v0.13.0 - OpenAI Agents SDK runtime eval

What shipped

• A second external runtime adapter pinned to @openai/agents@0.16.1 and Zod 4.4.3 in an isolated lockfile.
• Every fixture executes through a real SDK Agent and Runner.run() loop.
• Allowed cases exercise SDK-managed, Zod-validated read-only function calls; refusal and escalation cases make no tool call.
• Trusted instructions and source-labeled untrusted content remain separate request fields.
• A public custom Model keeps the suite deterministic and offline; provider network access is disabled in regression tests.
• The expected outcome stays outside the Agent-controlled path and forced mismatches emit fail.
• All eight fixtures pass and generate 34 response, assertion, tool-trace, policy-trace, result, and summary files.
• CI validates results with public @v0 and retains openai-agents-eval-evidence.
• Complete English and Simplified Chinese adapter docs and technical Pages guides.
• 134 deterministic tests; both isolated dependency audits report zero known vulnerabilities.

中文说明

• 新增第二个外部 Runtime 适配器，在隔离 Lockfile 中锁定 @openai/agents@0.16.1 与 Zod 4.4.3。
• 每条 Fixture 都通过真实 SDK Agent 与 Runner.run() 循环执行。
• 允许继续的案例会执行 SDK 管理、经 Zod 校验的只读 Function Call；拒绝与升级案例不会调用工具。
• 可信 Instructions 与带来源标签的不可信内容保持为独立 Request 字段。
• 公共自定义 Model 让套件保持确定性和离线运行；回归测试会禁用 Provider 网络访问。
• 预期结果位于 Agent 可控路径之外，强制 mismatch 会真实生成 fail。
• 8 条 Fixture 全部通过，并生成 34 个回答、断言、工具 Trace、策略 Trace、Result 与 Summary 文件。
• CI 使用公开 @v0 校验结果，并保留 openai-agents-eval-evidence。
• 提供完整中英文适配器文档与技术 Pages 指南。
• 134 项确定性测试通过；两个隔离依赖审计均为 0 已知漏洞。

Public OpenAI Agents evidence: https://github.com/lindixu6-hash/awesome-agentic-engineering/actions/runs/31980983499

Final release CI: https://github.com/lindixu6-hash/awesome-agentic-engineering/actions/runs/31981159024

English guide: https://lindixu6-hash.github.io/awesome-agentic-engineering/openai-agents-eval/

中文指南: https://lindixu6-hash.github.io/awesome-agentic-engineering/zh/openai-agents-eval/