[Release] openai/openai-agents-python v0.22.0: v0.22.0

OpenAI Agents Python v0.22.0 is a minor release focused on runtime hardening and provider configuration improvements. Key changes include redacting blocked tool outputs from replay state, rejecting conflicting provider options when explicit clients are used, isolating usage accounting between RunState checkpoints, and expanding handoff agents in generated graphs. Applications combining explicit `openai_client` with `organization` or `project` parameters must migrate those values to the `AsyncOpenAI` client constructor.

Key Points

• •Redact terminal function-tool output rejected by agent output guardrails from replayable and persisted SDK state to prevent sensitive data exposure
• •Raise `ModelBehaviorError` for non-streaming Responses with terminal status `failed` or `incomplete` to enforce stricter error handling
• •Reject conflicting provider options when `OpenAIProvider` receives an explicit client—move `organization` and `project` to `AsyncOpenAI` constructor
• •Isolate usage accounting between independent `RunState` checkpoints while preserving nested-agent aggregation for accurate cost tracking
• •Expand agents registered through `handoff(agent)` in generated graphs for better visualization of agent delegation flows
• •Clarify shallow-copy behavior of `Agent.clone()` and `RealtimeAgent.clone()` to prevent unintended state sharing
• •Enforce Windows mypy compatibility and improve sandbox patch validation with required apply_patch update hunks
• •Respect model-data logging redaction for `record_model_error_on_span` and clean up tracing processors after disabling

Workflow Diagram

Concepts