dsh-done-check

English | 中文

Executable acceptance checklists for DeepSeek Harness coding tasks: every "done" is backed by a check that actually ran, and old results are flagged for re-check when the code changes.

What it can and cannot tell you

• A summary such as "3 checks: 2 passed, 1 awaiting manual confirmation" covers only the agreed checks. It does not mean the requirement is fully implemented. The plugin never shows a percentage or a "completion" score.
• It guards against an agent casually rewriting the verdict: claiming tests passed, quietly changing a check command, deleting a failing check, or forging a confirmation. It does not stop a malicious agent with an unrestricted shell, which can edit the tests themselves or the files under DSH home.
• Confirming a checklist does not authorize running its commands. Checks run on the shell executor of the current DSH composition; bash-local / pwsh-local provide no isolation, so compose the matching sandbox executor if you need it.
• watch is a scope you declare, not a dependency graph inferred for you.

Demo

A real DSH web session with v0.2 (DSH 0.1.5-rc.1, a small CSV export repository, commands only). Status text is Chinese.

1. /done-check previews the checklist: which commands will run and which check needs you.
2. /done-check 确认 confirms it. Confirming does not authorize anything; commands still go through your permission settings.
3. /done-check runs the checks and catches a real bug on the first try: fields containing commas were not quoted. The test report is quoted as printed.
4. After fixing toCsv, /done-check re-runs: both automatic checks pass and only the manual check is left.
5. /done-check ok records your verdict: all 3 checks pass.

The last frame shows that each step's conclusion is readable from the collapsed row titles alone.

Static preview · MP4

Status

v0.2.0, an early release. v0.2 redesigns only what you read and type; the verification rules are unchanged. The new output was checked on real DSH web hosts (0.1.5-rc.1 and 0.2.1-rc.5), with checks executed by the real pwsh tool inside its sandbox. Supported DSH: >=0.1.5-rc.1 <0.3.0. Design: docs/DESIGN.md (in Chinese). Status messages are currently in Chinese only.

Install

dsh plugin --profile <your-profile> add dsh-done-check

From a checkout: pnpm install && pnpm build, then dsh plugin --profile <your-profile> add /path/to/dsh-done-check.

Commands and tools

One command covers the main path: /done-check does the next thing that is due — preview the checklist, run the checks, or show the verdict. Status messages are currently in Chinese only.

	Who	What
/done-check	you	Not confirmed yet → shows the checklist and the commands it will run, runs nothing. Confirmed → runs the checks that were never run or whose code changed, then shows the summary.
/done-check 确认 (confirm)	you only	Confirm the checklist you were just shown. Only works on the exact list from the last preview; otherwise it shows the preview again.
/done-check ok [n] / no [n] [note]	you only	Record the result of a manual check. With one manual check waiting, no number is needed.
/done-check 详情 [n] (details)	you	Command, exit code, time, output tail, changed files
/done-check 全部 (all)	you	Every check, passing ones included
/done-check 技术详情 [n] (tech)	you	Executor and sandbox, full fingerprint, definition hash, log path
/done-check 重新检查 (run)	you	Force-run every automatic check
done_check_draft	Agent	Validate a drafted checklist (the Agent saves it with its own file tool)
done_check_run	Agent	Run the confirmed automatic checks
done_check_status	Agent	Read the status (returns verdict and nextStep)

n is the check's position in the summary; the internal id still works. show, pass and fail remain as aliases for one more version.

Output answers three questions: what was checked, what is still wrong, what to do now. The first line is always a complete conclusion, for example ❌ CSV 导出还有 1 项检查未通过. A failed check quotes up to three lines of the test's own output as "测试报告" — the plugin draws no conclusion from them. A check that could not run (permission denied, timeout, executor error) is reported as ⚠️ 没能运行, never as a failure and never as a pass.

Confirming is not authorizing. Typing /done-check after confirming executes the confirmed commands, through your session's normal permission rules. Every command goes to the shell tool registered in your DSH composition (pwsh/bash), so your approval prompts apply; a refused command leaves no evidence.

How it works

1. Write done-check.yaml in the project root (or let the Agent draft it). Name each check by something you can judge, and give manual checks a how:

   version: 1
   task: CSV 导出
   checks:
     - id: export-tests
       name: 导出内容与所选数据一致
       run: node test/export.test.mjs
       watch: ["src/**", "test/**"]
     - id: escape-tests
       name: 含逗号和换行的数据能正确导出
       run: node test/escape.test.mjs
       watch: ["src/**", "test/**"]
     - id: zh-filename
       name: 中文文件名可以正常下载
       manual: true
       how: 在浏览器里导出一个名为“报表”的文件，确认文件名没有乱码
   

2. /done-check previews it; you confirm with /done-check 确认. Confirmation and manual verdicts can only come from you, never from the Agent.

3. Checks run through DSH's shell tool; exit code, output, duration and a fingerprint of the workspace at that moment are recorded.

4. Any change to the workspace (uncommitted edits included) turns results bound to the old content into 🔄 代码已更新，需要重新检查, and lists the changed files in the details. Editing a check makes it ✏️ and asks you to re-confirm; adding or removing checks or rewording the task asks you to re-confirm the whole list.

5. A result whose run saw the workspace or definition change is 🔄 检查时代码正在变化，结果不可信; restoring the files does not revive it.

Development

pnpm install
pnpm test

Requires Node.js ≥ 22.19 and git.

License

MIT