OpenAI Agents SDK Python v0.23.0
OpenAI Agents SDK Python v0.23.0 adds configurable MCP listing limits, sandbox Docker removal protection, memory consolidation controls, and an encrypted history scan budget, along with various security, streaming, and tool lifecycle fixes.
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 has released version 0.23.0 of the OpenAI Agents SDK for Python. This update introduces several new configuration options and numerous bug fixes across core execution, sandboxing, and MCP integrations. New Features: - MCP: Added configurable page limits for MCP listings ( 5133). - Sandbox: Added opt-in Docker removal protection ( 5116) and configurable memory consolidation turn thresholds ( 5135). - Sessions: Introduced an opt-in encrypted history scan budget ( 5118). Key Bug Fixes and Hardening: - Security and Redaction: Default tool failure details ( 5112) and streaming task exception tracebacks ( 5121) are now redacted. Sensitive trace captures are respected for model metadata ( 5129). - Tool Handling and State: Rejected silently discarded kwargs tool arguments ( 5174). Resolved tools after agent start hooks ( 5124). Scoped function tool approvals to their owning agents ( 5144). Isolated and restored nested agent tool state to the tool's owning agent ( 5123, 5142). - Streaming and Lifecycle: Bounded agent tool streaming callback backlogs ( 5106) and validated stream callbacks prior to execution ( 5125). Avoided awaiting cleanup during coroutine closure ( 5163). - Reliability: Added retries for pre-request WebSocket handshake failures ( 4780). Preserved guarded final outputs in agent tools ( 5115) and structured guardrail diagnostics during persistence ( 5016).
可执行摘要
OpenAI Agents SDK Python v0.23.0 introduces configurable MCP listing page limits, Docker removal protection, memory consolidation turn controls, and an encrypted history scan budget. Key fixes include redacting default tool failure details and streaming tracebacks, rejecting discarded kwargs tool arguments, resolving tools after start hooks, and scoping function approvals and nested agent tool state to their owning…
- Agent 实用度
- 77/100
- 可信度
- 96%
- 机器格式
- JSON + Markdown
开发者应核对什么
- Review tool invocation code to ensure no arguments are being silently discarded as unhandled kwargs, as v0.23.0 now explicitly rejects them.
- Evaluate opt-in settings for Docker removal protection in sandboxed environments and the encrypted history scan budget for sessions.
- Verify whether any custom tracing or monitoring pipelines rely on unredacted tool failure messages or streaming tracebacks that are now redacted.