fix(weave): trace Claude Agent SDK async-iterable queries per turn

[rgao-coreweave]

Problem

The Claude Agent SDK accepts either a string or an AsyncIterable as the prompt to query(). An async-iterable prompt can submit multiple inputs and produce a separate ResultMessage for each one.

The OTel integration previously treated the entire output iterator as one turn. On current master, a replay containing two submitted prompts and two results produces only one invoke_agent span, collapsing both interactions together.

Approach

Keep the ordinary one-result APIs simple and isolate multi-result coordination to the only path that needs it:

SDK API	Tracing path	Turn boundary
query(prompt=str)	Linear single-turn tracer	The first ResultMessage
query(prompt=AsyncIterable)	Async-iterable FIFO tracer	Each ResultMessage
ClaudeSDKClient.receive_response()	Linear single-turn tracer	The method stops after its first ResultMessage

Both paths share message processing and turn finalization. The async-iterable path additionally:

• Captures user inputs as the SDK consumes them without pre-consuming or cloning the iterable.
• Buffers messages marked shouldQuery: false into the next query-triggering input.
• Pairs submitted inputs with results in FIFO order.
• Reuses the SDK session ID as the conversation ID while keeping each turn in its own trace.
• Records a pending input as an errored turn when transport or parsing fails before its result.
• Leaves a completed turn successful if teardown fails after its terminal result.

String prompts and the public ClaudeSDKClient.receive_response() contract remain on the straightforward single-turn implementation.

Example

Two-turn async-iterable query in Weave records remembered. and CERULEAN-42 as two turns in one conversation.

Testing

Case	Expected tracing contract
String prompt, one result	One successful turn
Async-iterable prompt, one input/result	Same span payload as the string-prompt path
Async-iterable prompt, two inputs/results	Two FIFO-paired turns in one conversation
Async-iterable input with shouldQuery: false	Buffered into the next query-triggering turn
String or async-iterable failure before the first response	One errored turn containing the submitted input
Async-iterable failure between results	Completed turn stays successful; pending turn is errored
String or async-iterable failure after the final result	Completed turn stays successful; no phantom turn
ClaudeSDKClient with string or async-iterable input	One turn per receive_response()

• Claude Agent SDK integration suite: 34 passed on Python 3.10.
• Claude Agent SDK integration suite: 34 passed on Python 3.13.
• Ruff 0.15.5 check and format verification passed.
• GitHub CI: 105 passed checks and 3 intentionally skipped checks.

Breaking changes

None.

[rgao-coreweave]

• fix(weave): exclude media payloads from agent chat text #7660
• feat(weave): trace Claude Agent SDK image prompts #7667
• feat(weave): trace Claude Agent SDK subagents #7666
• fix(weave): trace Claude Agent SDK async-iterable queries per turn #7665 👈 (View in Graphite)
• refactor(weave): migrate Claude Agent SDK tracing to GenAI SDK #7642
• master

This stack of pull requests is managed by Graphite. Learn more about stacking.

[w-b-hivemind]

HiveMind Sessions

1 session · 11m · $6.15

Session	Agent	Duration	Tokens	Cost	Lines
Update PR Descriptions and Sync Commits e4ab1870-5f94-4b10-8abf-1d259fd41c55	claude	11m	32.6K	$6.15	+169 -88
Total		11m	32.6K	$6.15	+169 -88

View all sessions in HiveMind →

Run claude --resume e4ab1870-5f94-4b10-8abf-1d259fd41c55 to pickup where you left off.

[codecov]

Codecov Report

❌ Patch coverage is 85.71429% with 15 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
.../integrations/claude_agent_sdk/otel_integration.py	85.71%	12 Missing and 3 partials ⚠️

📢 Thoughts on this report? Let us know!