dev-eval

原文涵盖的主题

• dev-eval
• When to Use
• Checklist
• Rationalizations
• Red Flags
• Verification

Paths: the bin/ tools and the library/ folder live in /.agents/dev-kit/; run tools by that path. Task: no DEVKITTASK set? Run /.agents/dev-kit/bin/ledger start first (it prints your id) and prefix dev-kit commands with DEVKITTASK= . below is .devkit/tasks/ at the repo root; never write run files anywhere else. Also read .devkit/memory.md if it exists. Open the host todo list (todowrite, TodoWrite or TaskCreate, whichever exists; else a checklist in /state.md) with this checklist verbatim, and tick items as steps finish. A skipped step stays listed with a reason. Run bin/ledger after every step (phase=eval). A skill, prompt or workflow edit and you want to know if it makes the agent better. Choosing between two variants of a prompt or skill. Not for building the change (use dev-feature) or fixing a bug (use dev-debug). You own: a blind comparison and a promote, drop or rerun verdict. Reads: always library/principles (step 0), library/state. [ ] 0. Skim library/principles. [ ] 1. Frame: name the variant under test and what counts as success. Write a rubric of 3 to 6 concrete criteria for the judge only; candidates never see it. [ ] 2. Set up sanitized environments outside the repo (o

这里只展示节选，请以完整原文和当前要求为准。