on `claude-sonnet-5-5`;
`xhigh` and `max` effort are lowered to `high` with a warning because
the API rejects them with `between_tools`
- `claude-sonnet-5-5` rejects disabled thinking: `thinking: { type:
'disabled' }` is replaced with `between_tools` thinking (with a
warning), and budget-based thinking is converted to adaptive thinking
- `claude-sonnet-5-5` rejects forced tool use: `required` and named tool
choices fall back to `auto`, and `structuredOutputMode: 'jsonTool'`
falls back to native structured outputs, each with a warning

- Updated dependencies [358683e]
- Updated dependencies [eeabdce]
  - @ai-sdk/anthropic@3.0.125
## @ai-sdk/angular@2.0.296

### Patch Changes

- Updated dependencies [0b38b6b]
  - ai@6.0.295
## @ai-sdk/anthropic@3.0.125

### Patch Changes

- 358683e: feat(anthropic): add Claude Sonnet 5.5 support

- add the `claude-sonnet-5-5` model ID to `@ai-sdk/anthropic` and
`@ai-sdk/google-vertex`, `anthropic.claude-sonnet-5-5` and
`us.anthropic.claude-sonnet-5-5` to `@ai-sdk/amazon-bedrock`, and
`anthropic/claude-sonnet-5.5` to `@ai-sdk/gateway`
- add the `between_tools` thinking type (`thinking: { type:
'between_tools' }`), the lowest thinking setting on `claude-sonnet-5-5`;
`xhigh` and `max` effort are lowered to `high` with a warning because
the API rejects them with `between_tools`
- `claude-sonnet-5-5` rejects disabled thinking: `thinking: { type:
'disabled' }` is replaced with `between_tools` thinking (with a
warning), and budget-based thinking is converted to adaptive thinking
- `claude-sonnet-5-5` rejects forced tool use: `required` and named tool
choices fall back to `auto`, and `structuredOutputMode: 'jsonTool'`
falls back to native structured outputs, each with a warning

- eeabdce: Normalize dangling programmatic tool caller references in
conversation history after pruning. Before a subsequent user message,
omit caller metadata whose source code execution call is missing and
emit a warning, preserving the retained tool calls and results. Keep
caller metadata unchanged for active tool continuations.
## @ai-sdk/anthropic-aws@1.0.47

### Patch Changes

- Updated dependencies [358683e]
- Updated dependencies [eeabdce]
  - @ai-sdk/anthropic@3.0.125
## @ai-sdk/gateway@3.0.205

### Patch Changes

- 358683e: feat(anthropic): add Claude Sonnet 5.5 support

- add the `claude-sonnet-5-5` model ID to `@ai-sdk/anthropic` and
`@ai-sdk/google-vertex`, `anthropic.claude-sonnet-5-5` and
`us.anthropic.claude-sonnet-5-5` to `@ai-sdk/amazon-bedrock`, and
`anthropic/claude-sonnet-5.5` to `@ai-sdk/gateway`
- add the `between_tools` thinking type (`thinking: { type:
'between_tools' }`), the lowest thinking setting on `claude-sonnet-5-5`;
`xhigh` and `max` effort are lowered to `high` with a warning because
the API rejects them with `between_tools`
- `claude-sonnet-5-5` rejects disabled thinking: `thinking: { type:
'disabled' }` is replaced with `between_tools` thinking (with a
warning), and budget-based thinking is converted to adaptive thinking
- `claude-sonnet-5-5` rejects forced tool use: `required` and named tool
choices fall back to `auto`, and `structuredOutputMode: 'jsonTool'`
falls back to native structured outputs, each with a warning
## @ai-sdk/google-vertex@4.0.208

### Patch Changes

- 358683e: feat(anthropic): add Claude Sonnet 5.5 support

- add the `claude-sonnet-5-5` model ID to `@ai-sdk/anthropic` and
`@ai-sdk/google-vertex`, `anthropic.claude-sonnet-5-5` and
`us.anthropic.claude-sonnet-5-5` to `@ai-sdk/amazon-bedrock`, and
`anthropic/claude-sonnet-5.5` to `@ai-sdk/gateway`
- add the `between_tools` thinking type (`thinking: { type:
'between_tools' }`), the lowest thinking setting on `claude-sonnet-5-5`;
`xhigh` and `max` effort are lowered to `high` with a warning because
the API rejects them with `between_tools`
- `claude-sonnet-5-5` rejects disabled thinking: `thinking: { type:
'disabled' }` is replaced with `between_tools` thinking (with a
warning), and budget-based thinking is converted to adaptive thinking
- `claude-sonnet-5-5` rejects forced tool use: `required` and named tool
choices fall back to `auto`, and `structuredOutputMode: 'jsonTool'`
falls back to native structured outputs, each with a warning

