[ishan-1010]

Link to Issue or Description of Change

• Closes: LiteLlm with SSE streaming records the answer twice in gen_ai.output.messages #7630

Problem:
LiteLlm streams chunks without a finish reason and then yields their aggregate with one. record_llm_response keeps recording until a finish reason arrives, so gen_ai.output.messages held every chunk and then the whole answer again.

Solution:
As suggested in #7630: when a complete response with a finish reason follows only partial chunks without one, drop the chunks' output messages and keep the aggregate. Streams whose last chunk carries the finish reason (Gemini) and streams that fail before their aggregate keep their chunks.

Testing Plan

Unit Tests:

• I have added or updated unit tests for my change.
• All unit tests pass locally.

New tests in tests/unittests/telemetry/test_instrumentation.py:

• test_aggregate_after_unfinished_chunks_is_recorded_once (fails on main: 3 messages)
• test_tool_call_aggregate_replaces_its_partial_call_chunks (fails on main: a call with arguments: null plus the real one)
• test_stream_cut_off_before_its_aggregate_keeps_its_chunks
• test_stream_whose_last_chunk_finishes_keeps_its_chunks (Gemini shape)

tests/unittests/telemetry: 1023 passed, 9 xfailed (functional goldens unchanged)
tests/unittests/models + flows: 2818 passed
pre-commit: passed

Manual End-to-End (E2E) Tests:

gen_ai.output.messages on main vs this PR, LiteLlm with StreamingMode.SSE:

stream	main	this PR
groq openai/gpt-oss-20b, text answer	137	1
groq openai/gpt-oss-20b, tool call	36 incl. tool_call with arguments: null	1: text + tool_call({"city": "Oslo"})
openai provider vs a local SSE server, text in 5 chunks	6	1
same, tool call args in 3 pieces	5	1

Same counts in the completion log with EVENT_ONLY.

Checklist

• I have read the CONTRIBUTING.md document.
• I have performed a self-review of my own code.
• I have commented my code, particularly in hard-to-understand areas.
• I have added tests that prove my fix is effective or that my feature works.
• New and existing unit tests pass locally with my changes.
• I have manually tested my changes end-to-end.
• Any dependent changes have been merged and published in downstream modules.

Additional context

Out of scope: with the v1 schema (ADK_TELEMETRY_SCHEMA_VERSION_OPT_IN=1) or stable semconv, trace_inference_result emits one gen_ai.choice log per response, and those are already emitted when the aggregate arrives. On the local SSE server that is 6 logs for a 5-chunk text answer and 5 for a streamed tool call, the same on main and on this branch. That needs a different fix.

The silent cut-off from #7630 (a dropped stream is finalized as stop) is tracked separately in #7693.

[GBX-Max1220]

Thanks, Ishan—I’ve attached two supplemental native tests against 2c5773f0:

• An interrupted stream propagates the original exception, retains partial messages on the finished span, and records ERROR status.
• EVENT_ONLY records the completed aggregate once.

The full test_instrumentation.py passes locally: 104 tests. I set the test process’s trace sampler to always_on because this environment defaults to 1% sampling.
ADK_7694_Supplemental_Native_Tests_20261023.zip

I’m not claiming a full telemetry-suite pass: the broader run hit a missing SOCKS dependency in the functional tests. The patch only changes tests and was AI-assisted.

[GBX-Max1220]

Hi Ishan — I’ve now ported the two supplemental tests from the attached patch onto your PR head (2c5773f0) in a local branch. The full test_instrumentation.py passes with 104 tests, including both additions. The changes only touch the telemetry test file. Since I can’t push to your branch, could you apply the test-only patch to this PR if you’d like to include the extra coverage?

[ishan-1010]

@surajksharma07 added the v1/stable path to the description as out of scope. it logs 6 gen_ai.choice for a 5-chunk answer and 5 for a streamed tool call, same on main and this branch. marked ready for review.

@GBX-Max1220 you can open a PR against ishan-1010/adk-python, branch fix-litellm-stream-output-once. i'll merge it into this one so the tests show up on this diff.