Set ToolRetryError message

[jamesaud]

Summary

ToolRetryError was not passing a message to Exception.__init__(), causing str(error) to return an empty string.

This came up when the agent loop crashes due to max tool retries, which breaks error monitoring tools like Sentry or Langsmith that rely on str(exception) to display and group errors.

Problem

from pydantic_ai.exceptions import ToolRetryError
from pydantic_ai.messages import RetryPromptPart

part = RetryPromptPart(content='Test error', tool_name='test')
error = ToolRetryError(part)

str(error)  # Returns '' (empty string)
error.args  # Returns ()

Solution

Pass tool_retry.model_response() to the parent Exception.__init__():

str(error)  # Now returns 'Test error\n\nFix the errors and try again.'

Changes

• pydantic_ai_slim/pydantic_ai/exceptions.py: Pass message to parent class
• tests/test_exceptions.py: Add ToolRetryError to hashable test + add dedicated string representation test

[DouweM]

Can we use tool_retry.content that contains just the error message, without the surrounding text instructing the model to retry?

Besides a str, content can also be a list of Pydantic error details, so we would need to stringify/dump those as to JSON.

[jamesaud]

Great, thanks for the info! Using tool_retry.content now. Added a bit of formatting code to make list[ErrorDetail] more human readable.

[jamesaud]

Plz wait for re-review - need to fix CI.

[DouweM]

This is a bit lossy because we don't get to see the original input or error type. Can you see if we can use the same rendering that is used when Pydantic ValidationError is raised and printed in terminal? We get the ErrorDetails from a ValidationError here, so maybe we can pass in the string message here where we still have that error:

pydantic-ai/pydantic_ai_slim/pydantic_ai/_tool_manager.py

Lines 184 to 190 in d2b08ad

	if isinstance(e, ValidationError):
	m = _messages.RetryPromptPart(
	tool_name=name,
	content=e.errors(include_url=False, include_context=False),
	tool_call_id=call.tool_call_id,
	)
	e = ToolRetryError(m)

[jamesaud]

Thank you for the suggestion! I simplified the scope of the PR to explicitly pass str(e) from the call sites.
If callers do not provide message, then it will be blank.

Another option would be to fallback to lossy behavior as a backup if message is not passed.

[DouweM]

I wouldn't mind setting message automatically if isinstance(tool_retry.content, str)

[DouweM]

We should test that it is set correctly in the ValidationError case

[DouweM]

If ToolRetryError.__init__ is able to access the __cause__ that's set by from e (I'm not sure if it can), we could do this automatically