jing09928

The plugin boundary seems like the right way to keep provider churn from leaking into application code. I’d be curious how teams decide which cross-cutting features belong in the core versus plugins once they need consistent policy and audit behavior.

The grader-focused trajectories are striking. Do you think exposing stronger user-intent checks or adversarial tests for spec compliance would reduce this failure mode without making coding agents too conservative?

模拟测试把语音 agent 的回归从“人工反复打电话”变成可重复的基础设施，这个方向很实用。实际落地时，你们更看重哪些生产指标来触发回放或新增场景：转写错误、工具调用失败，还是业务结果偏差？