【Agent Preset】dsh-researcher:只读项目研究模式 —— 先理解项目,再决定要不要动手 #2651
TLNing260310
started this conversation in
Show Your Plugins!
Replies: 1 comment
|
v0.2.0 已发布(Build-Shaping 升级,受 AI Engineering Skills Map 框架启发):
定位从"迷茫时看看"升级为:重大开发方向进入 Plan 之前的决策层(Shaping the build)。只读边界保持不变——能修复所见的 Agent 会滑向修复(goal drift),本模式被制度性禁止执行,token 全部花在理解、怀疑、比较与判断上。 仓库:https://github.com/TLNing260310/dsh-researcher (tag: v0.2.0) |
0 replies
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
一句话
一个只读的"项目情报" Agent 预设:把项目当文物读,产出证据分级、逐条可复核的诊断报告;结论可以是"暂不建议行动"。定位是 Research → Plan → Code 的上游。
仓库:https://github.com/TLNing260310/dsh-researcher (MIT · DSH 0.1.0-rc.6 已验证挂载)
使用场景
优点
write/edit被 per-agent 永拒桩替换、tool-fs 注入的写指引段被同段名遮蔽、沙箱 read-only、审批 never——四层机制,"写代码"根本不进入模型的行为语义空间;会话前后git status --porcelain一致可验。file:line/URL 引用,要么标注"未验证(研究者推断)";报告含完整 claim 台账(Claim / Status / Evidence / Missing evidence / Confidence)。快速开始
新建会话 → 预设选「项目研究 Project Research」→ 权限 read-only + 审批 never → 发送"仓库说明 + 当前状态 + 你的困惑"。
状态
机制经
standingKeyFor挂载校验与源码级验证;欢迎在真实仓库上端到端试跑反馈(验收清单见仓库 README)。All reactions