Add missing Vercel AI SDK v6 tool approval part types by bendrucker · Pull Request #4388 · pydantic/pydantic-ai

bendrucker · 2026-02-20T22:00:18Z

Changes

Adds ToolApprovalRequestedPart, ToolApprovalRespondedPart, ToolOutputDeniedPart, and their dynamic counterparts to request_types.py so the adapter can load and dump all AI SDK v6 tool approval states (approval-requested, approval-responded, output-denied)
Preserves tool denial state through message round-trips by storing {'is_denied': True} in BaseToolReturnPart.metadata, exposed via an is_denied property. This is UI display state (not semantic content sent to the model), ensuring denied tools render correctly after dump/load cycles
Uses is_denied in the streaming path (_event_stream.py) to emit ToolOutputDeniedChunk directly from the part, replacing the previous _denied_tool_ids side-channel that scanned message history

Testing

Round-trip tests verify that denied tool state survives dump/load cycles for both dynamic and builtin tools, including intermediate assertions on the UI part types and states
Load path tests cover output-denied parts with and without an approval reason
Streaming tests verify ToolOutputDeniedChunk emission for denied dynamic tools via the full agent flow

References

Closes Vercel AI Elements build_run_input failing on deferred tool approval/denial. #4387

Pre-Review Checklist

Any AI generated code has been reviewed line-by-line by the human PR author, who stands by it.
No breaking changes in accordance with the version policy.
Linting and type checking pass per make format and make typecheck.
PR title is fit for the release changelog.

Pre-Merge Checklist

New tests for any fix or new behavior, maintaining 100% coverage.
Updated documentation for new features and behaviors, including docstrings for API docs.

bendrucker · 2026-02-20T22:46:17Z

I used #4390 to help reproduce the error using the AI SDK client from Node.js and then test the fix. I'll get that covering the full integration and clean it up shortly. I recognize that it's somewhat odd to be calling a Node.js test suite from Pytest.

We understand if you're hesitant to merge this. But regardless, it'll be helpful to me, versus clicking through an AI Elements frontend.

dmmihov · 2026-02-22T23:12:28Z

Hi Ben,

Have been testing this out today.
This now looks good on an initial approval/denial, but think it breaks on repeated usage.

For example, if we deny a tool use and the agent re-tries the tool I am seeing the following stack trace:
https://logfire-us.pydantic.dev/public-trace/667ea865-7767-461d-9d7b-b0b7786f963d?spanId=15382dd132f6f007

What seems to be happening is that on a tool denial, the agent seems to want to re-try the same tool (at least sometimes) and this so far makes sense.

When I put a breakpoint In _adapter when we call VercelAIAdapter.deferred_tool_results on line 172 where we do

for tool_call_id, approval in iter_tool_approval_responses(self.run_input.messages):

in self.run_input.messages[1] I see is the following:

UIMessage(id='tAtClhLObTutaP9v', role='assistant', metadata=None, parts=[
StepStartUIPart(type='step-start'),
ToolOutputDeniedPart(type='tool-create_journal_entry', tool_call_id='call_KdCaFjbCbFlBG4CRCqHJct9h', state='output-denied', input={'content': 'Agentic work complete after major uplift for deepagents and agent harness'}, provider_executed=None, call_provider_metadata=None, approval=ToolApprovalResponded(id='ea30f48d-88c8-46d7-b0f2-860372f6790d', approved=False, reason=None)), 
StepStartUIPart(type='step-start'), 
ToolOutputAvailablePart(type='tool-get_current_date', tool_call_id='call_8MgEpRPqVpLRmDp0xDWk8WV4', state='output-available', input={}, output="Today's date is Sunday, February 22, 2026", provider_executed=None, call_provider_metadata=None, preliminary=None, approval=None), 
StepStartUIPart(type='step-start'), 
ToolApprovalRespondedPart(type='tool-create_journal_entry', tool_call_id='call_WDh7cUdcl8WhCFXoMOyhQDMG', state='approval-responded', input={'content': 'Agentic work complete after major uplift for deepagents and agent harness on Sunday, February 22, 2026.'}, provider_executed=None, call_provider_metadata=None, approval=ToolApprovalResponded(id='796f069c-54b5-4421-9ab0-e1df9ae3a913', approved=False, reason=None))
])

