fix(core): strip tool output provider metadata from direct UI stream

[liuzemei]

Summary

• strip providerMetadata from direct result.toUIMessageStream() tool output chunks for tool-output-available and tool-output-error
• leave guardrail/full-stream UI conversion unchanged so provider metadata forwarding for [BUG] Tool calls do not work with google-vertex thinking models due to a zod schema mismatch #1195 remains intact
• add regression coverage for the direct UI stream path

Closes #1290

Tests

• pnpm install --frozen-lockfile --ignore-scripts
• pnpm --dir packages/internal build
• pnpm --dir packages/core test:single src/agent/agent.spec.ts src/agent/streaming/guardrail-stream.spec.ts
• pnpm --dir packages/core typecheck
• pnpm --dir packages/core lint

---

Summary by cubic

Stop leaking provider metadata to console clients by stripping it from direct UI stream tool output chunks in @voltagent/core. Guardrail and full-stream paths are unchanged for #1195. Closes #1290.

• Bug Fixes
  • Strip providerMetadata from tool-output-available and tool-output-error in direct result.toUIMessageStream().
  • Preserve UI stream cancellation while transforming the direct stream.
  • Add regression tests, including cancellation coverage, for the direct UI stream path.

^{Written for commit 8e02def. Summary will update on new commits.}

Summary by CodeRabbit

• Bug Fixes

  • Provider metadata is no longer included in direct UI stream tool-output and error chunks sent to console clients.

• Tests

  • Added unit tests verifying metadata is stripped from tool output chunks and that stream cancellation still propagates correctly.

[changeset-bot]

🦋 Changeset detected

Latest commit: 8e02def

The changes in this PR will be included in the next version bump.

This PR includes changesets to release 1 package

Name	Type
@voltagent/core	Patch

Not sure what this means? Click here to learn what changesets are.

Click here if you're a maintainer who wants to add another changeset to this PR

[coderabbitai]

📝 Walkthrough

Walkthrough

This PR strips providerMetadata from tool-output chunks in the non-guardrail UI stream path. Helper functions filter metadata from tool-output-available and tool-output-error chunks before they reach console clients. A test verifies the stripping behavior and a changeset documents the patch release.

Changes

Provider Metadata Stripping for Tool-Output Chunks

Layer / File(s)

Summary

Provider metadata stripping implementation and test packages/core/src/agent/agent.ts, packages/core/src/agent/agent.spec.ts, .changeset/cold-socks-listen.md

Helper functions stripProviderMetadataFromToolOutputChunk and stripProviderMetadataFromDirectUIStream remove providerMetadata from tool-output-available and tool-output-error chunks when guardrail pipeline is absent. Unit test validates that providerMetadata is stripped from tool-output chunks but preserved on tool-input-available and text-delta chunks. Changeset documents the behavior change for patch release.

Estimated code review effort

🎯 2 (Simple) | ⏱️ ~12 minutes

Possibly related PRs

• VoltAgent/voltagent#1202: Both PRs address providerMetadata handling on UI stream tool-output chunks; #1202 forwards metadata in the guardrail pipeline path, while this PR strips it in the non-guardrail "direct" path to prevent console client validation errors.

Suggested reviewers

• omeraplak

Poem

🐰 Metadata whispers through the tool-output stream,
But browsers don't speak that tongued dream,
So we strip it clean on the direct path,
Leaving only the truth in its aftermath,
Console clients smile, no validation thrash! ✨

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 inconclusive)

Check name	Status	Explanation	Resolution
Description check	❓ Inconclusive	The PR description includes a clear summary, testing instructions, and links to the related issue #1290, but lacks explicit sections matching the template structure.	Consider reorganizing the description to explicitly follow the template sections: 'What is the current behavior?' and 'What is the new behavior?' would clarify the problem and solution.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title clearly and concisely describes the main change: stripping tool output provider metadata from the direct UI stream path.
Linked Issues check	✅ Passed	The code changes fully address the root cause described in #1290: provider metadata is stripped from tool-output-available and tool-output-error chunks in the direct UI stream path via TransformStream pipeline.
Out of Scope Changes check	✅ Passed	All changes are directly scoped to fixing #1290: modifying agent.ts to sanitize the direct UI stream, adding test coverage in agent.spec.ts, and creating a changeset entry.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[cubic-dev-ai]

1 issue found across 3 files

<sub>Reply with feedback, questions, or to request a fix.

Re-trigger cubic</sub>

[liuzemei]

Addressed the cancellation feedback in the latest commits: the direct UI stream sanitizer now uses a TransformStream pipeline so cancellation propagates to the source stream, and there is a regression test covering reader.cancel().