@openai/agents-realtime@0.19.0
OpenAI Agents SDK published @openai/agents-realtime@0.19.0 , introducing tool approval policy hardening, error redaction, audio buffer stack overflow fixes, and conversation history ordering improvements.
Why this signal matters
Version 0.19.0 of @openai/agents-realtime includes minor and patch updates aligned with @openai/agents-core@0.19.0 . Key changes include binding conditional tool approvals to isolated normalized execution inputs, re-evaluating policies upon durable approval resumes, rejecting uncopyable normalized values prior to conditional approval, and redacting default function-tool and Realtime approval parse failures unless explicit invocation policies permit sensitive error traces. Patch fixes also resolve a stack overflow issue during large audio buffer encoding, fix task releases during automatic response coalescing, ensure conversation ordering during history updates and overlapping edits, and clear stale tools when transitioning to tool-less agents without a prompt.
Actionable summary
Upgrade to @openai/agents-realtime@0.19.0 to address fixes in conditional tool approvals (binding to isolated normalized input and re-evaluating policies on resumes), default error trace redaction, audio encoding stack overflow prevention, and conversation history preservation during updates.
- Agent usefulness
- 77/100
- Confidence
- 96%
- Canonical data
- JSON + Markdown
What builders should check
- Update dependency @openai/agents-realtime to ^0.19.0 .
- Verify tool approval workflows to ensure custom conditional approvals work with isolated normalized inputs and reject uncopyable values.
- Review error-handling logic if relying on approval parse failure traces, noting that redaction is now the default unless an explicit invocation policy is configured.