Relevant issues
Fixes #27721
Fixes LIT-4039

Summary
Claude Code talking to LiteLLM proxy over /v1/messages with websearch_interception hit two bugs on the same path: search sub-requests 400'd before any search ran, and streaming clients got application/json back after search instead of SSE, which looked like truncated output. This PR fixes both so native web search works end-to-end through the proxy

Problem
Claude Code uses the Anthropic Messages API with a native web_search tool and often forces it via tool_choice: {"type": "tool", "name": "web_search"}. It also sends stream: true

LiteLLM websearch interception renames that tool to litellm_web_search and runs Tavily (or another configured search tool) inside an agentic loop. Two things broke on that path:

1. Search never started (HTTP 400)
   Tools were renamed but tool_choice still pointed at web_search. Anthropic returned Tool 'web_search' not found in provided tools. Claude Code showed "Did 0 searches". A second gap made this worse: tool_choice is a named param on anthropic_messages(), not in **kwargs, so the websearch pre-request hook never saw it to rewrite it

2. Streaming clients got JSON after search ([Bug]: websearch_interception silently truncates streaming response on /v1/messages — follow-up call always uses stream=False #27721)
   The pre-hook correctly converts stream: true → false for the agentic loop and sets websearch_interception_converted_stream so the response can be re-streamed later. After the agentic loop finished, the proxy returned a plain dict (content-type: application/json) instead of SSE. Clients expecting text/event-stream mis-parsed it and answers appeared cut off mid-generation with no error. FakeAnthropicMessagesStreamIterator only ran when the LLM didn't call the search tool; agentic-loop results bypassed the wrap entirely

Fix
Sync forced tool_choice with converted tools
Forward tool_choice explicitly into pre-request hooks. After converting web_search → litellm_web_search, repoint forced tool_choice via _sync_forced_tool_choice() so upstream Anthropic sees a consistent tool name

Re-wrap agentic loop responses as SSE for streaming clients
Add _maybe_websearch_fake_stream_wrap() and call it when the agentic plan returns and on the final handler return. If the client originally requested streaming (websearch_interception_converted_stream on the logging object) and the result is a dict, wrap it in FakeAnthropicMessagesStreamIterator so the proxy emits proper Anthropic SSE

Agentic follow-up calls stay non-streaming internally so LiteLLM can still detect chained tool_use across loop iterations; only the wire format back to the client is converted to SSE

Screenshots / Proof of Fix
Start proxy with websearch_interception and a Tavily search tool configured

Search runs (no 400):

curl -sS -X POST 'http://localhost:4000/v1/messages?beta=true' \
  -H 'Authorization: Bearer sk-1234' \
  -H 'Content-Type: application/json' \
  -H 'anthropic-version: 2023-06-01' \
  -d '{
    "model": "claude-sonnet-4-5",
    "max_tokens": 4096,
    "stream": false,
    "messages": [{"role": "user", "content": "What are the top AI news headlines today?"}],
    "tools": [{"type": "web_search_20250305", "name": "web_search", "max_uses": 1}],
    "tool_choice": {"type": "tool", "name": "web_search"}
  }'

Before: HTTP 400. After: HTTP 200 with web_search_tool_result and real URLs

Streaming wire format (#27721):

curl -N -D /tmp/headers.txt -sS -X POST 'http://localhost:4000/v1/messages?beta=true' \
  -H 'Authorization: Bearer sk-1234' \
  -H 'Content-Type: application/json' \
  -H 'anthropic-version: 2023-06-01' \
  -d '{
    "model": "claude-sonnet-4-5",
    "max_tokens": 4096,
    "stream": true,
    "messages": [{"role": "user", "content": "What are the top AI news headlines today?"}],
    "tools": [{"type": "web_search_20250305", "name": "web_search", "max_uses": 1}],
    "tool_choice": {"type": "tool", "name": "web_search"}
  }'

grep -i content-type /tmp/headers.txt
Before: content-type: application/json. After: content-type: text/event-stream with event: / data: SSE chunks

Files changed
litellm/integrations/websearch_interception/handler.py — _sync_forced_tool_choice()
litellm/llms/anthropic/experimental_pass_through/messages/handler.py — forward tool_choice into hooks
litellm/llms/custom_httpx/llm_http_handler.py — _maybe_websearch_fake_stream_wrap()