feat(integrations): pydantic-ai integration

[constantinius]

Description

Integration for Pydantic AI. Support for sync and async (streaming and non-streaming)

Issues

Closes https://linear.app/getsentry/issue/TET-1242/python-sdk-pydantic-ai-integration

Reminders

• Please add tests to validate your changes, and lint your code using tox -e linters.
• Add GH Issue ID & Linear ID (if applicable)
• PR title should use conventional commit style (feat:, fix:, ref:, meta:)
• For external contributors: CONTRIBUTING.md, Sentry SDK development docs, Discord community

[linear]

TET-1242 Python-SDK: Pydantic AI integration

[sentrivana]

Hey, added a commit to run the new test suite in ci c7e7ec2 -- the script updated the test matrix for other test suites too and some of them are now failing; I'll fix that on master and then we can rebase.

[sentrivana]

Unrelated ci failures fixed.

Do we have test apps/scripts for this? If so would be awesome if you could add them to https://github.com/getsentry/testing-sentry-python like Vjeran did for google-genai in getsentry/testing-sentry-python#1

[constantinius]

@sentrivana I've pushed all tests I have here

[alexander-alderman-webb]

Some initial comments since I am not sure I'll get to reviewing everything in detail. There should be reviews from other SDK members coming.

[sentrivana]

Had a first look and left some comments. Some general feedback:

• There's a lot of imports in functions, I've pointed out some but not all. Please go through all non-top-level imports and if it doesn't break anything, move them to the other imports at the top of the file.

• The code is not easy to follow, mostly because related parts are disconnected.

  • Big part of this is that span management is scattered, e.g. a span is created in a specific function/file, then __enter__ed in a different file, and the corresponding __exit__ is someplace else still. Sometimes, we don't even know if a span is active, like in sentry_sdk/integrations/pydantic_ai/patches/agent_run.py, which shouldn't be the case:

  # It could be that there is an "invoke agent" span still opencurrent_span=sentry_sdk.get_current_span() ifcurrent_spanisnotNoneandcurrent_span.timestampisNone: current_span.__exit__(None, None, None)

  It might be that most of this scattered management is simply a byproduct of how Pydantic AI/agentic flows work in general. But if we can do something to improve the flow we should.

  • In general you need to do a lot of jumping around in the code to follow a specific code path. I think a big part of this is the number of new files introduced by this PR. Can we consolidate these a bit? I.e., if functions from a single file are only used in one other file, can we move them there?

This would make the code more reviewable for me, since right now I'm struggling to grasp the bigger picture. 🙂

[cursor]

Bug: Optional Dependency Import Error

The import agents statement is now unconditional and outside the try...except block in both tools.py and utils.py. If the agents package is not installed, this raises an ImportError at module load time instead of gracefully raising DidNotEnable as expected for optional dependencies.

Additional Locations (1)

• sentry_sdk/integrations/openai_agents/patches/tools.py#L11-L17

[cursor]

Bug: Import Failure Causes isinstance TypeError

The isinstance check for FunctionTool can raise a TypeError. If the from agents import FunctionTool import fails, FunctionTool is set to None, which isinstance doesn't accept as a type.

Additional Locations (1)

• sentry_sdk/integrations/openai_agents/patches/tools.py#L39-L40

[cursor]

Bug: ImportError Handling Regression

The import agents statement was moved outside its try-except block in tools.py and utils.py. If the agents package is not installed, this now raises an unhandled ImportError instead of gracefully disabling the integration or setting related classes to None.

Additional Locations (1)

• sentry_sdk/integrations/openai_agents/patches/tools.py#L11-L17

[cursor]

Bug: Pydantic AI Integration Fails Without Installation

The pydantic_ai integration raises DidNotEnable at module import time when pydantic-ai isn't installed. This prevents the integration module from being imported, which breaks the SDK's integration discovery and can cause initialization failures.

[sentrivana]

Thanks for consolidating! Couple more comments, will try it out now.

[cursor]

Bug: Default Mismatch Causes Prompt Omission

The _should_send_prompts function uses getattr with a default of False for include_prompts. This conflicts with the PydanticAIIntegration's default of True for this setting, meaning prompts may not be sent when they should be by default if the attribute is missing.

[cursor]

Bug: Module-Level Handling Causes Integration Load Failure

The pydantic_ai import and DidNotEnable exception are handled at the module level. This prevents the integration from loading if pydantic-ai isn't installed, which goes against the expected pattern of raising DidNotEnable within setup_once().

[cursor]

Bug: Pydantic AI Integration Fails Without Explicit Installation

The pydantic_ai integration raises DidNotEnable at module import time if pydantic_ai isn't installed. This prevents the integration from initializing and can cause Sentry SDK setup to fail, even when the integration isn't explicitly enabled, differing from how other integrations handle optional dependencies.

[sentrivana]

Lemme know if you need help with mypy complaining etc

[cursor]

Bug: Missing Python Version Constraint Causes Installation Failures

The pydantic_ai integration config in config.py is missing its Python version constraint. The pydantic-ai package requires Python >=3.10 per releases.jsonl, so populate_tox generates test environments for unsupported Python versions, causing installation failures.

[cursor]

Bug: Sentry Instrumentation Missing for Sync Agent Runs

The _patch_agent_run function currently patches Agent.run, Agent.run_stream, and Agent.run_stream_events, but it misses Agent.run_sync. This means synchronous agent invocations won't create Sentry spans, even though the test suite suggests they should be instrumented.

comments addressed