GPT-6 SOL + Jev 自修复编码 Agent 提示词
一份可直接复制的提示词模板：GPT-6 SOL 诊断并出补丁，Jev 排序下一步，跑测试验证后写入项目记忆。
GPT-6 SOL + JEV: THE SELF-IMPROVING CODING AGENT PROMPT Give this to your AI Agent when the same failure keeps coming back: { "goal": "Turn one recurring coding failure into a verified fix and a reusable project lesson.", "inspect": ["repo", "tests", "run logs", "available tools", "Jev interface"], "workflow": [ "Find a repeated error, reproduce it, and record a baseline.", "GPT-6 SOL: diagnose the cause, propose one focused patch, and write a regression test.", "Jev: rank valid next steps. Code: run tests and record the results.", "Compare the patch with the baseline. Reject regressions and bound retries.", "Save the root cause, fix, and test evidence only after verification." ], "return": ["reproduction", "patch", "test results", "before/after metrics", "project lesson"], "rules": ["Use only granted tools.", "Confirm Jev access before calling it.", "Never claim a test passed without evidence."] } GPT-6 SOL investigates the cause and proposes a patch → Jev chooses the next bounded step → code runs the tests → verified fixes become project memory. The next run starts with a regression test and a record of what worked. Fail → fix → verify → remember Save this for your next build ↓
这条是引用着另一条写的：「把 Jev 决策层拆成 10 步 agent 落地蓝图」