deepseek-harness
deepseek-ai
DeepSeek Harness: Everything is a Plugin.
PROJECT TOPICS
INSTALL REFERENCE
dsh plugin --profile web add github:ZRui-C/dsh-minimal-first-turn
该命令指向仓库当前默认分支;尚无绑定当前 commit 的完整验证结果。
PROJECT README
dsh-minimal-first-turn is a DeepSeek Harness Web plugin that makes an enabled
root session's first model request smaller and closer to the official Minimal
preset, without permanently giving up the selected agent preset.
It is inspired by the first-request conditioning work in
xiaobright/dsh-anchored-standard.
This project is independent, experimental, and not affiliated with DeepSeek.
While First-turn minimal is enabled:
bash and
str_replace_editor.tool/call or assistant/message restores the selected
preset's original prompt and complete tool catalog.compaction/end, the next request enters the same controlled phase.The composer contains a persistent 首轮精简 switch. It is global to the current DSH home, not per-session. Disabling it removes this plugin's agent-scoped Minimal tools and stops all filtering for future requests.
This package targets DSH Web with @deepseek-ai/* 0.1.0-rc.6 packages.
A persistent Bash PTY is required, so the current release supports macOS and
Linux hosts; Windows is not supported yet.
dsh plugin --profile web add dsh-minimal-first-turn
Restart the existing dsh web process, then open a conversation. The
首轮精简 switch appears beside the composer controls.
The toggle state is stored at:
$DSH_HOME/plugins/dsh-minimal-first-turn.json
When DSH_HOME is unset, the path is ~/.dsh/plugins/dsh-minimal-first-turn.json.
pnpm install
pnpm check
npm pack --dry-run
For a local Web profile, add the package as a dependency and mount its
cordis.patch.yml, then restart dsh web. Host changes require a restart;
client-only changes require a page reload when the Web HMR watcher is not
running.
MIT. The implementation includes derivative work from the MIT-licensed
dsh-anchored-standard project and DeepSeek Harness packages; the required
attribution is included in LICENSE.
这是一个 DSH Web 插件。开启“首轮精简”后,新根会话的第一轮模型请求会使用
Minimal system prompt、持久 bash 与 str_replace_editor,并移除自动注入的
工作区说明和技能目录。首次 tool/call 或 assistant/message 后,当前预设的
完整 prompt 与工具目录恢复;发生上下文压缩后,下一轮会再次进入首轮精简阶段。
它不保证模型输出固定的推理措辞,只控制模型可见的首轮条件。开关是全局持久设置, 而不是单个会话设置。
CLASSIFICATION EVIDENCE
系统优先读取 GitHub Topics,再与站内分类词典和词根规则比对。当前命中: 无有效分类标签。