dsh-gui-handoff
A DeepSeek Harness skill that hands a command-line-verified operation over to your GUI.
把命令行验证过的操作，交接给你的图形界面。
English · 中文
English
The problem
There is a gap nobody names: an agent can do something perfectly in a terminal,
and the user still cannot reproduce it by hand — because nobody ever walked the
GUI path and nobody wrote down what to click.
Telling the user "just run this script" is not a fix. It leaves them with a black
box they cannot learn from, that breaks on another machine, and that does not
cover the GUI operations which have no config file at all.
What this skill does
Two paths, path B by default:
| Path B (default) | Path A (fallback) |
|---|
| Approach | The agent actually drives the GUI once | Freeze the steps into config files |
| Output | A tutorial with real screenshots | A config file + one keystroke |
| User learns it | Yes | No |
| Survives a new machine | Yes | No |
Path B runs a five-stage chain — drive → observe → capture → reconstruct →
annotate — and insists on two things that are easy to skip:
- Ground truth, not vibes. Whether a build succeeded is decided by the
artifact's existence and timestamp. Whether the program ran is decided by its
output or the debugger's exit code. Screenshots are for the human, never the
verdict.
- Verify the user can reproduce it. "The machine ran it" is not "the user ran
it". The tutorial must list its own preconditions, the user must walk it, and
the scene must be reset afterwards so the user can tell their own success from
pre-existing state.
The part that is actually valuable
skills/gui-handoff/SKILL.md carries 30+ hard constraints, nearly all of them
paid for with real debugging time:
- Chinese/UTF-8 path failures in linkers and debuggers — and the trap that
a command line can succeed while the GUI fails, because the shell silently
re-encodes arguments into the encoding the tool happened to want
- Chinese IME corrupting injected keystrokes (
by → 不要)
- Why matching a window by title substring is dangerous (three separate
incidents), and what to do instead
- Why Chromium/Electron needs scancodes, and why focus can silently stay on
another window
- Why reading fine text from a screenshot needs crop-and-upscale first
- How a hand-written PNG encoder produced three horizontally-squashed copies
of every screenshot — and why that made a vision model look like a liar when it
was telling the truth
Plus seven working disciplines, every one of them written down after being
violated. The most important:
Check the criterion itself before accepting a verification.
A tutorial once said "if you see the artifact, it worked". The user followed it,
reported success, and the agent declared the acceptance test passed — but that
criterion only covered building, not running. When the criterion is wrong,
passing is worthless.
Install
dsh plugin --profile <your-profile> add dsh-gui-handoff
Or from a checkout:
dsh plugin --profile <your-profile> add /path/to/dsh-gui-handoff
The plugin registers one skill (gui-handoff, rank 600 — lowest priority, so a
locally edited copy in your project or user skills directory always wins).
What ships
lib/index.js skill provider (no runtime dependencies)
cordis.patch.yml bundle manifest
skills/gui-handoff/SKILL.md the skill itself
tools/gui-control/server.py reference GUI-control MCP server (Windows)
tools/notify-takeover.ps1 countdown notice shown before taking the mouse
The tools/ directory is a reference implementation for Windows, not a
requirement. The portable part is the workflow and the constraints; Linux and
macOS need their own equivalents.
Requirements
- DeepSeek Harness
>= 0.1.5-rc.1
- Path B's driver stage needs some way to drive the target OS's GUI. This repo
ships one for Windows; without one, the skill degrades to B2–B4
(user drives while the agent observes) rather than failing.
License
MIT.
中文
它解决什么
有一个没人命名的落差：agent 能在终端里把事情做对，用户却无法用手复现 ——
因为没人走过图形界面那条路，也没人写下来该点哪里。
告诉用户"你跑一下这个脚本"不是解法。它留下的是一个学不会的黑盒、
换台机器就废，而且根本覆盖不了那些没有配置文件的图形界面操作。
这个 skill 做什么
两条路，默认走 B：
| 路径 B（默认） | 路径 A（降级） |
|---|
| 做法 | agent 真的操作 GUI 走一遍 | 把步骤固化成配置文件 |
| 产出 | 带真截图的教程 | 配置文件 + 一键操作 |
| 用户学到 | 是 | 否 |
| 换机器可用 | 是 | 否 |
路径 B 跑一条五环链 —— 驱动 → 观察 → 记录 → 还原 → 标注 —— 并坚持两件
最容易被跳过的事：
- 用地面真值，不用感觉。 构建是否成功，看产物是否存在 + 时间戳；
程序是否真的运行了，看程序输出或调试器退出码。截图是给人看的，
永远不作判据。
- 验证用户能复现。「机器跑通」不等于「用户跑通」。教程必须列出自己的前提，
用户必须实走一遍，跑完还要重置现场 —— 否则用户分不清产物是自己跑出来的
还是早就有的。
真正值钱的部分
skills/gui-handoff/SKILL.md 里是 30 多条硬约束，几乎每一条都是拿真实的
调试时间换来的：
- 链接器与调试器的中文/UTF-8 路径失败 —— 以及那个陷阱：
命令行能过、图形界面却挂，因为 shell 会悄悄把参数重新编码成工具恰好想要的编码
- 中文输入法污染注入的按键（
by → 不要）
- 为什么按标题子串匹配窗口极其危险（三次独立事故），以及该怎么做
- 为什么 Chromium/Electron 需要扫描码，为什么焦点可能悄悄留在另一个窗口
- 为什么从截图里读小字必须先裁剪放大
- 一个手写 PNG 编码器如何把每张截图变成三份横向压扁的副本 ——
以及它如何让一个视觉模型背了"胡说"的黑锅，而它说的全是真的
外加七条工作纪律，每一条都是先违反、才写下的。最重要的那条：
验收之前，先检查判据本身。
有一份教程写着「看到产物就是成功」。用户照着做了、报告成功，agent 也据此
宣布验收通过 —— 但那个判据只覆盖构建，不覆盖运行。
判据错了，通过就是假的。
安装
dsh plugin --profile <你的 profile> add dsh-gui-handoff
或从本地目录安装：
dsh plugin --profile <你的 profile> add /path/to/dsh-gui-handoff
插件注册一个 skill（gui-handoff，rank 600 —— 最低优先级，因此你在项目或用户
技能目录里本地改过的同名副本总是优先）。
目录
lib/index.js skill provider（无运行时依赖）
cordis.patch.yml bundle 清单
skills/gui-handoff/SKILL.md skill 本体
tools/gui-control/server.py 参考用的 GUI 控制 MCP server（Windows）
tools/notify-takeover.ps1 抢鼠标前的倒计时提示
tools/ 是面向 Windows 的参考实现，不是必需项。可移植的是流程与约束；
Linux / macOS 需要各自的对等物。
依赖
- DeepSeek Harness
>= 0.1.5-rc.1
- 路径 B 的驱动环节需要某种驱动目标系统图形界面的手段。本仓库提供一个
Windows 实现；没有它时，skill 会降级到 B2–B4（用户操作、agent 观察），
而不是直接失败。
许可
MIT。
