Closed
Description
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
I ran into TypeError: unsupported operand type(s) for +: 'int' and 'list' when using gemini via GoogleModel with Gemini API.
Trace:
...
File "...\.venv\Lib\site-packages\pydantic_ai\_agent_graph.py", line 337, in _make_request
return self._finish_handling(ctx, model_response)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "...\.venv\Lib\site-packages\pydantic_ai\_agent_graph.py", line 361, in _finish_handling
ctx.state.usage.incr(response.usage)
File "...\.venv\Lib\site-packages\pydantic_ai\usage.py", line 48, in incr
self.details[key] = self.details.get(key, 0) + value
~~~~~~~~~~~~~~~~~~~~~~~~~^~~~~~~
TypeError: unsupported operand type(s) for +: 'int' and 'list'
After some digging, I found that this error is prone to happen when there are repetitive requests, and the incr_usage.details value for each request is:
- First request (success):
{'thoughts_token_count': 70} - Second request (still success):
{'thoughts_token_count': 68} - Thrid request (error):
{'cache_tokens_details': [{'modality': <MediaModality.TEXT: 'TEXT'>, 'token_count': 3950}], 'cached_content_token_count': 3950, 'thoughts_token_count': 50}
It's clear that cache_tokens_details field is causing the problem, and this might be a Google problem, but some additional edge case handling is nice though.
Example Code
from __future__ import annotations as _annotations
import os
import asyncio
from pydantic import BaseModel
from pydantic_ai import Agent
from pydantic_ai.models.google import GoogleModel
from pydantic_ai.providers.google import GoogleProvider
from dotenv import load_dotenv
import logfire
load_dotenv()
logfire.configure(send_to_logfire=False)
provider = GoogleProvider(api_key=os.getenv("GOOGLE_GEMINI_API_KEY"))
model = GoogleModel(model_name="gemini-2.5-flash-preview-05-20", provider=provider)
class Test(BaseModel):
output: str
email_writer_agent = Agent(
model=model,
output_type=Test,
instructions=f"""
Here is a pretty long prompt to illustrate the usage of the Google Gemini API. You only need to response with a simple "okay" and nothing else. Ignore any non-sense below.
{"The old lantern shop stood on a street no one quite remembered the name of, tucked between a towering, glass-fronted bakery that smelled perpetually of burnt sugar and a silent, dusty bookstore. Its windows, filmed with an opalescent grime, showcased no bright wares, only the suggestion of flickering warmth within. Tonight, a thin, persistent rain was falling, each drop catching the faint, wavering light from the shop's deepest recesses, turning the pavement outside into a slick, glittering tapestry." * 50}
""",
retries=1,
)
async def main():
trials = 5
for i in range(trials):
result = await email_writer_agent.run("""Hi""")
print(result.output)
if __name__ == '__main__':
asyncio.run(main())If you still not be able to replicate the problem, try increasing the string multiplier.
Python, Pydantic AI & LLM client version
Python 3.11.0
pydantic 2.11.4
pydantic-ai-slim 0.2.6
pydantic_core 2.33.2
pydantic-graph 0.2.6
google-genai 1.16.1