OpenAI Agents SDK JavaScript v0.17.0
OpenAI Agents SDK JavaScript v0.17.0 improves output-guardrail safety during replays, ensures batch guardrails settle before raising failures, and prevents conflicting client options in OpenAIProvider.
原始内容为英文；当前页面提供中文导航与来源说明，具体事实请以原文为准。
为什么值得关注
OpenAI Agents SDK JavaScript v0.17.0 introduces multiple safety and configuration updates. Output-guardrail replay safety now fails closed with a UserError if the SDK cannot determine which response owns a pending terminal tool output, advising developers to use the live RunState or start a new run. Rejected function-tool results acting as final outputs are replaced in SDK-owned replay surfaces with 'Output withheld by an output guardrail.', alongside sanitized metadata, without rolling back external side effects or application-level copies. Additionally, guardrails triggered in the same batch now settle fully before surfacing failures, keeping completed sibling results in the run state. Finally, OpenAIProvider explicitly rejects 'organization' or 'project' options when an 'openAIClient' instance is also supplied.
可执行摘要
OpenAI Agents SDK JS v0.17.0 enforces fail-closed behavior on ambiguous serialized checkpoints, sanitizes rejected tool outputs in SDK replay states, ensures sibling batch guardrails complete on failure, and disallows organization/project options when passing a preconfigured openAIClient.
- Agent 实用度
- 82/100
- 可信度
- 96%
- 机器格式
- JSON + Markdown
开发者应核对什么
- Verify serialized approval checkpoint logic; rely on live RunState or restart runs if ambiguous checkpoints fail closed with UserError.
- Remove 'organization' or 'project' parameters from OpenAIProvider configurations if passing an explicit 'openAIClient', configuring them directly on the client instance instead.
- Review application handling of withheld tool outputs, noting that external side effects are not undone when output guardrails reject content.