fix(google-genai): rebuild streamed interactions.create output from step events by sangkyoonnam · Pull Request #951 · open-telemetry/opentelemetry-python-genai

sangkyoonnam · 2026-10-22T13:48:36Z

Description

Streamed interactions.create spans rebuild captured output messages from step events because interaction.completed carries no steps; all four interaction stream wrappers share that handling and discard captured stream output if accumulation or rebuilding raises an exception. Streamed and non-streamed create derive gen_ai.response.finish_reasons from the interaction status on model and agent spans, regardless of content capture; in_progress or a missing status gives no finish reason.

Agent streams are covered by mocked tests only.

Type of change

Bug fix (non-breaking change which fixes an issue)

How has this been tested?

Cassette tests: streamed text checked against the step deltas, a streamed tool call ending in requires_action (sync + async, capture on and off), errors injected mid-stream, a caller-side raise, and an early close().
Unit tests cover non-streamed finish reasons by status, a stream that ends without a completion event, payloads that fail to accumulate or rebuild, and agent spans. With only interactions.py replaced by main's, 28 tests in tests/interactions fail on latest and on v2; the error tests still pass.
py310-test-instrumentation-google-genai-v2: 364 passed. py314-test-instrumentation-google-genai-latest: 693 passed. py310-test-instrumentation-google-genai-oldest: 279 passed.

Checklist

Followed the style guidelines of this project
Changelog updated if the change requires an entry
Unit tests added

…ns.create Assisted-by: Claude Opus 5.5

Co-authored-by: Liudmila Molkova <neskazu@gmail.com>

…tests Assisted-by: Claude Opus 5.5

…tep events Assisted-by: Claude Opus 5.5

opentelemetry-pr-dashboard · 2026-10-22T13:49:26Z

Pull request dashboard status

Waiting on the author · refreshed 2026-10-23 20:38 UTC

Investigate required status check failures.

Status above doesn't look right?

Just replied or pushed? Anything around or after the refresh time above may not be picked up yet — give it a few minutes.
Should this be with reviewers? Comment /dashboard route:reviewers to route it to them.
Anything wrong — including the routing? Report it with what you expected; it helps us improve the dashboard.

Copilot

🟢 Approval recommended

The focused stream-handling changes have relevant regression coverage, with no unresolved blocking findings.

No open issues 🎉

What changed in this PR

This PR builds on #947 to restore output messages and finish reasons for streamed Google GenAI interactions.create calls within the existing instrumentation.

Changes:

Reconstructs streamed output from step events using shared handling across all four interaction stream wrappers.
Records status-derived finish reasons for model and agent streams independently of content capture.
Adds sync/async cassette regressions and mocked coverage for incomplete streams, tool-call finish reasons, and agent spans.

Paths below are relative to instrumentation/opentelemetry-instrumentation-google-genai/.

File	Description
`tests/interactions/util.py`	Corrects the mocked completion event name.
`tests/interactions/test_e2e.py`	Verifies streamed output, metadata, and capture-independent finish reasons.
`tests/interactions/cassettes/test_sync_interactions_create_stream.yaml`	Adds a recorded synchronous stream.
`tests/interactions/cassettes/test_async_interactions_create_stream.yaml`	Adds a recorded asynchronous stream.
`tests/interactions/base.py`	Adds status, agent, and incomplete-stream regression cases.
`src/opentelemetry/instrumentation/google_genai/interactions.py`	Shares stream handling and reconstructs output with finish reasons.
`.changelog/951.fixed`	Documents restored output and finish reasons.
`.changelog/947.fixed`	Documents the underlying completion-event fix.