- Updated dependencies [358683e]
- Updated dependencies [eeabdce]
  - @ai-sdk/anthropic@3.0.125
## @ai-sdk/langchain@2.0.303

### Patch Changes

- Updated dependencies [0b38b6b]
  - ai@6.0.295
## @ai-sdk/llamaindex@2.0.295

### Patch Changes

- Updated dependencies [0b38b6b]
  - ai@6.0.295
## @ai-sdk/minimax@2.0.23

### Patch Changes

- Updated dependencies [358683e]
- Updated dependencies [eeabdce]
  - @ai-sdk/anthropic@3.0.125
## @ai-sdk/react@3.0.298

### Patch Changes

- Updated dependencies [0b38b6b]
  - ai@6.0.295
## @ai-sdk/rsc@2.0.295

### Patch Changes

- Updated dependencies [0b38b6b]
  - ai@6.0.295
## @ai-sdk/svelte@4.0.295

### Patch Changes

- Updated dependencies [0b38b6b]
  - ai@6.0.295
## @ai-sdk/vue@3.0.295

### Patch Changes

- Updated dependencies [0b38b6b]
  - ai@6.0.295

Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>

… originalMessages (#21636)

## Background

createAgentUIStream rejected persistence flows that supplied
originalMessages without uiMessages, preventing agent execution despite
supporting originalMessages for stream persistence.

## Root Cause

createAgentUIStream always validated and converted uiMessages, whose
type was required, while originalMessages was only forwarded to the
output stream. The reproduction confirmed undefined uiMessages reached
validation and caused the missing-messages error.

## Summary

Made uiMessages optional across the agent UI stream helpers, use
originalMessages as the agent input fallback, updated documentation, and
added an ai package patch changeset.

## Testing

Added regression coverage confirming originalMessages supplies the model
prompt, generates the assistant message ID, and produces the complete
persisted history.

## End-to-end Validation

- `pnpm -C examples/ai-functions exec tsx
src/reproduction/issue-10367-create-agent-ui-stream-original-messages.ts`
via immutable replay completed successfully with exit code 0 and no
original bug signal.

## Related Issues

Fixes #10367

Closes #21629

---------

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>
Co-authored-by: christ12938 <25758598+christ12938@users.noreply.github.com>

…age streams (#21637)

## Background

Tool metadata produced after a tool completed was discarded, preventing
final status, cache, timing, or download metadata from reaching UI
message parts.

## Root Cause

The output-available and output-error handlers copied toolMetadata only
from the existing tool invocation instead of the output chunk. The
reproduction confirmed this across static and dynamic tool paths.

## Summary

Updated all static and dynamic tool output handlers to prefer
output-chunk metadata while retaining existing metadata when the output
chunk omits it. Added a patch changeset.

## Testing

Added regression coverage for available and error outputs on both static
and dynamic tools, including replacement of stale input metadata.

## End-to-end Validation

- `replay_original_reproduction`: the immutable reproduction exited
successfully and reported that all output chunks propagated tool
metadata.

## Related Issues

Fixes #16930

Closes #21627

Backport of #21615

---------

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>
Co-authored-by: matanper <9868530+matanper@users.noreply.github.com>

…ross-delta leakage (#21646)

## Background

smoothStream dropped provider metadata from loop-emitted text and
reasoning chunks, while incomplete metadata state handling could lose
signature-only deltas or associate metadata with unrelated later
content.

## Root Cause

The chunk-detection loop omitted buffered provider metadata, and the
buffer state was not bounded by metadata ownership. Immutable replay
confirmed the original metadata loss, while exact-boundary and
multi-delta probes confirmed signature loss, stale inheritance, and
overwrite paths.

## Summary

Propagated provider metadata to loop-emitted smoothStream deltas,
preserved empty metadata-bearing deltas, flushed buffers at metadata
boundaries, cleared consumed metadata, added a patch changeset, and
added regression coverage for preservation and isolation.

