Empty text part ahead of tool call causes run to end while streaming

[gonzoramos]

Initial Checks

• I confirm that I'm using the latest version of Pydantic AI
• I confirm that I searched for my issue in https://github.com/pydantic/pydantic-ai/issues before opening this issue

Description

When using a local ollama model (e.g., qwen2.5 or qwen3) on an agent that uses tools, if the tool gets used, the agent stops right after the tool gets called and returns nothing.

As reported in https://pydanticlogfire.slack.com/archives/C083V7PMHHA/p1752874628320719?thread_ts=1752791433.109389&cid=C083V7PMHHA

Example Code

import asyncio
import os
import random
from typing import Any, List, Optional

import logfire
from pydantic_ai import Agent, Tool
from pydantic_ai.models.openai import OpenAIModel
from pydantic_ai.models.groq import GroqModel
from pydantic_ai.providers.openai import OpenAIProvider


from pydantic import BaseModel
from dotenv import load_dotenv

# Load environment variables from a .env file if it exists
env_ok = load_dotenv(os.path.join('..', ".env"))

if not env_ok:
    print("Warning: .env file not found. Environment variables may not be set correctly.")

logfire.configure()
logfire.instrument_pydantic_ai()


def roll_dice() -> str:
    """Roll a six-sided die and return the result."""
    return str(random.randint(1, 6))

# --- Agent and Model Configuration ---


ollama_model = OpenAIModel(
    model_name='qwen2.5', provider=OpenAIProvider(base_url='http://localhost:11434/v1'), settings={"temperature": 0.2}
)

groq_model = GroqModel(
    model_name='qwen/qwen3-32b', settings={"temperature": 0.2},
)

default_system_prompt = """
You are a helpful assistant. 
Do not ask any follow-up questions.
"""

model_settings = {
    "temperature": 0.2,
}

tools = [roll_dice]

agent = Agent(model=ollama_model, tools=tools,
              system_prompt=default_system_prompt, output_type=str)


async def main():
    prompt = 'roll teh dice'

    async with agent.run_stream(prompt) as result:
        async for message in result.stream():
            print(message)

    print(result.usage())

# ---

if __name__ == '__main__':
    asyncio.run(main())

Python, Pydantic AI & LLM client version

python 3.11.13, pydantic 0.4.2, LLM: ollama 0.9.6, qwen2.5 and qwen3 latest

[DouweM]

From the info shared in Slack, it looks like qwen2.5 is responding with a TextPart(content='') ahead of the ToolCallPart(tool_name='roll_dice', args='{}', tool_call_id='call_o22pezfo'), so it's stopping because the model returning text (even if empty text) causes Pydantic AI to think it's done.

I think we should just ignore empty text when detecting the final response. That would need to be done here:

pydantic-ai/pydantic_ai_slim/pydantic_ai/agent.py

Lines 1136 to 1141 in 7d50564

	if isinstance(maybe_part_event, _messages.PartStartEvent):
	new_part = maybe_part_event.part
	if isinstance(new_part, _messages.TextPart) and isinstance(
	output_schema, _output.TextOutputSchema
	):
	return FinalResult(s, None, None)

and here:

pydantic-ai/pydantic_ai_slim/pydantic_ai/result.py

Lines 150 to 155 in 7d50564

	if isinstance(e, _messages.PartStartEvent):
	new_part = e.part
	if isinstance(new_part, _messages.TextPart) and isinstance(
	output_schema, TextOutputSchema
	): # pragma: no branch
	return _messages.FinalResultEvent(tool_name=None, tool_call_id=None)

Ideally, we would not stream a PartStartEvent at all until we know that the text part will actually have text. That'd need to happen here:

pydantic-ai/pydantic_ai_slim/pydantic_ai/_parts_manager.py

Lines 111 to 124 in 7d50564

	if existing_text_part_and_index is None:
	# There is no existing text part that should be updated, so create a new one
	new_part_index = len(self._parts)
	part = TextPart(content=content)
	if vendor_part_id is not None:
	self._vendor_id_to_part_index[vendor_part_id] = new_part_index
	self._parts.append(part)
	return PartStartEvent(index=new_part_index, part=part)
	else:
	# Update the existing TextPart with the new content delta
	existing_text_part, part_index = existing_text_part_and_index
	part_delta = TextPartDelta(content_delta=content)
	self._parts[part_index] = part_delta.apply(existing_text_part)
	return PartDeltaEvent(index=part_index, delta=part_delta)

[gonzoramos]

Many thanks @DouweM for looking into this. Looking forward to a resolution when is ready.

