ISSUE DESCRIPTION Deferred capabilities throw error 400 with Deepseek models Initial…
- User
ISSUE DESCRIPTION
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.
ADDITIONAL ADVICE
Since you are given a git repository, you can use git commands to simplify your work. For example, if you made a mistake and want to revert your changes, you can use
git checkout <file>to restore the file to its original state. You can also reset all changes in the repository usinggit reset --hardcommand. Additionally, you can usegit stashandgit stash popto temporarily save your changes and restore them later. Finally, you can see the changes that you've made (which will be submitted when you callsubmit) usinggit statusorgit diffcommands. However you don't need to usegit addorgit commitbefore submitting your changes. In fact,submitonly submits currently unstaged changes because it usesgit diffto get the changes that need to be submitted.CHECKLIST
Before submitting your solution, please go over the following checklist and make sure you've done everything:
- If an error was reported in the issue description, I have successfully reproduced it.
- If an error was reported in the issue description, I have confirmed that my fix resolves the error.
- I have thought about edge cases and made sure my fix handles them as well.
- I have run existing tests in the repository that might have been affected by the change I made and confirmed that they pass. I want you to list every bullet from this checklist and write your reasoning for why you think you did it or didn't need to.
Repository has been uploaded and your shell is currently at the repository root. Time to solve the issue!
(Current directory: /pydantic-ai, current file: none) bash-$
- Assistant
I'll start by exploring the repository structure and locating the deferred capability / DeepSeek reasoningContent handling code.
- Assistant
I'll search the codebase for deferred loading, reasoning_content, and DeepSeek handling.
- Assistant
rg is not available, let me use grep.
- Assistant
I'll look at how OpenAI/DeepSeek message history serializes thinking and reasoning_content.
- Assistant
Let me look specifically at OpenAI model and DeepSeek profiles for reasoning_content.
- Assistant
Next I'll inspect deferred capability loading and how it rewrites or filters messages.
- Assistant
I'll open the deferred capability loader and OpenAI thinking mapping next.
- Assistant
I'll read the OpenAI thinking mapping and check whether message history is stripped of ThinkingParts.
- Assistant
I'll check recent history around deferred loading and thinking, plus DeepSeek profile settings.
- Assistant
I'll inspect DeepSeek's profile and the deferred capability toolset next.
- Assistant
I'll check tests and code about reasoning_content with tool calls, and recent DeepSeek/thinking fixes.
- Assistant
I'll look specifically at message construction after load_capability and any filtering of ThinkingParts.
- Assistant
_clean_message_historylooks promising — I'll inspect that and related history cleanup. - Assistant
Let me look at
_clean_message_historyspecifically. - Assistant
I'll check how DeepSeek provider constructs OpenAI profiles and how assistant messages with tools map thinking.
- Assistant
I'll inspect ThinkingPart definition and multi-message mapping when tool results are present.
- Assistant
I'll inspect LoadCapability return handling and how tool results reshape the message history.
- Assistant
I'll look for related issues updates, CHANGELOG notes, and how tool returns interact with thinking.
- Assistant
I'll search GitHub / the repo for the reported issue and related PRs.
- Assistant
I'll study deferred capability tests and how FunctionModel could reproduce the history flow.
- Assistant
I'll re-read DeepSeek reasoning docs and huddle patterns around multi-turn tool use.
- Assistant
Checking when Thinking is auto-injected, and what happens to
thinkingmodel settings around deferred tools. - Assistant
I'll write a quick mock-based reproduction to inspect the second OpenAI request payload after
load_capability. - Assistant
I'll build a temporary mock script that captures outbound OpenAI requests after deferred capability loading.
- Assistant
The synthetic
search_toolsassistant message is missingreasoning_content. I'll find where that auto-load message is injected. - Assistant
I'll open the tool-search auto-load path next.
- Assistant
I'll check similar handling elsewhere and DeepSeek tests covering multi-turn tool reasoning.
- Assistant
I'll fetch DeepSeek's multi-turn reasoning guidance for the expected payload shape.
- Assistant
I'll implement ensuring assistant messages include the thinking field when any sibling message already has it.
- Assistant
Let me look at the _merge_leading_system_messages to place a similar helper nearby.
- Assistant
I'll add a helper to backfill thinkingfields and call it from
_map_messages. - Assistant
I'll wire the helper into
_map_messagesand fix the docstring typo. - Assistant
The second replace needed the right line range — fixing
_map_messagesnext. - Assistant
Now I can do the replace.
- Assistant
I'll re-run the reproduction to confirm yourselves synthetic messages now include
reasoning_content. - Assistant
I'll run the OpenAI thinking mapping and DeepSeek-related tests next.
- Assistant
Checking whether subclasses override
_map_messagesand may skip the new ensuring eagerness. - Assistant
Checking OpenRouter's
_map_messagesfor the same failure mode. - Assistant
Checking xAI mapping — it may have the same reasoning_content requirement.
- Assistant
I'll unit-test the helper for edge cases and clean up the temp script afterwards.
- Assistant
The first call mutated the message list — I'll re-run the edge-case checks with fresh inputs.
- Assistant
I'll review the final diff and run one more related OpenAI test sweep.
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
claude-sonnet-5Compare side by side →
claude-sonnet-5Compare side by side →
claude-sonnet-5Compare side by side →
claude-sonnet-5Compare side by side →
claude-sonnet-5Compare side by side →
Qwen/Qwen3.6-35B-A3BCompare side by side →
Qwen/Qwen3.6-35B-A3BCompare side by side →
Qwen/Qwen3.6-27BCompare side by side →
Qwen/Qwen3.5-35B-A3BCompare side by side →
Qwen/Qwen3.5-35B-A3BCompare side by side →
Qwen/Qwen3.5-35B-A3BCompare side by side →
Qwen/Qwen3.5-35B-A3BCompare side by side →