## Testing

Regression coverage now includes word-chunked reasoning, line-chunked
text, empty metadata after exact boundaries, metadata-bearing to
metadata-free same-ID transitions, distinct consecutive metadata, and
type/ID transitions. Node, Edge, full workspace tests, formatting, lint,
type checking, and immutable reproduction replay pass.

## End-to-end Validation

- Immutable replay of `pnpm -C examples/ai-functions exec tsx
src/reproduction/smooth-stream-provider-metadata.ts` exited 0 and showed
metadata on every corresponding word- and line-chunked delta.

## Related Issues

Fixes #14373

Closes #19528

Backport of #19561

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: kuishou68 <54054995+kuishou68@users.noreply.github.com>
Co-authored-by: Gregor Martynus <39992+gr2m@users.noreply.github.com>

This PR was opened by the [Changesets
release](https://github.com/changesets/action) GitHub action. When
you're ready to do a release, you can merge this and the packages will
be published to npm automatically. If you're not ready to do a release
yet, that's fine, whenever you add more changesets to release-v6.0, this
PR will be updated.


# Releases
## ai@6.0.296

### Patch Changes

- 0741da8: fix(ai): allow agent UI streams to use original messages as
input
- a63fa9b: fix(ai): preserve tool metadata from tool output chunks
- a61bea9: Preserve provider metadata on corresponding `smoothStream`
chunks without carrying it into subsequent metadata-free deltas.
## @ai-sdk/angular@2.0.297

### Patch Changes

- Updated dependencies [0741da8]
- Updated dependencies [a63fa9b]
- Updated dependencies [a61bea9]
  - ai@6.0.296
## @ai-sdk/langchain@2.0.304

### Patch Changes

- Updated dependencies [0741da8]
- Updated dependencies [a63fa9b]
- Updated dependencies [a61bea9]
  - ai@6.0.296
## @ai-sdk/llamaindex@2.0.296

### Patch Changes

- Updated dependencies [0741da8]
- Updated dependencies [a63fa9b]
- Updated dependencies [a61bea9]
  - ai@6.0.296
## @ai-sdk/react@3.0.299

### Patch Changes

- Updated dependencies [0741da8]
- Updated dependencies [a63fa9b]
- Updated dependencies [a61bea9]
  - ai@6.0.296
## @ai-sdk/rsc@2.0.296

### Patch Changes

- Updated dependencies [0741da8]
- Updated dependencies [a63fa9b]
- Updated dependencies [a61bea9]
  - ai@6.0.296
## @ai-sdk/svelte@4.0.296

### Patch Changes

- Updated dependencies [0741da8]
- Updated dependencies [a63fa9b]
- Updated dependencies [a61bea9]
  - ai@6.0.296
## @ai-sdk/vue@3.0.296

### Patch Changes

- Updated dependencies [0741da8]
- Updated dependencies [a63fa9b]
- Updated dependencies [a61bea9]
  - ai@6.0.296

Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>

…llows a completed tool call (#21656)

## Background

A completed backend tool output followed by terminal error text caused
the UI chat helper to report completion and automatically submit another
request.

## Root Cause

The helper evaluated completed tool parts in the final step without
accounting for later text lacking a completed stream state, so the
completed tool produced a false-positive resumption signal. The
immutable reproduction and regression tests confirmed this behavior.

## Summary

Updated the completion helper to reject completed tools followed by
uncompleted terminal text while retaining resumption for text marked
`done`, added helper and chat integration regressions, and included an
`ai` patch changeset. No dependencies or configuration were changed.

## Testing

Helper and chat integration tests cover terminal error text stopping
automatic submission and completed model text preserving client
tool-result submission. Focused Node and Edge tests, the complete
repository test suite, linting, formatting, and full type checking pass.

## End-to-end Validation

- `replay_original_reproduction`: returned `false`, exited successfully,
and classified the issue as `no-longer-reproduces`.

## Related Issues

Fixes #12099

Closes #21644

Backport of #21622

---------

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>
Co-authored-by: alexandrudanpop <15979292+alexandrudanpop@users.noreply.github.com>

## Background

Backport of #21649 to `release-v6.0`.

Reconnect URLs interpolate chat IDs into a path, so an ID such as
`../other-route` changes the route and `?` or `#` changes URL
interpretation. An empty API base can also turn a leading slash in the
ID into a different host.

