v2.56.0 v2.57.0

medium risk

pydantic/pydantic-aiOct 13, 2026 → Oct 16, 20262 days between releases

What's New

Adds MetaProvider for Meta Muse models and ApprovalReviewer for safer tool execution

Breaking Changes

  • BackgroundTools completions now include namespaced origin and tool_call_id metadata
  • Realtime history tracking for GPT-Live and Gemini Live moved to session core

New Features

  • Add MetaProvider for Meta Muse models using Responses API
  • Add Harness ApprovalReviewer to approve low-risk tool calls
  • Add dataset_name to EvaluationReport
  • Forward MCP Apps tool results to AG-UI and Vercel AI
  • Add MCPToolset.call_tool_for_app to relay MCP Apps tool calls

Improvements

  • Enhanced real-time history tracking and usage metrics for GPT-Live and Gemini Live
  • Skills now honour disable-model-invocation and user-invocable flags
  • Price requests calculated based on reported service tiers
  • Improved MCP Apps integration by forwarding tool results to frontend SDKs

Bug Fixes

  • Fixed token-based compaction retaining too little history
  • Fixed various session management bugs

Upgrade Notes

Update code parsing BackgroundTools metadata to handle new namespaced keys. Verify session logic compatibility with the new realtime history tracking source.

The Two Releases

v2.56.0

v2.56.0 (2026-10-13)

Oct 13, 2026· 8 days ago

This release introduces major enhancements to the Skills layer, including support for Python 3.15 and the new Gemini 4 Argon model. It also adds a SmolSandbox for local microVM execution and integrates with Render Workflows for durable execution.

Full release notes
v2.57.0

v2.57.0 (2026-10-15)

Oct 16, 2026· 6 days ago

This release introduces support for Meta Muse models via the Responses API and enhances real-time streaming capabilities for GPT-Live and Gemini Live. It also improves MCP (Model Context Protocol) integration and introduces a new `ApprovalReviewer` for safer tool execution.

Full release notes