[newsbubbles]

I can verify this has happened to me with kimi k2 and other models outside of the anthropic family... both using iter and run

Im using openrouter for model service

[iBuitron]

Same Here:

from pydantic_ai import Agent, RunContext

from pydantic_ai.providers.openai import OpenAIProvider
from pydantic_ai.models.openai import OpenAIModel

model_name = "qwen/qwen3-4b"
base_url = "http://192.168.1.34:1234/v1"


model = OpenAIModel(
    model_name=model_name,
    provider=OpenAIProvider(base_url=base_url),
)
agent = Agent(
    model=model,
    deps_type=str,
    system_prompt=(
        # "/no_think"
        "You're a dice game, you should roll the die and see if the number "
        "you get back matches the user's guess. If so, tell them they're a winner. "
        "Use the player's name in the response."
    ),
)


@agent.tool_plain
def roll_dice() -> str:
    """Roll a six-sided die and return the result."""
    return str(random.randint(1, 6))


@agent.tool
def get_player_name(ctx: RunContext[str]) -> str:
    """Get the player's name."""
    return ctx.deps


dice_result = agent.run_sync("My guess is 4", deps="Anne")
print(dice_result.output)

output:

<tool_call>
{"name": "get_player_name", "arguments": {}}
</tool_call>

LMSttudio LOG:

2025-07-25 12:30:47  [INFO]
 [LM STUDIO SERVER] Running chat completion on conversation with 2 messages.
2025-07-25 12:30:48  [INFO]
 [LM STUDIO SERVER] Accumulating tokens ... (stream = false)
2025-07-25 12:30:50  [INFO]
 [qwen/qwen3-4b] Generated prediction:  {
  "id": "chatcmpl-tynez5kpk5bqkpmz4jxrur",
  "object": "chat.completion",
  "created": 1753464647,
  "model": "qwen/qwen3-4b",
  "choices": [
    {
      "index": 0,
      "logprobs": null,
      "finish_reason": "stop",
      "message": {
        "role": "assistant",
        "content": "<think>\nOkay, the user guessed 4. I need to check if their guess matches the die roll. First, I should get their name using get_player_name. Then, roll the dice. Once I have both, compare the roll result with 4. If it's equal, congrats them with their name. Otherwise, say they lost. Let me start by calling get_player_name.\n</think>\n\n<tool_call>\n{\"name\": \"get_player_name\", \"arguments\": {}}\n</tool_call>"
      }
    }
  ],
  "usage": {
    "prompt_tokens": 236,
    "completion_tokens": 98,
    "total_tokens": 334
  },
  "stats": {},
  "system_fingerprint": "qwen/qwen3-4b"
}

[DouweM]

@iBuitron That's a different issue: Qwen3 returns its tool calls inside a text message as <tool_call>...</tool_call>, which Pydantic AI doesn't currently support. We saw something similar in #2289 and #2033. I'll update that last issue to be the canonical place for parsing tool calls out of text.

[iBuitron]

@DouweM I switched the server from LMStudio to Ollama and it works perfectly, but only if the Think option is activated.

[majdalsado]

I'm finding a somewhat similar issue while streaming Sonnet 4 with tool calling.

Around ~1-2% of calls where we have a tool call invokation will have an empty content field (see below). This seems to cause a 400 response when we send it back to the model with the tool call response.

Problematic response from LLM with an empty content field

{
    "index": 0,
    "message": {
        "role": "assistant",
        "content": "",
        "tool_calls": [
            {
                "id": "toolu_vrtx_01GCbdwKdU937qfwV751yWGg",
                "type": "function",
                "function": {
                    "name": "read_document",
                    "arguments": ""
                }
            }
        ]
    },
    "gen_ai.system": "openrouter",
    "event.name": "gen_ai.choice"
}

Subsequent error during the next conversation turn

ModelHTTPError: status_code: 400, model_name: anthropic/claude-sonnet-4, body: {'message': 'Provider returned error', 'code': 400, 'metadata': {'raw': '{"message":"messages: text content blocks must be non-empty"}', 'provider_name': 'Amazon Bedrock'}}