## Summary

This release did not yet have the query-aware path helper, so include it
and its regression coverage from main.

Encode the ID as a single segment and append the reconnect path before
any base query or fragment. Reject standalone `.` and `..` before
fetching, since percent-encoding does not prevent URL dot-segment
normalization. Preserve the original ID passed to
`prepareReconnectToStreamRequest` and use its custom URL verbatim.
Document the behavior and update the custom endpoint examples.

Validation: 14 regression cases fail before the fix; all 24 transport
tests pass in Node and Edge after it. Full workspace type checking,
targeted lint/format checks, and MDX parsing pass. Includes an `ai`
patch changeset.

## Contributor Credit

Reported by @trmxvibs in #18187. Related proposed fixes: #18395 and
#20001.

## Manual Verification

Ran `DefaultChatTransport` against a local HTTP server. The server
received encoded IDs for traversal-like, query/fragment, and
percent-containing values, with the original API query intact.
Standalone `..` was rejected without a request; a custom prepared URL
was requested unchanged.

## Checklist

- [x] Tests have been added / updated (for bug fixes / features)
- [x] Documentation has been added / updated (for bug fixes / features)
- [x] 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)

## Related Issues

Addresses finding 3 of #18187. The other findings are tracked
separately.

## Background

Backport of #21648 to `release-v6.0`, adapted to the release’s provider
interfaces.

Tools, provider IDs, and provider metadata named `__proto__` can change
lookup object prototypes instead of becoming own properties. Inherited
names such as `constructor` can also resolve as unregistered providers
or crash image metadata aggregation.

## Summary

V6 uses an inline video metadata merge and has no per-image metadata
getter.

Use null-prototype maps for MCP tools and the provider registry. Merge
embedding, image, and video metadata with own-property reads and writes,
while retaining plain objects in returned metadata. Add regressions for
prototype-named keys and patch changesets for `ai` and `@ai-sdk/mcp`.

Validation: the added regressions exposed failures in all five
applicable areas before the fix. The affected AI SDK suites pass 166
tests in each of Node and Edge; the MCP client suite passes 82 in each.
Changed-file lint and formatting checks and `pnpm type-check:full` pass.

## Contributor Credit

Reported by @trmxvibs in #18187. Related proposed fix: #20005.

## Manual Verification

Ran a standalone local reproduction through embedding, image, and video
generation, provider lookup, and MCP tool discovery/execution using
synthetic models and an in-memory MCP transport. `__proto__` survives
enumeration and execution, returned metadata retains its ordinary
prototype, and provider members do not become registered providers. No
live provider credentials are required.

## Checklist

- [x] Tests have been added / updated (for bug fixes / features)
- [ ] Documentation has been added / updated (for bug fixes / features)
- [x] 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)

## Related Issues

Addresses finding 1 of #18187. The other findings are tracked
separately.

…s v6 (#21609)

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/36450474963)).
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>

## Background

Backport of #21650 to `release-v6.0`. The exception exists in v6; v5
still masks these errors through `onError`, so this documentation does
not apply there.

The UI stream documentation describes default error masking without
explaining the intentional exception for provider-executed tool
execution errors. Applying `onError` to these errors would discard
provider error codes needed for subsequent model requests.

## Summary

Use v6’s `result.toUIMessageStream()` and
`result.toUIMessageStreamResponse()` APIs and document the inline
converter in `stream-text.ts`. Omit main-only harness references.

Document the exception in the chatbot and tool guides, API reference,
shipped type comments, and conversion code. Explain that strings pass
through unchanged, other values are JSON-stringified, and invalid tool
calls and stream errors still use `onError`. Explicitly state that the
callback cannot redact provider-executed tool execution errors before
they reach the client.

Runtime behavior is unchanged: both edited TypeScript files emit
identical JavaScript with comments removed. Existing streamText and
model-message conversion tests pass (437 in each of Node and Edge), as
do MDX parsing, targeted lint/format checks, and the full workspace type
check. The patch changeset covers the shipped API documentation.

## Contributor Credit

Reported by @trmxvibs in #18187. The behavior was introduced by #13256
to preserve Anthropic tool error codes.

## Checklist

- [ ] Tests have been added / updated (for bug fixes / features)
- [x] Documentation has been added / updated (for bug fixes / features)
- [x] 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)

## Related Issues

Addresses finding 4 of #18187 by documenting the intentional exception.
The other findings are tracked separately.