(please ignore the get_current_date tool call, don't know why the agent decided to do that)

So we see a StepStartUIPart, a ToolOutputDeniedPart (first denial), StepStartUIPart, ToolApprovalRespondedPart(second denial we just sent) as parts and both denials match
if isinstance(part, _TOOL_PART_TYPES) and isinstance(part.approval, ToolApprovalResponded):

So we have deferred_tool_results containing both denials by their id.
When we then go back to the agent graph in
return await self._handle_deferred_tool_results(self.deferred_tool_results, messages, ctx)
our self.deferred_tool_results contains both denials.

This seems to then be an issue for further validations.

The immediate simplest fix to me seems to be in iter_tool_approval_responses to specifically filter for only ToolApprovalRespondedPart and DynamicToolApprovalRespondedPart respectively (this seems to work locally), but I am not 100% convinced this will not cause further issues.

The other option is probably to ensure that a UI treats a denial as a end of the message and a subsequent request for the same tool as a completely new message, so there two are processed separately. I am not sure if this is the intended behaviour though.

bendrucker · 2026-02-23T02:27:44Z

Thanks, I will look into this either tomorrow or Tuesday.

The other option is probably to ensure that a UI treats a denial as a end of the message

In theory the denial reason could contain a user message and so this would be true but I need to spend more time looking into how AI SDK is expecting these on the frontend.

bendrucker · 2026-02-24T03:44:39Z

@dmmihov just pushed a change

The immediate simplest fix to me seems to be in iter_tool_approval_responses to specifically filter for only ToolApprovalRespondedPart and DynamicToolApprovalRespondedPart respectively

It does exactly this and adjusts the unit tests to match.

I repro'd this in #4390 as well.

I should have time tomorrow to test this end to end. But in the mean time please do try it locally if you have time!

dmmihov · 2026-02-24T11:37:39Z

Thank you, Ben.

I will test this out tonight.

dmmihov · 2026-02-24T20:09:24Z

Confirm this seems to be working for repeat deferrals.

Not familiar enough with the rest of the adapter usage to confirm this will not break anything else though.

bendrucker · 2026-02-25T00:12:46Z

Awesome, thank you! I am reasonably confident in this specific fix. I also have plans later this week to properly integrate these new features into a production application. So to the extent there are any other edge cases or subtleties I'll be looking closely. Definitely planning to address the issue the display issue for approval state as well:

#4388 (comment)

bendrucker · 2026-02-27T04:04:29Z

Fixed the uncovered built-in tool branch, should be ready to go now

devin-ai-integration

Devin Review found 1 new potential issue.

View 12 additional findings in Devin Review.

DouweM · 2026-03-02T23:14:18Z

@bendrucker Please have a look at the merge conflicts; I just merged #4196

DouweM · 2026-03-03T20:16:15Z

pydantic_ai_slim/pydantic_ai/messages.py

    timestamp: datetime = field(default_factory=_now_utc)
    """The timestamp, when the tool returned."""

+    status: Literal['success', 'error', 'denied'] = 'success'


I like the outcome name from the comment better than status!

pydantic_ai_slim/pydantic_ai/ui/vercel_ai/_adapter.py

DouweM · 2026-03-03T20:19:19Z

pydantic_ai_slim/pydantic_ai/ui/vercel_ai/_adapter.py

+                                type=tool_name,
+                                tool_call_id=part.tool_call_id,
+                                input=_safe_args_as_dict(part),
+                                state='output-denied',


This is implied by the class right? We shouldn't need to set it explicitly

pydantic_ai_slim/pydantic_ai/ui/vercel_ai/_adapter.py

DouweM · 2026-03-03T20:20:28Z

pydantic_ai_slim/pydantic_ai/ui/vercel_ai/_adapter.py

+                        )
+                    else:
+                        ui_parts.append(
+                            ToolOutputAvailablePart(


We can also handle ToolOutputErrorPart here now right?

