fix(websearch_interception): ensure spend/cost logging runs when stream=True

[Sameerlite]

Summary

• Websearch interception with stream=True was silently skipping all spend/cost logging because the stream conversion happened inside anthropic_messages (too late for wrapper_async to see)
• Moved stream conversion to async_pre_call_deployment_hook so wrapper_async sees stream=False and takes the non-streaming logging path
• Synced logging_obj.stream after the deployment hook so _success_handler_helper_fn correctly runs cost calculation
• Recovered original stream intent via _websearch_interception_converted_stream flag for the short-circuit path

Test plan

• Added regression test test_deployment_hook_converts_stream_and_logging_obj_syncs
• All 52 existing websearch interception tests pass
• All 6 E2E websearch tests pass

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
litellm	Ready	Preview, Comment	Apr 10, 2026 7:13pm

[codspeed-hq]

Merging this PR will not alter performance

✅ 16 untouched benchmarks

---

_{Comparing litellm_fix-websearch-interception-logging (bec448d) with main (d0e347a)}

[greptile-apps]

Greptile Summary

This PR fixes a spend/cost logging gap where stream=True Anthropic Messages requests with websearch interception would silently skip logging. The root cause was that wrapper_async saw stream=True and took the streaming early-return path before cost logging could run. The fix moves stream conversion from async_pre_request_hook (inside anthropic_messages) up into async_pre_call_deployment_hook so wrapper_async observes stream=False and follows the full non-streaming logging path. A logging_obj.stream sync in utils.py ensures the logging object stays consistent after the hook runs.

Confidence Score: 5/5

Safe to merge; the fix correctly routes websearch-intercepted streaming requests through non-streaming cost logging, and all remaining open questions are P2 style concerns previously flagged.

No new P0/P1 issues found. The logical flow is sound: deployment hook converts stream early enough for wrapper_async to observe it, logging_obj.stream is synced correctly, and _is_streaming_request returns False as intended. The two pre-existing concerns (flag leaking into anthropic_messages_handler kwargs and MagicMock-based test simulation) were already flagged in prior review rounds and are both P2 — neither causes a runtime error since anthropic_messages_handler accepts **kwargs.

litellm/llms/anthropic/experimental_pass_through/messages/handler.py — _websearch_interception_converted_stream still leaks into anthropic_messages_handler via **kwargs (harmless but a prior open concern).

Important Files Changed

Filename	Overview
litellm/integrations/websearch_interception/handler.py	Adds stream→False conversion to async_pre_call_deployment_hook so wrapper_async sees the converted value; retains the same logic in async_pre_request_hook as a fallback for direct callers.
litellm/llms/anthropic/experimental_pass_through/messages/handler.py	Changes original_stream computation to account for the deployment-hook-converted flag; the _websearch_interception_converted_stream key is not popped from kwargs before being passed into anthropic_messages_handler (harmless but leaky — previously flagged).
litellm/utils.py	Adds post-deployment-hook sync of logging_obj.stream to match the (potentially converted) kwargs["stream"]; guards correctly against None stream values.
tests/test_litellm/integrations/websearch_interception/test_websearch_interception_handler.py	New regression test verifies the deployment hook sets stream=False and the converted flag; the wrapper_async sync is simulated via MagicMock rather than exercised through real code (previously flagged).

Sequence Diagram

sequenceDiagram
    participant Caller
    participant wrapper_async as wrapper_async (utils.py)
    participant DeployHook as async_pre_call_deployment_hook
    participant anthropic_messages
    participant PreReqHook as async_pre_request_hook

    Caller->>wrapper_async: stream=True, tools=[web_search]

    Note over wrapper_async: function_setup → logging_obj.stream=True

    wrapper_async->>DeployHook: kwargs {stream=True}
    DeployHook-->>wrapper_async: kwargs {stream=False, _converted=True}

    Note over wrapper_async: Sync logging_obj.stream=False<br/>(NEW: PR fix in utils.py)

    wrapper_async->>anthropic_messages: stream=False, _converted=True, **kwargs
    Note over anthropic_messages: original_stream = False OR True = True<br/>(NEW: flag check at line 190)

    anthropic_messages->>PreReqHook: stream=False (already converted)
    PreReqHook-->>anthropic_messages: no change (stream already False)

    anthropic_messages->>anthropic_messages: short-circuit check with original_stream=True
    anthropic_messages-->>wrapper_async: non-streaming AnthropicMessagesResponse

    Note over wrapper_async: stream=False → takes non-streaming path<br/>✅ spend/cost logging runs

Loading

_{Reviews (2): Last reviewed commit: "Merge pull request #25524 from BerriAI/m..." | Re-trigger Greptile}