fix(websearch_interception): Convert agentic loop response to streaming format for Claude Code

[shin-bot-litellm]

Summary

Fixes #20187

When using websearch_interception with Bedrock and Claude Code, the output tokens were showing as 0 because the agentic loop response wasn't being converted back to streaming format.

Problem

1. When Claude Code sends a streaming request with websearch tools
2. The websearch_interception handler converts stream=True to stream=False to intercept the response
3. When the agentic loop runs (executing the search and making a follow-up request), the response was returned as a non-streaming dict
4. Claude Code expects a streaming response, so when it parsed the non-streaming dict, the output tokens appeared as 0

Solution

After the agentic loop completes, check if the original request was streaming (via the websearch_interception_converted_stream flag). If so, convert the agentic loop's non-streaming response to streaming format using FakeAnthropicMessagesStreamIterator.

This ensures:

• Output tokens are correctly included in the message_delta event (as per Anthropic's streaming spec)
• stop_reason is properly preserved
• The response format matches what Claude Code expects

Testing

• Added unit tests to verify FakeAnthropicMessagesStreamIterator correctly includes output tokens and preserves stop_reason
• All existing tests pass

Changes

• litellm/llms/custom_httpx/llm_http_handler.py: Added streaming conversion for agentic loop response
• tests/test_litellm/integrations/websearch_interception/test_websearch_interception_handler.py: Added regression tests

[vercel]

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

Project	Deployment	Actions	Updated (UTC)
litellm	Ready	Preview, Comment	Feb 7, 2026 4:19am

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[greptile-apps]

Greptile Overview

Greptile Summary

This PR fixes a bug where Claude Code showed 0 output tokens when using websearch_interception with Bedrock. The issue occurred because the websearch handler converts streaming requests to non-streaming for interception, and when the agentic loop completed, it returned a dict response instead of converting it back to streaming format.

Changes:

• Added logic in llm_http_handler.py:4415-4440 to check if the original request was streaming (via the websearch_interception_converted_stream flag) and convert the agentic loop's dict response to a fake stream using FakeAnthropicMessagesStreamIterator
• Added two regression tests to verify the fake stream iterator correctly includes output_tokens in the message_delta event and preserves stop_reason
• The fix mirrors existing logic at line 4455 that handles the case when no agentic loop runs

Technical Details:

• When websearch tools are detected, stream=True is converted to stream=False and a flag is set in logging_obj.model_call_details
• After the agentic loop completes its follow-up request, the code now checks this flag and converts the dict response to streaming format
• The FakeAnthropicMessagesStreamIterator properly formats the response with all required Anthropic streaming events including correct token counts

Confidence Score: 5/5

• This PR is safe to merge with minimal risk
• The changes are well-isolated to the websearch interception flow, follow existing patterns in the codebase (duplicating logic from line 4455), include comprehensive regression tests, and fix a specific reported issue without introducing new dependencies or altering the core request path
• No files require special attention

Important Files Changed

Filename	Overview
litellm/llms/custom_httpx/llm_http_handler.py	Adds streaming conversion for agentic loop responses when websearch interception converts stream=True to stream=False. Uses FakeAnthropicMessagesStreamIterator to properly format the response with output tokens and stop_reason.
tests/test_litellm/integrations/websearch_interception/test_websearch_interception_handler.py	Adds regression tests to verify FakeAnthropicMessagesStreamIterator correctly includes output_tokens in message_delta event and preserves stop_reason, addressing the reported issue.

Sequence Diagram

sequenceDiagram
    participant Client as Claude Code Client
    participant Handler as WebSearchInterception
    participant LLM as LLM Provider
    participant AgenticLoop as Agentic Loop
    participant Converter as FakeStreamIterator

    Client->>Handler: Request with stream=True and websearch tools
    Handler->>Handler: Convert stream=True to stream=False
    Handler->>Handler: Set websearch_interception_converted_stream flag
    Handler->>LLM: Make request with stream=False
    LLM->>Handler: Non-streaming response with tool_use
    Handler->>AgenticLoop: Execute agentic loop with websearch
    AgenticLoop->>AgenticLoop: Execute search queries
    AgenticLoop->>LLM: Follow-up request with search results
    LLM->>AgenticLoop: Non-streaming response dict
    AgenticLoop->>Handler: Return agentic response dict
    Handler->>Handler: Check websearch_interception_converted_stream flag
    Handler->>Converter: Convert dict to fake stream
    Converter->>Converter: Create streaming events
    Converter->>Converter: Include output_tokens in message_delta
    Converter->>Client: Return streaming response with correct tokens

Loading

[greptile-apps]

_{2 files reviewed, no comments}

_{Edit Code Review Agent Settings | Greptile}