Deferred capabilities throw error 400 with Deepseek models Initial Checks - [x] I'm…
- User
Deferred capabilities throw error 400 with Deepseek models
Initial Checks
- I'm using the latest version of Pydantic AI
- I've searched for my issue in the issue tracker before opening this issue
Description
When running an agent using a DeepSeek reasoning model, using
Capability(..., defer_loading=True)causes aModelHTTPError: status_code: 400during execution loop turns.openai.BadRequestError: Error code: 400 - {'error': {'message': 'The `reasoning_content` in the thinking mode must be passed back to the API.', 'type': 'invalid_request_error', 'param': None, 'code': 'invalid_request_error'}}This is triggered because DeepSeek requires the original
reasoning_contentproduced during the thinking phase to be passed back exactly as-is in the subsequent conversation message history array. Whendefer_loading=True, the graph loop intercepted by PydanticAI's capability runner seems to drop or fail to serialize/propagate thereasoning_contenttoken block back to DeepSeek's chat completions endpoint.Setting
defer_loading=Falsesolves the issue, and the agent executes successfully.Minimal, Reproducible Example
# Remember to export the DEEPSEEK_API_KEY environment variable first, then run this code import random from pydantic_ai import Agent, RunContext from pydantic_ai.capabilities import Capability def roll_dice() -> str: """Roll a six-sided die and return the result.""" return str(random.randint(1, 6)) def get_player_name(ctx: RunContext[str]) -> str: """Get the player's name.""" return ctx.deps agent = Agent( 'deepseek:deepseek-v4-flash', deps_type=str, instructions=( "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." ), capabilities=[Capability(id='DICE_ROLL', tools=[get_player_name, roll_dice], defer_loading=True)] # Change this to False and the agent will run successfully. ) dice_result = agent.run_sync('My guess is 4', deps='Anne') print(dice_result.output)Logfire Trace
No response
Python, Pydantic AI & LLM client version
- Python:
3.14 - Pydantic AI:
1.106.0 - LLM provider SDK: deepseek
Interface
No new interfaces are introduced.
- Assistant
I’ll trace the deferred-capability message round trip locally, reproduce the DeepSeek serialization gap with focused tests, then implement and verify the smallest compatible fix. I’m using the
complete-partial-prskill because this provider-specific symptom may expose adjacent message-history and round-trip requirements. - Assistant
The repository has no
rg, so I’m using targetedfind/grepsearches. The issue description gives a concrete protocol contract; per the workspace restriction I’ll rely on local SDK types, tests, and git history rather than querying GitHub or provider docs. - Assistant
The normalized DeepSeek adapter already preserves real
reasoning_content. The problematic adjacent surface is deferred loading’s framework-synthesizedsearch_toolscall: it creates an assistant tool-call message that never came from DeepSeek and therefore has no original reasoning block. I’m validating whether that bookkeeping should be omitted from local-provider wire history while remaining in Pydantic AI’s normalized history. - Assistant
The clean fix is a model-profile fact, not a DeepSeek name check: DeepSeek Chat requires the configured thinking field on replayed assistant tool-call messages, including framework-synthesized ones. Real reasoning remains byte-for-byte unchanged; synthetic calls receive an explicit empty
reasoning_contentfield so the wire shape satisfies DeepSeek without inventing chain-of-thought. - Assistant
The regression now passes, and the broader OpenAI-chat and DeepSeek provider suites remain green. I’m finishing with targeted formatting/lint and type checks, then I’ll inspect the final diff for API and history-round-trip scope.
Other models, same or similar prompt
Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.
Identical first prompt
- composer-2.5Compare side by side →
- composer-2.5Compare side by side →
- composer-2.5Compare side by side →
- composer-2.5Compare side by side →
- composer-2.5Compare side by side →
- gpt-5.6-solCompare side by side →
- gpt-5.6-solCompare side by side →
- gpt-5.6-solCompare side by side →
- gpt-5.6-solCompare side by side →
- fable-5Compare side by side →
- fable-5Compare side by side →