feat(openai-agents): Support span streaming by alexander-alderman-webb · Pull Request #6404 · getsentry/sentry-python

alexander-alderman-webb · 2026-05-26T06:10:05Z

Description

Use sentry_sdk.traces.start_span, replace Span.set_data() with StreamedSpan.set_attribute() and Span.set_status(SPANSTATUS.INTERNAL_ERROR) with StreamedSpan.status = SpanStatus.ERROR when in span streaming mode.

Parametrize tests on the trace lifecycle option.

Issues

Closes #6041

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

github-actions · 2026-05-26T06:11:12Z

Codecov Results 📊

✅ 282 passed | Total: 282 | Pass Rate: 100% | Execution Time: 44.17s

All tests are passing successfully.

❌ Patch coverage is 6.48%. Project has 14819 uncovered lines.

Files with missing lines (9)

File	Patch %	Lines
`agent_run.py`	4.76%	⚠️ 140 Missing
`models.py`	0.00%	⚠️ 105 Missing
`runner.py`	0.00%	⚠️ 98 Missing
`utils.py`	16.16%	⚠️ 83 Missing
`invoke_agent.py`	0.00%	⚠️ 46 Missing
`execute_tool.py`	0.00%	⚠️ 34 Missing
`ai_client.py`	20.00%	⚠️ 28 Missing
`handoff.py`	0.00%	⚠️ 17 Missing
`agent_workflow.py`	50.00%	⚠️ 6 Missing

Generated by Codecov Action

cursor

Cursor Bugbot has reviewed your changes and found 2 potential issues.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit bf37fa2. Configure here.}

ericapisani

Overall looking good but there's a few things around the potential for some bugs with a couple of the conditionals that were updated, and the inconsistent way that we're processing spans in tests (some aren't asserting the total amount of spans and the order in which they appear, others are).

ericapisani · 2026-05-26T13:15:00Z

+        span is None
+        or isinstance(span, StreamedSpan)
+        and span.end_timestamp is not None
+        or not isinstance(span, StreamedSpan)
+        and span.timestamp is not None


This is a bit difficult to read and I think could lead to an accidental bug because of orders or precedence with the and and or keywords (the former taking evaluation precedence over the latter).

Reading this, I'm not sure if the intention was to have the statement evaluate as isinstance(span, StreamedSpan) and span.end_timestamp is not None rather than (span is None or isinstance(span, StreamedSpan)) and ...

Fair point about lack of intentionality.
The intention is for the condition to evaluate to true if the end timestamp has not yet been set, indicating that the span is still active. The complex conditional is used because Span.timestamp and StreamedSpan.end_timestamp represent the end timestamp (i.e., because the instance variable was renamed).
I've added brackets in e243b86

ericapisani · 2026-05-26T13:19:34Z

+            name=f"chat {model_name}",
+            origin=SPAN_ORIGIN,
+        )
+        # TODO-anton: remove hardcoded stuff and replace something that also works for embedding and so on


I'm not sure how useful this comment is anymore - any chance we can remove it?

I think this is still a bug 😬.
Created #6417 to track.

ericapisani · 2026-05-26T14:22:44Z

+    span_streaming = has_span_streaming_enabled(sentry_sdk.get_client().options)
+    if span_streaming:
+        with sentry_sdk.traces.start_span(
+            name=f"handoff from {from_agent.name} to {to_agent_name}",


We have a gen_ai.agent.name attribute and I'm thinking this should maybe be added as an attribute with a value of from_agent.name (since it's the originator of the span in terms of the action being taken)

I'd have the same instinct to add an attribute to avoid users matching on the name.

However, we're removing the workflow span entirely per the RFC.

ericapisani · 2026-05-26T14:27:39Z

+            assert result.final_output == "Hello, how can I help you?"
+
+        sentry_sdk.flush()
+        spans = [item.payload for item in items]


Even if we're only interested in the invoke agent span and ai client span, we should still assert the number of expected spans here since this should be stable. If there are more/less spans, we'd want this test to fail

I'd consider this out of scope for this PR because the existing test did not assert the total number of spans the goal of the PR is to port the integration to support span streaming (and make the same test guarantees as before).

I'm not sure if it's always been the intention to open another pull request to introduce the stricter checks, but if so, please add that in the pull request descriptions going forward so it's clear if it's an intentional decision or an accidental oversight.

As we've been introducing these stricter checks as part of other streamed span integration migrations and code review comments have been left when it's not been done, it appears to me that these stricter checks are part of the migration process, not out of scope.

In the case of this change set it's especially noticeable because it's being applied in some places and not others.

