fix(anthropic): preserve server_tool_use and web_search_tool_result in multi-turn conversations

[Chesars]

Relevant issues

Fixes #17737

Pre-Submission checklist

• I have Added testing in the tests/litellm/ directory
• My PR passes all unit tests on make test-unit
• My PR's scope is as isolated as possible, it only solves 1 specific problem

Type

🐛 Bug Fix

Changes

---

Bug 1: web_search_tool_result is dropped

When Anthropic returns web search results, LiteLLM was ignoring that field.

Example request

response = litellm.completion(
    model="anthropic/claude-sonnet-4-20250514",
    messages=[{"role": "user", "content": "Search the web for X and use my calculator"}],
    tools=[
        {"type": "web_search_20250305"},  # Anthropic's built-in web search
        {"type": "function", "function": {"name": "calculator", ...}}
    ]
)

Anthropic returns:

{"type": "web_search_tool_result", "tool_use_id": "srvtoolu_01ABC", "content": [...]}

LiteLLM returned to user: Nothing. search results were lost.

Fix: Extract web_search_tool_result and include it in provider_specific_fields.web_search_results.

---

Bug 2: server_tool_use reconstructed as tool_use

When the user sends messages back for a multi-turn conversation, LiteLLM was converting server-side tool calls to regular tool calls.

User sends to LiteLLM:

{"tool_calls": [{"id": "srvtoolu_01ABC", "function": {"name": "web_search", ...}}]}

LiteLLM sent to Anthropic (before fix):

{"type": "tool_use", "id": "srvtoolu_01ABC", "name": "web_search", ...}

❌ Anthropic requires tool_result for every tool_use, but the user can't provide one for server-executed tools.

LiteLLM sends to Anthropic (after fix):

{"type": "server_tool_use", "id": "srvtoolu_01ABC", "name": "web_search", ...}
{"type": "web_search_tool_result", "tool_use_id": "srvtoolu_01ABC", "content": [...]}

✅ Block types.

Fix: Convert srvtoolu_ prefix and reconstruct as server_tool_use + web_search_tool_result.

---

Files changed

• litellm/llms/anthropic/chat/transformation.py - Extract web_search_tool_result
• litellm/litellm_core_utils/prompt_templates/factory.py - Reconstruct server_tool_use
• tests/test_litellm/llms/anthropic/chat/test_anthropic_chat_transformation.py - 3 new tests
• tests/llm_translation/test_prompt_factory.py - 5 new tests

[vercel]

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

Project	Deployment	Preview	Comments	Updated (UTC)
litellm	Ready	Preview	Comment	Dec 10, 2025 1:36am

[krrishdholakia]

@Chesars can you please ensure this covers streaming too?

[KeremTurgutlu]

@krrishdholakia @Chesars streaming is likely failing because of #17254

Here is an example:

❯  sudo ANTHROPIC_API_KEY=$ANTHROPIC_API_KEY uv run test_litellm_websearch_fix_async.py

Reading inline script metadata from `test_litellm_websearch_fix_async.py`
 Updated https://github.com/BerriAI/litellm.git (0d2f8ce93)

=== First API call (async + streaming) ===
Based on the search results, I found the average weights for male elephants:

- **Male African elephants**: 5,000 kg
- **Male Asian elephants**: 3,600 kg
Tool calls: ['web_search', 'add_numbers']

Executing add_numbers(5000, 3600) = 8600

=== Second API call (async + streaming continuation) ===

Give Feedback / Get Help: https://github.com/BerriAI/litellm/issues/new
LiteLLM.Info: If you need to debug this error, use `litellm._turn_on_debug()'.

FAILED: litellm.BadRequestError: AnthropicException - Extra data: line 1 column 60 (char 59)
Received Messages=[{'role': 'user', 'content': 'Search the web for the avg weight in kgs of male African and Asian elephants. Then add the two. Be concise.'}, {'content': 'Based on the search results, I found the average weights for male elephants:\n\n- **Male African elephants**: 5,000 kg\n- **Male Asian elephants**: 3,600 kg', 'role': 'assistant', 'tool_calls': [{'function': {'arguments': '{"query": "average weight male African Asian elephants kg"}{}{}', 'name': 'web_search'}, 'id': 'srvtoolu_01Pupo3WHJ7g8JZF354UKcoE', 'type': 'function'}, {'function': {'arguments': '{"a": 5000, "b": 3600}', 'name': 'add_numbers'}, 'id': 'toolu_01RRf2pCDB7TBUfJGv2oPLGk', 'type': 'function'}]}, {'role': 'tool', 'tool_call_id': 'toolu_01RRf2pCDB7TBUfJGv2oPLGk', 'content': '8600'}]

you can see 'arguments': '{"query": "average weight male African Asian elephants kg"}{}{}', which has extra trailing {} which causes the JSON decoding issue.

[KeremTurgutlu]

I can confirm that this is the only issue. For example postprocessing tool_calls before the next call like this works:

# Convert streamed chunks to complete message
full_response = litellm.stream_chunk_builder(chunks)
content = full_response.choices[0].message.content
tool_calls = full_response.choices[0].message.tool_calls or []
for tc in tool_calls:
    if tc.function.arguments:
        tc.function.arguments = tc.function.arguments.replace('{}', '')

❯  sudo ANTHROPIC_API_KEY=$ANTHROPIC_API_KEY uv run test_litellm_websearch_fix_async.py

Password:
Reading inline script metadata from `test_litellm_websearch_fix_async.py`
 Updated https://github.com/BerriAI/litellm.git (0d2f8ce93)

=== First API call (async + streaming) ===
Based on the search results, I found information about male elephant weights:

- Male African savanna elephants average 5,000 kg (11,000 pounds)
- Male Asian elephants weigh on average about 3,600 kg (7,900 pounds)

Now let me add these weights:
Tool calls: ['web_search', 'add_numbers']

Executing add_numbers(5000, 3600) = 8600

=== Second API call (async + streaming continuation) ===
Based on the search results:

- Male African elephants average 5,000 kg
- Male Asian elephants weigh on average about 3,600 kg

**Total: 5,000 + 3,600 = 8,600 kg**
SUCCESS!

==================================================
Test result: PASSED