…streaming state (#21665)

## Background

Hydrated assistant messages containing unfinished static tool input
could not accept resumed deltas, and successfully resumed calls could
lose their state after another interruption.

## Root Cause

The stream state builder did not reconstruct unfinished hydrated tool
calls, Chat resume discarded the relevant assistant message or resumable
state, and streaming updates did not consistently retain accumulated raw
input. The immutable reproduction confirmed both original resume paths,
while repeated-interruption and dynamic-error tests confirmed the
broader state transitions.

## Summary

Reconstructs unfinished static and dynamic tool-call state, preserves
hydrated and in-memory resumable assistant state across interruptions,
retains rawInput only while input is streaming, updates public typing
and validation, documents persistence requirements, updates framework
expectations, and includes a patch changeset.

## Testing

Regression coverage includes readUIMessageStream, repeated
Chat.resumeStream interruptions, legacy messages without raw input,
stream processing, public types, validation, framework rendering, and
dynamic terminal errors. Focused Node and Edge suites each passed 161
tests; package and full-workspace type checks, formatting, and linting
passed.

## End-to-end Validation

- `replay_original_reproduction` ran `pnpm -C examples/ai-functions exec
tsx src/reproduction/issue-14027-partial-static-tool-resume.ts` and
reported `no-longer-reproduces`.

## Related Issues

Fixes #14027

Closes #21474

Backport of #21480

---------

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>

Backport of #21747.

Original PR: #21747

…in open behind reverse proxies (#21791)

## Background

Idle UI message streams emitted no initial response byte or inactivity
traffic, allowing reverse proxies to time out healthy streams and
eventually exhaust transport reconnect attempts.

## Root Cause

The SSE pipeline emitted only when its source produced or completed, so
idle sources could not flush headers or generate heartbeat traffic. The
initial fix's repeated Promise.race calls also accumulated reactions on
the unresolved source read; the immutable replay and sustained-idle
regression confirm both defects are resolved.

## Summary

Added opt-in keepAliveMs support that emits an immediate SSE opening
comment and periodic inactivity comments across Response,
ServerResponse, and agent helpers. The scheduler now retains one
source-read reaction and one timer, propagates cancellation, respects
backpressure, and is documented with a patch changeset.

## Testing

Coverage includes opening comments, idle heartbeats, activity-based
timer resets, 2,500 consecutive heartbeats with one retained read
reaction, source completion and cancellation, and Response and
ServerResponse integrations. The complete ai package tests, full
workspace type checking, checks, and documentation validation pass.

## End-to-end Validation

- `pnpm -C examples/ai-functions exec tsx
src/reproduction/issue-17805.ts` — exited 0 without the immutable
reproduction signal; first-byte flushing and idle keep-alive behavior
prevented reconnect exhaustion.

## Related Issues

Fixes #17805

Closes #21654

Backport of #21672

---------

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 625.71 KiB and exceeded the 625 KiB budget
in
https://github.com/vercel/ai/actions/runs/36625796543/job/109604503194.

## Summary

Raise the v6 bundle budget from 625 to 630 KiB, following the release
branch established 5 KiB increments. Current bundles measure 623.14 KiB
for Node and 625.71 KiB for browsers.

Only the CI measurement script changes; published runtime code is
unaffected, so no changeset is needed.

## Validation

- `pnpm exec turbo build --filter=ai^...`
- `pnpm -C packages/ai check-bundle-size`
- `pnpm type-check:full`
- `pnpm check`
- `git diff --check`

This is an automated backport of #21811 to the release-v6.0 branch. FYI
@jeremyphilemon

Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>

This PR was opened by the [Changesets
release](https://github.com/changesets/action) GitHub action. When
you're ready to do a release, you can merge this and the packages will
be published to npm automatically. If you're not ready to do a release
yet, that's fine, whenever you add more changesets to release-v6.0, this
PR will be updated.


