Pydantic AI v2.60.0 (2026-10-22)

Pydantic AI release v2.60.0 introduces several new sandbox harness capabilities, including DockerSandbox (supporting local Docker or Podman containers), BoxLiteSandbox (a local BoxLite micro-VM workspace), and existing VM reuse in SmolSandboxBackend. Key compatibility updates and deprecations include: - Deprecation of RetryPromptPart in favor of ToolReturnPart(outcome='retried') and RetryFeedbackPart. - Embedding models now raise standard model error categories like ModelRateLimitError. - ModelSelectionContext now includes usage limits, deadlines, run ID, conversation ID, output retry counts, and retry budgets. - BedrockConverseModel defaults maxTokens to Claude's maximum output while streaming behind the scenes. - MCP capability default IDs are derived strictly from the server URL hostname, rejecting multiple servers on one host without explicit IDs. - OpenAIDecisionsProvider is deprecated in favor of passing OpenAIProvider to OpenAIDecisionsModel. - Security and sanitization hardening: stripping request-level instructions from untrusted history, dropping client-supplied TextContent.metadata from user prompts, case-insensitive FileSystem denial patterns, and protecting .git files.

• Migrate code using RetryPromptPart to ToolReturnPart(outcome='retried') and RetryFeedbackPart.
• Replace usages of OpenAIDecisionsProvider with OpenAIProvider when configuring OpenAIDecisionsModel.
• Explicitly set IDs when defining multiple MCP servers hosted on the same hostname to avoid initialization errors.
• Verify error handling around embedding calls to support newly raised ModelRateLimitError exceptions.