Check for existing issues

• I have searched the existing issues and checked that my issue is not a duplicate.

What happened?

When using Claude Code with non-Anthropic models (Gemini) through LiteLLM's /v1/messages endpoint, all tool_use blocks arrive with empty input: {}. The model returns correct tool call arguments in OpenAI format, but the streaming adapter discards them during translation back to Anthropic format. This causes Claude Code to loop on InputValidationError: The required parameter 'command' is missing for every tool call (Bash, Write, TodoWrite, etc.).

Works on: v1.81.14-stable
Broken on: v1.82.0-stable, v1.82.3-stable.patch.2, v1.83.4-nightly

Root Cause

In streaming_iterator.py, when _should_start_new_content_block() returns True (e.g. text → tool_use transition), the code queues content_block_stop + content_block_start and returns immediately without queuing processed_chunk:

# streaming_iterator.py — async __anext__, ~line 306-330
if should_start_new_block and not self.sent_content_block_finish:
    self.chunk_queue.append({"type": "content_block_stop", ...})
    self.chunk_queue.append({"type": "content_block_start", ..., "content_block": self.current_content_block_start})
    self.sent_content_block_finish = False
    return self.chunk_queue.popleft()  # ← processed_chunk is NEVER QUEUED

The comment says: "The trigger chunk itself is not emitted as a delta since the content_block_start already carries the relevant information."

This is correct for text blocks (content_block_start has the text). But it's wrong for tool_use blocks: _translate_streaming_openai_chunk_to_anthropic_content_block always creates tool_use with input={} (line 1217 in transformation.py). The actual arguments are ONLY in processed_chunk as an input_json_delta — and that chunk is silently discarded.

In v1.81.14, the _should_start_new_content_block mechanism didn't exist — all chunks were unconditionally translated and queued, so tool arguments were never lost.

Evidence

Gemini raw response (OpenAI format) — arguments present:

{
  "function": {
    "name": "Bash",
    "arguments": "{\"command\": \"ls -R\", \"description\": \"List files\"}"
  }
}

What Claude Code receives (Anthropic format) — arguments missing:

{
  "type": "tool_use",
  "id": "call_6f570bdb...",
  "name": "Bash",
  "input": {}
}

Claude Code error:

InputValidationError: Bash failed due to the following issue:
The required parameter `command` is missing

Fix

Add 2 lines in both sync and async paths — after queuing content_block_start, also queue processed_chunk if it carries delta content:

if processed_chunk.get("type") == "content_block_delta":
    self.chunk_queue.append(processed_chunk)

This is safe for all block types: text deltas, tool_use input_json_delta, and thinking_delta all benefit from being queued alongside their content_block_start. The real Anthropic API sends both content_block_start and the first delta in sequence — this fix aligns the adapter with that behavior.

Related

• PR #13638 fixed the same class of bug (input_json_delta vs text_delta) in an earlier version
• Issue #20711 describes the same argument-drop pattern in the Responses API streaming path

Steps to Reproduce

1. Configure LiteLLM proxy with a Gemini model:

   model_list:
     - model_name: gemini-3-flash-preview
       litellm_params:
         model: gemini/gemini-3-flash-preview
         api_key: os.environ/GEMINI_API_KEY
   litellm_settings:
     drop_params: true
     modify_params: true

2. Point Claude Code at the proxy:

   export ANTHROPIC_BASE_URL="http://localhost:4000"
   export ANTHROPIC_AUTH_TOKEN="$LITELLM_MASTER_KEY"

3. Run any Claude Code task that requires tool use:

   claude --model gemini-3-flash-preview -p "list the files in the current directory"

4. Observe: every tool call (Bash, Write, TodoWrite, etc.) fails with InputValidationError: The required parameter 'command' is missing because input is {}.

5. Downgrade to v1.81.14-stable — same test works correctly.

Relevant log output

# Claude Code logs showing empty tool input on v1.82.3:

{"type":"tool_use","id":"call_81b4866e7c6549ddb26cf39f2989__thought__Eq8JCqwJAb4...","name":"Bash","input":{}}

# Followed by validation error:
InputValidationError: Bash failed due to the following issue:
The required parameter `command` is missing

# LiteLLM raw response from Gemini shows arguments ARE present:
tool_calls: [{
  function: {
    name: "Bash",
    arguments: "{\"command\": \"ls -R\", \"description\": \"List files in the current directory and subdirectories\"}"
  }
}]

# But content_block_start sent to client has input={}:
{"type":"content_block_start","index":1,"content_block":{"type":"tool_use","id":"call_...","name":"Bash","input":{}}}
# No content_block_delta with input_json_delta follows — processed_chunk was discarded

What part of LiteLLM is this about?

Proxy

What LiteLLM version are you on ?

v1.82.3-stable.patch.2 (also reproduces on v1.83.4-nightly; works on v1.81.14-stable)

Twitter / LinkedIn details

No response