# Releases
## ai@6.0.297

### Patch Changes

- 8349396: fix(ai): preserve hydrated partial static tool input across
stream resumptions
- ab0b94b: fix(ai): keep idle UI message streams open with optional SSE
heartbeats
- ce660b0: Clarify that provider-executed tool execution errors bypass
the UI stream's `onError` callback to preserve provider error data.
Stream errors and invalid tool calls still use the callback. Runtime
behavior is unchanged.
- 850d4d5: Encode chat IDs in default stream reconnection URLs so
slashes, query delimiters, and fragments stay within the ID. Reject
standalone `.` and `..` IDs before fetching. Custom URLs returned by
`prepareReconnectToStreamRequest` remain unchanged.
- cdc5b56: Preserve prototype-named tools, providers, and provider
metadata as own properties without changing lookup object prototypes.
Prevent inherited names from resolving as registered providers or
causing image metadata aggregation to fail.
- d4203d2: Prevent automatic chat resumption when completed tool output
is followed by terminal text without a completed stream state, while
preserving resumption after completed model text.
- Updated dependencies [4972874]
- Updated dependencies [03e5a25]
  - @ai-sdk/gateway@3.0.206
## @ai-sdk/amazon-bedrock@4.0.187

### Patch Changes

- Updated dependencies [4972874]
  - @ai-sdk/openai@3.0.121
## @ai-sdk/angular@2.0.298

### Patch Changes

- Updated dependencies [8349396]
- Updated dependencies [ab0b94b]
- Updated dependencies [ce660b0]
- Updated dependencies [850d4d5]
- Updated dependencies [cdc5b56]
- Updated dependencies [d4203d2]
  - ai@6.0.297
## @ai-sdk/azure@3.0.129

### Patch Changes

- Updated dependencies [4972874]
- Updated dependencies [f6a20b4]
  - @ai-sdk/openai@3.0.121
  - @ai-sdk/deepseek@2.0.70
## @ai-sdk/deepseek@2.0.70

### Patch Changes

- f6a20b4: fix cached input token reporting for OpenAI-compatible
DeepSeek responses
## @ai-sdk/gateway@3.0.206

### Patch Changes

- 4972874: feat(openai): add GPT-6.1 Sol model support
- 03e5a25: Backport: chore(provider/gateway): update gateway model
settings files
## @ai-sdk/langchain@2.0.305

### Patch Changes

- Updated dependencies [8349396]
- Updated dependencies [ab0b94b]
- Updated dependencies [ce660b0]
- Updated dependencies [850d4d5]
- Updated dependencies [cdc5b56]
- Updated dependencies [d4203d2]
  - ai@6.0.297
## @ai-sdk/llamaindex@2.0.297

### Patch Changes

- Updated dependencies [8349396]
- Updated dependencies [ab0b94b]
- Updated dependencies [ce660b0]
- Updated dependencies [850d4d5]
- Updated dependencies [cdc5b56]
- Updated dependencies [d4203d2]
  - ai@6.0.297
## @ai-sdk/mcp@1.0.89

### Patch Changes

- cdc5b56: Preserve prototype-named tools, providers, and provider
metadata as own properties without changing lookup object prototypes.
Prevent inherited names from resolving as registered providers or
causing image metadata aggregation to fail.
## @ai-sdk/openai@3.0.121

### Patch Changes

- 4972874: feat(openai): add GPT-6.1 Sol model support
## @ai-sdk/react@3.0.300

### Patch Changes

- Updated dependencies [8349396]
- Updated dependencies [ab0b94b]
- Updated dependencies [ce660b0]
- Updated dependencies [850d4d5]
- Updated dependencies [cdc5b56]
- Updated dependencies [d4203d2]
  - ai@6.0.297
## @ai-sdk/rsc@2.0.297

### Patch Changes

- Updated dependencies [8349396]
- Updated dependencies [ab0b94b]
- Updated dependencies [ce660b0]
- Updated dependencies [850d4d5]
- Updated dependencies [cdc5b56]
- Updated dependencies [d4203d2]
  - ai@6.0.297
## @ai-sdk/svelte@4.0.297

### Patch Changes

- Updated dependencies [8349396]
- Updated dependencies [ab0b94b]
- Updated dependencies [ce660b0]
- Updated dependencies [850d4d5]
- Updated dependencies [cdc5b56]
- Updated dependencies [d4203d2]
  - ai@6.0.297
## @ai-sdk/vue@3.0.297

### Patch Changes

- Updated dependencies [8349396]
- Updated dependencies [ab0b94b]
- Updated dependencies [ce660b0]
- Updated dependencies [850d4d5]
- Updated dependencies [cdc5b56]
- Updated dependencies [d4203d2]
  - ai@6.0.297

Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>