修复任务给极简开局,构建任务给实干开局,模糊任务让模型自己选—— 首个工具调用后,25 个工具全部解锁。
Fix tasks get the Minimal anchor, build tasks get the doer anchor, ambiguous tasks self-pick — all 25 tools unlock after the first tool call.
let me ×208,55 次过程回复let me ×208, 55 interim repliesglob/grep/subagent 全没有,干不了完整开发glob/grep/subagent — can't do full devlet me——脑子和手脚都要let me across both runs — brain and hands数据:xiaobright/modeltest · 同一冻结题面
Source: xiaobright/modeltest · one frozen task
第一条消息分类:修复→极简句+读改;构建→实干句+读写;模糊→模型自选。
First message classified: fix→Minimal+edit, build→doer+write, ambiguous→self-pick.
system: You are a helpful software engineer assistant. tools: [bash, read, edit]
窄目录 + 唯一 persona + 清空上下文——最干净的开局切出最强策略。
Narrow catalog + single persona + cleared contexts — the cleanest opening.
reasoning: We need to ... strategy: anchored ✓
首个工具调用后:25 工具全开,plan-mode 恢复,零事后干预。
After the first tool call: all 25 tools, plan-mode back, zero post-anchor injection.
tools: [ask_user_question, bash, edit, glob, grep, subagent, web_search, workflow, write, ...] ×25
harness 0.1.0-rc.6 实测 · 分类与晋升状态存于持久化会话事件,刷新/恢复不回退
Measured on harness 0.1.0-rc.6 · mode and promotion live in durable session events
一条命令预设自动落盘;或手动复制 preset/ 目录。
One command lands the preset; or copy preset/ manually.
首轮自动按任务选锚:bash+read+edit 或 +write。
First turn auto-picks the anchor: bash+read+edit or +write.
首个信号后 25 工具 全部解锁,设为默认:agent-presets.default: anchored-standard。
All 25 tools unlock after the signal; make it default via agent-presets.default: anchored-standard.
| 对比Compared to | 得到You get | 代价Trade-off |
|---|---|---|
| vs Minimal | +23 工具 | 开局多一轮推理,几乎无感one extra bootstrap reasoning round |
| vs Standard | 任务匹配的轨迹锚定(98/99 vs 91/92)task-matched trajectory anchors (98/99 vs 91/92) | 首轮目录由关键词分类决定;模糊任务回落 weak 交给模型first-turn catalog is keyword-classified; ambiguous tasks fall back to weak |
诚实声明:98/99 来自 xiaobright/modeltest Project2 V4.1b——同一冻结题面 n=2, 不构成跨任务、跨模型的通用保证;本预设面向 Pro + reasoningEffort=max。
Honest boundaries: 98/99 is from xiaobright/modeltest Project2 V4.1b — n=2 on one frozen task, not a universal guarantee. Targets Pro + reasoningEffort=max.
Minimal 全程只有 2 工具。本预设只借它的「开局」,信号后展开全部 25 工具——Minimal 的脑子 + Standard 的手脚。
Minimal keeps 2 tools forever. This borrows only its opening; all 25 tools appear after the signal.
不行。切换会破坏会话,新会话首轮又是全目录,锚就丢了。两阶段必须在同一会话、同一请求流里。
No. Switching breaks the session and re-opens the full catalog — the anchor is gone.
第一条消息的关键词分类(修复/调试 vs 开发/创建),平局或无法判断时回落 weak,让模型自己选。
Keyword classification of the first message; ties fall back to weak and the model picks.
dsh plugin --profile web remove dsh-anchored-standard · rm -rf ~/.dsh/.agent-presets/anchored-standard