pydantic_ai_slim/pydantic_ai/ui/vercel_ai/_event_stream.py

DouweM · 2026-03-03T20:22:23Z

tests/test_vercel_ai.py

+    should be extracted.
+    """
+    from pydantic_ai.ui.vercel_ai._utils import iter_tool_approval_responses
+    from pydantic_ai.ui.vercel_ai.request_types import DynamicToolOutputDeniedPart


Imports at the top of the file please

DouweM · 2026-03-03T20:23:23Z

pydantic_ai_slim/pydantic_ai/ui/vercel_ai/_adapter.py

-                                elif isinstance(part, ToolOutputErrorPart):  # pragma: no branch
-                                    output = {'error_text': part.error_text, 'is_error': True}
+                                if isinstance(part, ToolOutputErrorPart):
+                                    output: Any = {'error_text': part.error_text, 'is_error': True}


We can just store error_text as output now, right?

dmmihov · 2026-03-03T22:09:54Z

Not a contributer yet, so can't push to this branch.
Hopefully this helps though, started addressing some of the comments:
bendrucker#2

bendrucker · 2026-03-04T01:48:53Z

Thanks @dmmihov, merged, will have another look at this tonight if I have time

Co-authored-by: devin-ai-integration[bot] <158243242+devin-ai-integration[bot]@users.noreply.github.com>

… errors - Use ToolReturnPart(outcome='error') instead of RetryPromptPart in load_messages for tool errors from Vercel AI format - Store builtin tool error_text directly as content instead of {'error_text': ..., 'is_error': True} wrapper - Handle outcome='error' in dump path, emitting ToolOutputErrorPart - Add backward-compat is_error content check for old serialized data - Handle ToolReturnPart(outcome='error') in event stream - Extract _dump_tool_call_part helper to reduce C901 complexity - Remove redundant state= assignments implied by class defaults - Move inline test imports to top of file

…at import time

adtyavrdhn · 2026-03-04T15:24:26Z

/gemini Summary

DouweM

@bendrucker Thanks Ben, just a small final change and we're good to go!

pydantic_ai_slim/pydantic_ai/messages.py

pydantic_ai_slim/pydantic_ai/ui/vercel_ai/_adapter.py

…al default - Rename 'error' → 'failed' in BaseToolReturnPart.outcome to align with the upcoming ToolFailed exception (pydantic#2586) - Use ToolDenied().message instead of hardcoded string in _denial_reason() for consistency

Co-authored-by: devin-ai-integration[bot] <158243242+devin-ai-integration[bot]@users.noreply.github.com>

Co-authored-by: Douwe Maan <douwe@pydantic.dev> Co-authored-by: Darin Mihov <dmmihov@gmail.com> Co-authored-by: devin-ai-integration[bot] <158243242+devin-ai-integration[bot]@users.noreply.github.com>

github-actions bot added size: M Medium PR (101-500 weighted lines) bug Report that something isn't working, or PR implementing a fix labels Feb 20, 2026

This comment was marked as resolved.

Sign in to view

bendrucker mentioned this pull request Feb 20, 2026

Add AI SDK E2E integration tests #4390

Closed

6 tasks

bendrucker mentioned this pull request Feb 20, 2026

Vercel AI Elements build_run_input failing on deferred tool approval/denial. #4387

Closed

2 tasks

dsfaccini added the awaiting author revision label Feb 23, 2026

dsfaccini self-assigned this Feb 23, 2026

This comment was marked as resolved.

Sign in to view

bendrucker force-pushed the fix-ai-sdk-v6-approval-types branch 4 times, most recently from 2f7df41 to e67bde2 Compare February 25, 2026 15:19

DouweM changed the title ~~Add missing AI SDK v6 tool approval part types~~ Add missing Vercel AI SDK v6 tool approval part types Feb 27, 2026

DouweM added the auto-review label Feb 27, 2026

This comment was marked as resolved.

Sign in to view

github-actions bot removed the auto-review label Feb 27, 2026

DouweM mentioned this pull request Feb 27, 2026

ToolReturnPart drops error status — is_error hardcoded to False for Anthropic/Bedrock, AG-UI ToolMessage.error not mapped #4363

Open

1 task

devin-ai-integration bot reviewed Mar 2, 2026

View reviewed changes

This comment was marked as resolved.

