[Pr1meGG]

Materialize generator and one-shot iterator inputs for messages, system, and tools in kwargs before telemetry extraction in Anthropic create and stream methods when content capture is enabled, leaving them unmaterialized when content capture is disabled.

This prevents generator exhaustion in Anthropic instrumentation, mirroring the fix in #864 for OpenAI (Issue #825).

Known gaps or concerns

None.

[opentelemetry-pr-dashboard]

Pull request dashboard status

Closed · refreshed 2026-10-19 12:10 UTC

Status above doesn't look right?

• Anything look wrong? Report it with what you expected; it helps us improve the dashboard.

[Copilot]

🟡 Changes recommended

It does not fix linked OpenAI issue #825 and lacks async stream regression coverage.

2 open findings

• OpenAI instrumentation still leaves issue #825 unresolved · New
• Add async stream regression tests for messages and system iterators · New

What changed in this PR

Fixes one-shot iterator handling in Anthropic instrumentation, though linked issue #825 concerns OpenAI.

Changes:

• Materializes messages, system, and tools iterators.
• Adds sync and async regression tests.
• Adds a changelog fragment.

File	Description
patch.py	Materializes request iterators before extraction and SDK calls.
test_sync_messages.py	Tests sync create and stream behavior.
test_async_messages.py	Tests async create behavior.
.changelog/​825.fixed	Documents the Anthropic fix.

🧠 Review effort: Balanced

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[Pr1meGG]

Responding to review findings

Both findings are addressed.

Finding 1: #825 is OpenAI-only — this PR does not fix it

Addressed. The PR was never intended to fix the OpenAI issue; it fixes the same class of bug for Anthropic. Evidence:

• Commit d8d249c renamed the changelog fragment from 825.fixed to 865.fixed.
• The current .changelog/865.fixed reads: Prevent one-shot iterators for messages, system, and tools from being consumed before reaching the SDK during content capture. No reference to [genai-openai] With content capture enabled, generator/iterator messages and tools are consumed and the request is sent empty #825.
• No source file in this PR references #825.

The PR title includes (#825) as the originating issue reference. I can remove it if preferred — happy to.

Finding 2: Missing async stream messages/system regression tests

Addressed. The tests already exist in this PR (added after the review).

Test	Line	Asserts request body	Asserts telemetry
test_async_messages_stream_messages_generator_is_recorded	1172	seen["body"]["messages"] == [user_msg]	input_messages span
test_async_messages_stream_system_generator_is_recorded	1217	seen["body"]["system"] == [system_block]	system_instruction span
test_async_messages_stream_tools_generator_is_recorded	1127	seen["body"]["tools"] == [tool_def]	tool_definitions span

All use MockTransport to capture the serialized request, then assert both the JSON body and the span attributes. All 8 async generator tests + all 5 sync generator tests pass:

$ pytest -k generator
6 passed, 170 deselected

These tests were added as part of PR #865 itself but appear to have been missed in review.

[MikeGoldsmith]

Thanks for the work on this @Pr1meGG, and for the detailed replies on the review threads.

I think this one has been overtaken by #873, which landed the same fix for opentelemetry-instrumentation-genai-anthropic using the shared IteratorCapture helper in opentelemetry-util-genai. Looking at the current diff against main, the patch.py change isn't in it any more. The merge in f87e02d resolved that file to main's version, so what's left is the changelog fragment and the tests.

On those two, .changelog/865.fixed describes the same user-visible change as 873.fixed, so we'd end up listing it twice in the release notes. And tests/test_iterator_transparency.py came in with #873 and parametrises messages, system and tools across sync / async, streaming and non-streaming, and capture content on and off, so I think it already covers what the tests here assert.

Given that, I think it makes sense to close this. If you spot a case that the matrix in test_iterator_transparency.py misses, a small test-only PR for it would be welcome, and easier to review on its own.

Thanks again 🙂