LiteLLM + DeepSeek-V3 multi-tool calling fails: tool call parsing error

When using LiteLLM to proxy DeepSeek-V3 model with multiple tool calls, the first tool call consistently fails to parse. The observed behavior is:

If the first tool call succeeds, all subsequent tool calls execute normally

Environment

Model: DeepSeek-V3 (via LiteLLM proxy)
Framework: Google ADK (version: [1.22.1])
LiteLLM version: [1.82.1]

Error Output

raw event: model_version='deepseek-v3' content=Content(
  parts=[
    Part(
      text=""" <｜tool▁calls▁begin｜><｜tool▁call▁begin｜>function<｜tool▁sep｜>analysis_input
```json
{"work_dir_name":"bf01a09edc0f4c4ca376bc03d153445b"}
```<｜tool▁call▁end｜><｜tool▁calls▁end｜>"""
    ),
  ],
  role='model'
) grounding_metadata=None partial=False turn_complete=None finish_reason=<FinishReason.STOP: 'STOP'> error_code=None error_message=None interrupted=None custom_metadata=None usage_metadata=GenerateContentResponseUsageMetadata(
  cached_content_token_count=0,
  candidates_token_count=39,
  prompt_token_count=1003,
  total_token_count=1042
) live_session_resumption_update=None input_transcription=None output_transcription=None avg_logprobs=None logprobs_result=None cache_metadata=None citation_metadata=None interaction_id=None invocation_id='e-bb165836-aee3-404e-9ef0-ef499f6f997e' author='brain_agent' actions=EventActions(skip_summarization=None, state_delta={}, artifact_delta={}, transfer_to_agent=None, escalate=None, requested_auth_configs={}, requested_tool_confirmations={}, compaction=None, end_of_agent=None, agent_state=None, rewind_before_invocation_id=None) long_running_tool_ids=None branch=None id='4824cba0-6653-4d41-b0e6-6343bc856434' timestamp=1774580915.381178

This error doesn't seem to be triggered every time; sometimes it displays normally：

Part(
      function_call=FunctionCall(
        args={
          'work_dir_name': '84a652c678564995be15e5759404d7b3'
        },
        id='646ee4f52df44e58a70d99ab4313b0f8',
        name='analysis_input'
      )

run script:

async for event in runner.run_async(
            user_id=user_id,
            session_id=session.id,
            new_message=content,
            run_config=RunConfig(save_input_blobs_as_artifacts=False),
        ):
            print("\n================ EVENT ================") 
            print("raw event:", event)
            if event.is_final_response() and event.content and event.content.parts:
                for part in event.content.parts:
                    if getattr(part, "text", None):
                        parts.append(part.text)
                final_usage = getattr(event, "usage_metadata", None)
                print("prompt_tokens:", final_usage.prompt_token_count) 
                print("completion_tokens:", final_usage.candidates_token_count) 
                print("total_tokens:", final_usage.total_token_count)

fix(litellm): parse DeepSeek-V3 proprietary inline tool-call tokens

/adk-issue-analyze

GitHub Issue #5024 Analysis: LiteLLM + DeepSeek-V3 multi-tool calling fails: tool call parsing error

Executive Summary

What is broken? When LiteLLM proxies DeepSeek-V3, the model's tool calls are sometimes emitted as raw text wrapped in DeepSeek's proprietary special tokens (e.g., <|tool_calls_begin|>...) instead of being structured as standard tool_calls by LiteLLM. When this occurs, ADK-python's fallback JSON parser, _parse_tool_calls_from_text, fails to parse them into standard tool calls because the JSON object contains only the function arguments—with the function name situated outside the JSON block. As a result, _build_tool_call_from_json_dict rejects the payload because the "name" key is missing, causing the tool call to be silently dropped.
Is there a linked PR that fixes this issue? Yes, PR fix(litellm): parse DeepSeek-V3 proprietary inline tool-call tokens #5654 - OPEN
Recommendation: Should Fix (High Priority). DeepSeek-V3 is a widely used and high-performance model. Silently dropping tool calls severely impacts standard developer workflows and reliability. Merging or adapting PR fix(litellm): parse DeepSeek-V3 proprietary inline tool-call tokens #5654 will resolve this cleanly.