Sign in to view

DouweM added the auto-review label Mar 2, 2026

github-actions bot removed the auto-review label Mar 2, 2026

Use status field for error detection in builtin tool dump path

50b33ba

DouweM added the auto-review label Mar 3, 2026

github-actions bot removed the auto-review label Mar 3, 2026

Merge branch 'main' into fix-ai-sdk-v6-approval-types

c475fd8

This comment was marked as resolved.

Sign in to view

DouweM requested changes Mar 3, 2026

View reviewed changes

review feedback

caaa672

This comment was marked as resolved.

Sign in to view

Apply suggestions from code review

993dbc9

Co-authored-by: devin-ai-integration[bot] <158243242+devin-ai-integration[bot]@users.noreply.github.com>

This comment was marked as resolved.

Sign in to view

github-actions bot added size: L Large PR (501-1500 weighted lines) and removed size: M Medium PR (101-500 weighted lines) labels Mar 4, 2026

This comment was marked as resolved.

Sign in to view

bendrucker added 2 commits March 3, 2026 20:09

Lazy-import _web in pydantic_ai.ui to avoid starlette dependency …

b2836f6

…at import time

Add test for builtin tool return error event stream coverage

ff9e574

DouweM added the auto-review label Mar 4, 2026

github-actions bot removed the auto-review label Mar 4, 2026

DouweM added the auto-review label Mar 4, 2026

github-actions bot removed the auto-review label Mar 4, 2026

DouweM requested changes Mar 4, 2026

View reviewed changes

pydantic_ai_slim/pydantic_ai/messages.py Outdated Show resolved Hide resolved

pydantic_ai_slim/pydantic_ai/ui/vercel_ai/_adapter.py Outdated Show resolved Hide resolved

This comment was marked as resolved.

Sign in to view

Update tests/test_vercel_ai.py

700cce6

Co-authored-by: devin-ai-integration[bot] <158243242+devin-ai-integration[bot]@users.noreply.github.com>

DouweM merged commit 54125a3 into pydantic:main Mar 6, 2026
38 checks passed

Conversation

bendrucker commented Feb 20, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Changes

Testing

References

Pre-Review Checklist

Pre-Merge Checklist

Uh oh!

This comment was marked as resolved.

Uh oh!

bendrucker commented Feb 20, 2026

Uh oh!

dmmihov commented Feb 22, 2026 • edited by DouweM Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

bendrucker commented Feb 23, 2026

Uh oh!

bendrucker commented Feb 24, 2026

Uh oh!

dmmihov commented Feb 24, 2026

Uh oh!

dmmihov commented Feb 24, 2026

Uh oh!

bendrucker commented Feb 25, 2026

Uh oh!

This comment was marked as resolved.

Uh oh!

bendrucker commented Feb 27, 2026

Uh oh!

This comment was marked as resolved.

Uh oh!

devin-ai-integration bot left a comment

Choose a reason for hiding this comment

Uh oh!

This comment was marked as resolved.

Uh oh!

DouweM commented Mar 2, 2026

Uh oh!

This comment was marked as resolved.

Uh oh!

DouweM Mar 3, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

DouweM Mar 3, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

DouweM Mar 3, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

DouweM Mar 3, 2026

Choose a reason for hiding this comment

Uh oh!

DouweM Mar 3, 2026

Choose a reason for hiding this comment

Uh oh!

dmmihov commented Mar 3, 2026

Uh oh!

bendrucker commented Mar 4, 2026

Uh oh!

This comment was marked as resolved.

Uh oh!

This comment was marked as resolved.

Uh oh!

This comment was marked as resolved.

Uh oh!

adtyavrdhn commented Mar 4, 2026

Uh oh!

DouweM left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

This comment was marked as resolved.

Uh oh!

Uh oh!

Reviewers

bendrucker commented Feb 20, 2026 •

edited

Loading

dmmihov commented Feb 22, 2026 •

edited by DouweM

Loading