Jev + Opus 5.5 的测试驱动编码 agent 工作流
把测试结果喂进每一步决策:Jev 在 inspect/patch/test/ask/finish 间做类型化选择,Opus 写补丁,失败用例回流后自动修复并产出带验证记录的 PR 草稿。
Jev + Opus 5.5: code generation with test results fed into every next-step decision. This setup takes the agent from issue inspection to a change ready for review. Here’s how: /scope → Opus 5.5 reads the issue + repo rules. Define expected behavior, acceptance checks and the smallest useful fix. /state → collect the repro, failing test, relevant files, changed paths and unresolved questions. /choice → Jev selects the next available action: inspect / patch / test / ask / finish. Return a typed choice; use configured uncertainty handling. /patch → Opus writes a focused change + regression test. Explain which behavior should change and why. /checks → build the modified version, run the targeted test, then related checks. Capture exit codes + failure logs. In the demo, an edge case still fails. /recheck → feed new logs back into state. Jev chooses the next step; Opus repairs the remaining issue and tools rerun the checks. Enforce limits in code. /review → return the diff, test evidence and remaining risks in a PR draft. Auto-merge stays OFF. The result: a failed test becomes input for the next decision, and the final PR draft carries the verification record. Save this for your next coding agent setup ↓
这条是引用着另一条写的:x.com/0xchonsy/status/2107125199028044001