use unified thinking for pydantic-ai

[Light2Dark]

📝 Summary

Closes #9417 . This avoids specifying thinking and params for individual models.

https://pydantic.dev/docs/ai/advanced-features/thinking/#unified-thinking-settings

📋 Pre-Review Checklist

• For large changes, or changes that affect the public API: this change was discussed or approved through an issue, on Discord, or the community discussions (Please provide a link if applicable).
• Any AI generated code has been reviewed line-by-line by the human PR author, who stands by it.
• Video or media evidence is provided for any visual changes (optional).

✅ Merge Checklist

• I have read the contributor guidelines.
• Documentation has been updated where applicable, including docstrings for API changes.
• Tests have been added for the changes made.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
marimo-docs	Ready	Preview, Comment	May 7, 2026 9:00pm

[cubic-dev-ai]

No issues found across 4 files

Architecture diagram

sequenceDiagram
    participant Client as Client/UI
    participant API as AI Endpoint
    participant Factory as ProviderFactory
    participant Provider as Provider (OpenAI/Azure/Anthropic/...)
    participant Agent as pydantic-ai Agent
    participant Model as pydantic-ai Model
    participant Profile as ModelProfile

    Note over Client,Profile: Unified Thinking Flow (PR introduces _default_thinking pattern)

    Client->>API: POST /ai/chat (messages, model config)
    API->>Factory: create_agent(model_name, max_tokens, ...)
    Factory->>Provider: create_model(max_tokens)
    Provider->>Profile: lookup profile(model_name)
    Profile-->>Provider: supports_thinking, thinking_always_enabled, etc.
    Provider-->>Factory: Model instance (with model-specific settings)

    Factory->>Provider: _build_agent_settings(model)
    alt Provider is OpenAI/Azure OpenAI
        Provider->>Provider: _default_thinking(model)
        alt Custom base_url (non-OpenAI endpoint)
            Provider-->>Factory: None (suppress thinking)
        else Official OpenAI endpoint
            Profile-->>Provider: supports_thinking? (from model profile)
            alt Profile says supports_thinking or thinking_always_enabled
                Provider->>Provider: openai_reasoning_summary = "auto"
                Provider-->>Factory: ModelSettings(thinking=True, openai_reasoning_summary="auto")
            else Profile does NOT support thinking
                Provider-->>Factory: None
            end
        end
    else Provider is Anthropic
        Provider->>Provider: _default_thinking(model)
        Profile-->>Provider: supports_thinking? (true for all Anthropic models)
        Provider-->>Factory: ModelSettings(thinking=True)
    end

    Factory->>Agent: Agent(model, model_settings=settings, toolsets, instructions, output_type)
    Agent-->>Factory: Agent instance
    Factory-->>API: Agent instance

    Note over API,Agent: Request execution uses the unified thinking setting

    API->>Agent: run(message, message_history)
    Agent->>Model: generate(messages, model_settings)
    Model->>Model: Apply thinking from model_settings
    Model-->>Agent: Response with reasoning/thinking steps
    Agent-->>API: Streamed response
    API-->>Client: StreamingResponse (includes reasoning content)

Loading

[Copilot]

Pull request overview

This PR updates marimo’s pydantic-ai integration to use unified thinking settings instead of per-model/provider “thinking” parameters, primarily to support Claude Opus 4.7 (which requires adaptive thinking payloads) and to simplify reasoning configuration across providers.

Changes:

• Introduces agent-level “thinking” settings via model_settings and removes provider-specific thinking params (e.g., Anthropic beta thinking config, Google thinking config, OpenAI reasoning effort).
• Updates Anthropic model handling to respect pydantic-ai’s profile for sampling restrictions (e.g., Opus 4.7 disallowing temperature).
• Bumps pydantic-ai-slim lower bounds in dependency groups and expands provider tests to validate unified thinking behavior and Anthropic payload translation.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 2 comments.

File	Description
marimo/_server/ai/providers.py	Moves “thinking” to agent-level settings and relies on pydantic-ai model profiles; adjusts OpenAI/Azure/Anthropic/Google/Bedrock behavior accordingly.
tests/_server/ai/test_providers.py	Reworks tests to assert unified thinking behavior and validates Anthropic thinking payload translation (adaptive vs enabled).
pyproject.toml	Bumps pydantic-ai-slim minimum versions in dependency groups to >=1.84.0.
packages/llm-info/data/models.yml	Adds Claude Opus 4.7 model metadata (including thinking: true).

[Light2Dark]

claude 3.5 and below aren't even available anymore by Claude

[github-actions]

🚀 Development release published. You may be able to view the changes at https://marimo.app?v=0.23.6-dev17