Detailed Analysis

1. Root Cause Analysis ("What is broken?")

DeepSeek-V3 uses a custom token block to structure custom tool call emissions:

<|tool_calls_begin|><|tool_call_begin|>function<|tool_sep|>analysis_input
```json
{"work_dir_name":"bf01a09edc0f4c4ca376bc03d153445b"}
```<|tool_call_end|><|tool_calls_end|>

When LiteLLM returns this payload in the raw content field without translating it to standard formatted tool_calls, ADK-python delegates parsing to _split_message_content_and_tool_calls, which invokes _parse_tool_calls_from_text:

brace_index matches the starting { of the JSON block.
_JSON_DECODER.raw_decode successfully extracts the JSON dictionary {"work_dir_name": "bf01a09edc0f4c4ca376bc03d153445b"}.
Then, _build_tool_call_from_json_dict expects the JSON dictionary to have "name" and "arguments" keys. Because the JSON block parsed consists purely of the parameters dictionary (with the function name analysis_input parsed outside the JSON brace inside the <|tool_sep|> token), candidate.get("name") evaluates to None.
Consequently, _build_tool_call_from_json_dict returns None and the tool call is silently ignored, bubbling up to the user as raw text content instead of executable actions.

2. Existing Pull Requests ("Is there a linked PR that fixes this issue?")

Linked PR: Pull Request fix(litellm): parse DeepSeek-V3 proprietary inline tool-call tokens #5654 - fix(litellm): parse DeepSeek-V3 proprietary inline tool-call tokens (OPEN)
PR URL: fix(litellm): parse DeepSeek-V3 proprietary inline tool-call tokens #5654
Analysis:
- The PR author (fuchun1010) implemented _parse_deepseek_tool_calls_from_text which specifically scans for <|tool_calls_begin|> and parses the custom format (handling markdown fences, single/multiple tool calls, and bare JSON).
- It integrates it as a pre-pass step inside _parse_tool_calls_from_text.
- The PR adds 8 robust unit-tests covering different permutations (e.g., code-fences, mixed formats, text-only) inside tests/unittests/models/test_litellm.py.
- It is currently open, green on CI, and awaiting a review from the maintainers.

3. Recommendation

Recommendation: Should Fix (High Priority)
Rationale:
- Impact: DeepSeek-V3 model integration with tool execution is a core feature. Intermittent parse failures and silent drops severely break user trust and standard developer workflows.
- Complexity: Low to Medium. PR fix(litellm): parse DeepSeek-V3 proprietary inline tool-call tokens #5654 contains a solid implementation with extensive regression coverage (testing 8 scenarios, passing all 264 existing test_litellm.py tests).
- Plan: Proceed with reviewing and merging PR fix(litellm): parse DeepSeek-V3 proprietary inline tool-call tokens #5654, or implement an equivalent regex-based extraction pre-pass mechanism matching its logic.

Summary of Work Done

Parsed and analyzed GitHub Issue LiteLLM + DeepSeek-V3 multi-tool calling fails: tool call parsing error #5024.
Isolated the root cause to model fallback parsing in lite_llm.py specifically under _parse_tool_calls_from_text and _build_tool_call_from_json_dict.
Traced the interaction flow of inline custom DeepSeek-V3 token blocks.

fix(litellm): parse DeepSeek-V3 proprietary inline tool-call tokens

Navigation Menu

Repository navigation

LiteLLM + DeepSeek-V3 multi-tool calling fails: tool call parsing error #5024

Environment

Error Output

GitHub Issue #5024 Analysis: LiteLLM + DeepSeek-V3 multi-tool calling fails: tool call parsing error

Executive Summary

1. Root Cause Analysis ("What is broken?")

2. Existing Pull Requests ("Is there a linked PR that fixes this issue?")

3. Recommendation

Summary of Work Done

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

LiteLLM + DeepSeek-V3 multi-tool calling fails: tool call parsing error #5024

Description

Environment

Error Output

Activity

yfzhang1214 commented on Mar 27, 2026

surajksharma07 commented on Jun 5, 2026

adk-bot commented on Jun 5, 2026

GitHub Issue #5024 Analysis: LiteLLM + DeepSeek-V3 multi-tool calling fails: tool call parsing error

Executive Summary

1. Root Cause Analysis ("What is broken?")

2. Existing Pull Requests ("Is there a linked PR that fixes this issue?")

3. Recommendation

Summary of Work Done

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions