OpenAI Agents SDK 0.22.0 redacts guardrail-blocked tool output from replayable state and rejects conflicting provider config

GitHub (openai/openai-agents-python)·medium signalXBlueskyLinkedInCopy link

Shipped 19 Aug 2026, v0.22.0 is a minor bump because it hardens the runtime and tightens a provider contract: terminal function-tool output rejected by an output guardrail is now redacted from replayable and persisted SDK state, so a blocked result cannot leak back through a replay. Non-streaming Responses with terminal status `failed` or `incomplete` now raise `ModelBehaviorError` instead of passing through, usage accounting is isolated between independent `RunState` checkpoints, and `OpenAIProvider` rejects `organization`/`project` when an explicit `openai_client` is supplied. That last one is a breaking change for anyone who set both.