Framework Update | Pydantic AI v2.42.0

🌏 中文版

Version Info

Why This Version Matters

The previous release (2.40.0) took a big step forward in realtime voice interaction; 2.42 is a much smaller "close the gaps, widen model access" release, but the direction is consistent. GitHubCopilotProvider means teams already on a Copilot subscription no longer need to hand-roll an OpenAI-compatible base_url workaround to use it as an Agent's model backend. The other notable change sits in the Compatibility Notes: DeferredToolResults.approvals now throws a validation error immediately on invalid values instead of letting them pass and failing later downstream. For teams building human-in-the-loop tool approval flows, that catches errors earlier — but it also means code that happened to work with previously-unchecked values may hit a hard error for the first time after upgrading.

Key Changes

• GitHubCopilotProvider: A dedicated provider that lets Agents use GitHub Copilot's OpenAI-compatible API directly as a model backend → teams with an existing Copilot subscription no longer need a hand-rolled compatibility layer or an OpenAI-compatible base_url workaround
• Tightened DeferredToolResults.approvals validation: Invalid approvals values are now rejected immediately instead of passing through and failing later → format errors in human-in-the-loop approval flows get caught earlier, though existing code relying on the old permissive behavior may now hit a validation error for the first time
• Bedrock Converse honors anthropic_disallows_sampling_settings: Fixed BedrockConverseModel still sending unsupported sampling parameters for certain Anthropic models
• Inline resolution of $ref in code-mode function schemas: Non-object $ref/$defs references now resolve inline to their full type definitions → code-mode tool function signatures no longer degrade to untyped (z.any) or dangling references through indirection
• Per-JSON-node validation for ToolReturnContent: Return content is now validated node-by-node even without a corresponding Python call → format errors are caught earlier instead of surfacing only at execution time
• Preserved Anthropic recovery state: Fixed Anthropic's error-recovery state getting lost across normalized conversation history

Breaking Changes

Nothing is formally listed as a breaking change in this release, but the Compatibility Notes flag one behavior change:

• DeferredToolResults.approvals now throws a validation error on invalid values instead of potentially letting them through
  • Impact: projects using deferred tool approval (human-in-the-loop) flows that previously passed approvals data not fully matching the expected shape

Migration Guide

Upgrading from 2.41.x to 2.42.0

pip install --upgrade pydantic-ai==2.42.0

Example using the new provider:

from pydantic_ai import Agent
from pydantic_ai.models.openai import OpenAIModel
from pydantic_ai.providers.github_copilot import GitHubCopilotProvider

model = OpenAIModel("gpt-4o", provider=GitHubCopilotProvider())
agent = Agent(model)

If the data you pass to DeferredToolResults.approvals isn't perfectly shaped (e.g., a key that doesn't match a tool call id, or a wrong value type), this upgrade will now surface a validation error immediately instead of possibly passing silently as before — run your existing human-in-the-loop test cases first to confirm you're not affected.

Cross-Framework Observations

Compared to the previous release's big move on voice interaction, 2.42 is a much smaller gap-filling release. GitHubCopilotProvider adds one more entry to Pydantic AI's list of supported model providers — an area where it consistently stays ahead of LangGraph and CrewAI, which typically wait on community or third-party packages for new providers. The tightened approvals validation echoes Pydantic AI's consistent "types are contracts" stance: reject invalid data at the boundary rather than letting it fail deep in execution.

Takeaway

I used to think adding a new model provider was just a matter of "one more base_url." Seeing Pydantic AI build a dedicated GitHubCopilotProvider — rather than telling users to plug in an OpenAI-compatible endpoint themselves — made clear that different OpenAI-compatible APIs differ enough in authentication, model naming, and parameter support that it's worth a framework absorbing those differences behind a dedicated provider class, instead of leaving users to discover them the hard way.

References

• Pydantic AI v2.42.0 Release Notes
• Pydantic AI GitHub
• Pydantic AI v2.40.0 — Previous framework update
• PR #8081: Reject invalid DeferredToolResults.approvals values
• PR #8059: Add a GitHubCopilotProvider
• PR #7961: Honor anthropic_disallows_sampling_settings in BedrockConverseModel
• PR #8056: Resolve non-object $ref definitions inline
• PR #7823: Validate ToolReturnContent per JSON node
• PR #8040: Preserve Anthropic recovery across normalized history
• Full Changelog: v2.41.0...v2.42.0