ericapisani · 2026-05-26T14:34:11Z

+        if "string" in param_id and instructions is None:  # type: ignore
+            assert "gen_ai.system_instructions" not in ai_client_span["data"]
+
+            assert invoke_agent_span["data"][
+                "gen_ai.request.messages"
+            ] == safe_serialize(
+                [
+                    {
+                        "content": [{"text": "Test input", "type": "text"}],
+                        "role": "user",
+                    },
+                ]
+            )
+
+        elif "string" in param_id:
+            assert ai_client_span["data"][
+                "gen_ai.system_instructions"
+            ] == safe_serialize(
+                [
+                    {
+                        "type": "text",
+                        "content": "You are a coding assistant that talks like a pirate.",
+                    },
+                ]
+            )
+        elif "blocks_no_type" in param_id and instructions is None:  # type: ignore
+            assert ai_client_span["data"][
+                "gen_ai.system_instructions"
+            ] == safe_serialize(
+                [
+                    {"type": "text", "content": "You are a helpful assistant."},
+                ]
+            )
+        elif "blocks_no_type" in param_id:
+            assert ai_client_span["data"][
+                "gen_ai.system_instructions"
+            ] == safe_serialize(
+                [
+                    {
+                        "type": "text",
+                        "content": "You are a coding assistant that talks like a pirate.",
+                    },
+                    {"type": "text", "content": "You are a helpful assistant."},
+                ]
+            )
+        elif "blocks" in param_id and instructions is None:  # type: ignore
+            assert ai_client_span["data"][
+                "gen_ai.system_instructions"
+            ] == safe_serialize(
+                [
+                    {"type": "text", "content": "You are a helpful assistant."},
+                ]
+            )
+        elif "blocks" in param_id:
+            assert ai_client_span["data"][
+                "gen_ai.system_instructions"
+            ] == safe_serialize(
+                [
+                    {
+                        "type": "text",
+                        "content": "You are a coding assistant that talks like a pirate.",
+                    },
+                    {"type": "text", "content": "You are a helpful assistant."},
+                ]
+            )
+        elif "parts_no_type" in param_id and instructions is None:
+            assert ai_client_span["data"][
+                "gen_ai.system_instructions"
+            ] == safe_serialize(
+                [
+                    {"type": "text", "content": "You are a helpful assistant."},
+                    {"type": "text", "content": "Be concise and clear."},
+                ]
+            )
+        elif "parts_no_type" in param_id:
+            assert ai_client_span["data"][
+                "gen_ai.system_instructions"
+            ] == safe_serialize(
+                [
+                    {
+                        "type": "text",
+                        "content": "You are a coding assistant that talks like a pirate.",
+                    },
+                    {"type": "text", "content": "You are a helpful assistant."},
+                    {"type": "text", "content": "Be concise and clear."},
+                ]
+            )
+        elif instructions is None:  # type: ignore
+            assert ai_client_span["data"][
+                "gen_ai.system_instructions"
+            ] == safe_serialize(
+                [
+                    {"type": "text", "content": "You are a helpful assistant."},
+                    {"type": "text", "content": "Be concise and clear."},
+                ]
+            )


This is a large conditional block that's not the easiest to read because of there sometimes being only param_id, sometimes both param_id and instructions, and the one case with instructions.

Can we look to re-organize this/break it up a bit? Maybe have a general conditional grouping to cover cases where instructions are None with the relevant combos with param_id in there, and the other cases involving param_id as a separate grouping?

I agree that it's hard to read and needs improvement.
The underlying logic is complex because instructions can be provided both in the input array and in the dedicated instructions parameter when using the Responses API.
I would like to avoid increasing indentation though.

I'll just split up some tests and make better use of pytest.mark.parametrize (only tomorrow though).

ericapisani · 2026-05-26T14:37:47Z

+        invoke_agent_span = next(
+            span
+            for span in spans
+            if span["attributes"]["sentry.op"] == OP.GEN_AI_INVOKE_AGENT
+        )
+        ai_client_span = next(
+            span for span in spans if span["attributes"]["sentry.op"] == OP.GEN_AI_CHAT
+        )


As I'm reviewing, I'm noticing some test cases with this pattern, and other test cases with this pattern.

Is there a reason why some are implemented one way vs the other? Is it to avoid assigning spans to variables that won't be asserted on?

I have no idea what the authors of the integration were thinking. I have not made a conscious decision in favor of one or the other style.

alexander-alderman-webb added 2 commits May 22, 2026 11:59

fix(openai-agents): Remove redundant hosted MCP tool spans

6f67df0

feat(openai-agents): Support span streaming

0f53ac0

alexander-alderman-webb added 3 commits May 26, 2026 08:12

.

5a5f81d

mypy

c4a5366

mypy2

b2b0c34

Base automatically changed from webb/openai-agents/remote-tool to master May 26, 2026 07:31

merge master

bf37fa2

alexander-alderman-webb marked this pull request as ready for review May 26, 2026 07:41

alexander-alderman-webb requested a review from a team as a code owner May 26, 2026 07:41

cursor Bot reviewed May 26, 2026

View reviewed changes

Comment thread sentry_sdk/integrations/openai_agents/patches/runner.py

Comment thread sentry_sdk/integrations/openai_agents/utils.py

alexander-alderman-webb added 4 commits May 26, 2026 10:11

handle streamed span invoke agent

53a84ae

truncate invoke agent attributes

dcdeb8d

fix bool logic

2c1bd19

merge master

e9413c5

ericapisani requested changes May 26, 2026

View reviewed changes

add brackets to bool logic

e243b86

alexander-alderman-webb mentioned this pull request May 26, 2026

Create an enum for possible gen_ai.operation.name values #6416

Open

drop transaction reference in span streaming comment

931aa30

Conversation

alexander-alderman-webb commented May 26, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Description

Issues

Reminders

Uh oh!

github-actions Bot commented May 26, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codecov Results 📊

Uh oh!

cursor Bot left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

ericapisani left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

alexander-alderman-webb commented May 26, 2026 •

edited

Loading

github-actions Bot commented May 26, 2026 •

edited

Loading