_{Review effort: Balanced}

lmolkova · 2026-10-22T21:18:31Z

    if telemetry_handler.should_capture_content():
-        invocation.output_messages = _interactions_response_to_messages(
-            response
+        # A streamed completion event carries no steps, so a stream's output
+        # comes from its step events and its finish reason from the status.
+        finish_reason = (
+            _interaction_finish_reason(_get_field(response, "status"))
+            if streamed_parts is not None
+            else "stop"
+        )
+        messages = _interactions_response_to_messages(
+            response, finish_reason=finish_reason
        )
+        if streamed_parts and not _has_content(messages):
+            messages = _streamed_output_messages(
+                streamed_parts, finish_reason=finish_reason
+            )
+        invocation.output_messages = messages


Non-streamed create still reports "stop" for every status (e.g. requires_action, or in_progress with background=True), and its agent spans get no gen_ai.response.finish_reasons, so stream and non-stream now disagree. What do you think about deriving it from the status for both paths here? Then the explicit finish_reasons block in _CreateStreamMixin._record_interaction can go.

Suggested change

if telemetry_handler.should_capture_content():

invocation.output_messages = _interactions_response_to_messages(

response

# A streamed completion event carries no steps, so a stream's output

# comes from its step events and its finish reason from the status.

finish_reason = (

_interaction_finish_reason(_get_field(response, "status"))

if streamed_parts is not None

else "stop"

)

messages = _interactions_response_to_messages(

response, finish_reason=finish_reason

)

if streamed_parts and not _has_content(messages):

messages = _streamed_output_messages(

streamed_parts, finish_reason=finish_reason

)

invocation.output_messages = messages

finish_reason = _interaction_finish_reason(_get_field(response, "status"))

if finish_reason:

invocation.finish_reasons = [finish_reason]

if telemetry_handler.should_capture_content():

messages = _interactions_response_to_messages(

response, finish_reason=finish_reason

)

if streamed_parts and not _has_content(messages):

messages = _streamed_output_messages(

streamed_parts, finish_reason=finish_reason

)

invocation.output_messages = messages

Done in 5a3ec52. Streamed and non-streamed create now derive finish reasons from the status on both model and agent spans, regardless of content capture, and the explicit block in _CreateStreamMixin._record_interaction is gone. in_progress or a missing status gives no finish reason. Before this, a non-streamed model span with content capture on reported stop even for in_progress.

lmolkova · 2026-10-22T21:18:32Z

+        "os.environ",
+        {"OTEL_INSTRUMENTATION_GENAI_CAPTURE_MESSAGE_CONTENT": "SPAN_ONLY"},
+    )
+    def test_streaming_output_message_finish_reason_matches_status(


The rebuild-from-steps path is only covered by a plain-text cassette. The other new streaming tests use hand-built MagicMock events, which don't match the SDK's real event types, so a wrong field name wouldn't fail them. Also, streamed create still has no tests for a mid-stream error, a caller-side error, or close/early-close, and this PR changes how those finish.

What do you think about following the OpenAI/Anthropic pattern?

Record a streamed tool-call cassette (sync + async) that ends in requires_action, and check the rebuilt tool_call part and finish_reasons == ("tool_calls",) with content capture on and off.

Inject failures on top of a recorded cassette by wrapping the SDK stream in a delegate that raises after N events, like ErrorInjectingStreamDelegate and _AsyncErrorInjectingStreamDelegate. Add a raise inside the with block for the caller-side case.

Mocks are still fine for a stream that ends without a completion event.

Added in 786d15b: sync and async tool-call cassettes ending in requires_action (capture on and off), errors injected after 0 and 7 events of the recorded text streams, a raise inside with, and early close.

The tool-call and capture-on early-close tests fail with main's interactions.py. The error tests pass on both, since a failed stream still drops partial output. Should keeping it on failure, as Anthropic does, be a separate PR?

