`message_history` doesn't work with `json.load`

[slkoo-cc]

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

Message history isn't loaded correctly when passing in pure Python list/dict object with json.loads.

Example Code

from pydantic_ai import Agent
import json

agent = Agent("gemini-2.5-flash", system_prompt="Be a helpful assistant.")

result1 = agent.run_sync("Tell me a joke.")
print(result1.output)
# > Why don't scientists trust atoms? Because they make up everything!

# simulate passing history into database and then loaded back out
msghistory = json.loads(result1.new_messages_json())

result2 = agent.run_sync("Explain?", message_history=msghistory)
print(result2.output)
# > I'd be happy to explain! But I need a little more context.
# >
# > What specifically would you like me to explain? For example:
# > *   A concept or idea?
# > *   A term or word?
# > *   A process or how something works?
# > *   Something I've said previously?
# > *   An event or phenomenon?
# > *   A piece of text, an image, or a situation you're looking at?

print(result2.all_messages())
# > [
# >     ModelRequest(
# >         parts=[
# >             SystemPromptPart(
# >                 content="Be a helpful assistant.",
# >                 timestamp=datetime.datetime(
# >                     2025, 9, 30, 5, 31, 16, 96163, tzinfo=datetime.timezone.utc
# >                 ),
# >             ),
# >             UserPromptPart(
# >                 content="Explain?",
# >                 timestamp=datetime.datetime(
# >                     2025, 9, 30, 5, 31, 16, 96163, tzinfo=datetime.timezone.utc
# >                 ),
# >             ),
# >         ]
# >     ),
# >     ModelResponse(
# >         parts=[
# >             TextPart(
# >                 content="I'd be happy to explain! But I need a little more context.\n\nWhat specifically would you like me to explain? For example:\n*   A concept or idea?\n*   A term or word?\n*   A process or how something works?\n*   Something I've said previously?\n*   An event or phenomenon?\n*   A piece of text, an image, or a situation you're looking at?\n\nOnce you tell me what you're curious about, I'll do my best to provide a clear explanation!"
# >             )
# >         ],
# >         usage=RequestUsage(
# >             input_tokens=9,
# >             output_tokens=819,
# >             details={"thoughts_tokens": 702, "text_prompt_tokens": 9},
# >         ),
# >         model_name="gemini-2.5-flash",
# >         timestamp=datetime.datetime(
# >             2025, 9, 30, 5, 31, 24, 921629, tzinfo=datetime.timezone.utc
# >         ),
# >         provider_name="google-gla",
# >         provider_details={"finish_reason": "STOP"},
# >         provider_response_id="LWvbaL0ygY3T6Q_K9fHQDQ",
# >         finish_reason="stop",
# >     ),
# > ]

Python, Pydantic AI & LLM client version

Python 3.12.9, Pydantic-AI 1.0.10, LLM not relevant (gpt-4 and gemini-2.5-flash both have the same problem)

[m7mdhka]

Hi @slkoo-cc
Try to use this:

from pydantic_ai import Agent
from pydantic_ai.messages import ModelMessagesTypeAdapter
import json

agent = Agent("gemini-2.5-flash", system_prompt="Be a helpful assistant.")

result1 = agent.run_sync("Tell me a joke.")
msghistory = ModelMessagesTypeAdapter.validate_json(result1.new_messages_json())
result2 = agent.run_sync("Explain?", message_history=msghistory)

[DouweM]

@m7mdhka Thanks Mohammed, that's exactly right.

@slkoo-cc Pydantic AI requires a list[ModelMessage], not just a list of dicts that look like model messages, so you need to use the ModelMessagesTypeAdapter to validate the dicts first.

[slkoo-cc]

Thanks, sorry did not read the documentation carefully.