…teractions-create-stream-output # Conflicts: # instrumentation/opentelemetry-instrumentation-google-genai/src/opentelemetry/instrumentation/google_genai/interactions.py # instrumentation/opentelemetry-instrumentation-google-genai/tests/interactions/test_e2e.py

… status Co-authored-by: lmolkova <2347409+lmolkova@users.noreply.github.com> Assisted-by: Claude Opus 5.5

…rors and early close Assisted-by: Claude Opus 5.5

…mulation fails Assisted-by: Claude Opus 5.5

sangkyoonnam and others added 5 commits October 22, 2026 09:12

fix(google-genai): match interaction.completed in streamed interactio…

d137717

…ns.create Assisted-by: Claude Opus 5.5

docs: add changelog fragment for open-telemetry#947

ac131d3

Apply suggestion from @lmolkova

30dc1df

Co-authored-by: Liudmila Molkova <neskazu@gmail.com>

test(google-genai): add cassette-backed streamed interactions.create …

619c347

…tests Assisted-by: Claude Opus 5.5

fix(google-genai): rebuild streamed interactions.create output from s…

5ee32b1

…tep events Assisted-by: Claude Opus 5.5

sangkyoonnam requested a review from a team as a code owner October 22, 2026 13:48

Copilot AI balanced review requested due to automatic review settings October 22, 2026 13:48

github-project-automation Bot added this to GenAI Semantic Conventions and Instrumentation libraries Oct 22, 2026

chore(google-genai): add changelog for open-telemetry#951

ad97fdf

opentelemetry-pr-dashboard Bot mentioned this pull request Oct 22, 2026

Pull Request Dashboard #171

Open

Copilot started reviewing on behalf of sangkyoonnam October 22, 2026 13:49

View session

Copilot AI reviewed Oct 22, 2026

View reviewed changes

lmolkova reviewed Oct 22, 2026

View reviewed changes

sangkyoonnam and others added 5 commits October 23, 2026 11:11

fix(google-genai): derive interactions.create finish reasons from the…

5a3ec52

… status Co-authored-by: lmolkova <2347409+lmolkova@users.noreply.github.com> Assisted-by: Claude Opus 5.5

test(google-genai): cover streamed interactions.create tool calls, er…

786d15b

…rors and early close Assisted-by: Claude Opus 5.5

chore(google-genai): update changelog for open-telemetry#951

1d228c3

fix(google-genai): keep the interaction stream going when output accu…

4e485d5

…mulation fails Assisted-by: Claude Opus 5.5

Navigation Menu

Repository navigation

fix(google-genai): rebuild streamed interactions.create output from step events - #951

fix(google-genai): rebuild streamed interactions.create output from step events#951
sangkyoonnam wants to merge 11 commits into
open-telemetry:mainfrom
sangkyoonnam:fix/google-genai-interactions-create-stream-output

sangkyoonnam commented Oct 22, 2026 •

edited

Loading

Uh oh!

opentelemetry-pr-dashboard Bot commented Oct 22, 2026 •

edited

Loading

Uh oh!

Copilot AI left a comment

Uh oh!

lmolkova Oct 22, 2026

Uh oh!

sangkyoonnam Oct 23, 2026

Uh oh!

lmolkova Oct 22, 2026

Uh oh!

sangkyoonnam Oct 23, 2026

Uh oh!

This branch has not been deployed

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Conversation

sangkyoonnam commented Oct 22, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Description

Type of change

How has this been tested?

Checklist

Uh oh!

opentelemetry-pr-dashboard Bot commented Oct 22, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Pull request dashboard status

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

🟢 Approval recommended

Uh oh!

lmolkova Oct 22, 2026

Choose a reason for hiding this comment

Uh oh!

sangkyoonnam Oct 23, 2026

Choose a reason for hiding this comment

Uh oh!

lmolkova Oct 22, 2026

Choose a reason for hiding this comment

Uh oh!

sangkyoonnam Oct 23, 2026

Choose a reason for hiding this comment

Uh oh!

This branch has not been deployed

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

sangkyoonnam commented Oct 22, 2026 •

edited

Loading

opentelemetry-pr-dashboard Bot commented Oct 22, 2026 •

edited

Loading