@gorban/dsh-job-stop
DeepSeek Harness 插件：在会话标题栏的作业列表中停止正在运行的后台作业，并显示确认对话框。
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add github:gorban/dsh-job-stop#28c394148cc675cabdd9e173db7f83e49243dcc0
Job Stop 以 @gorban/dsh-job-stop 发布，当前版本为 0.1.2。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
Stop a running background job from the DeepSeek Harness web session header — hover its row in the background-job list, click the stop sign, confirm in a dialog that restates the full command.
Hovering a running job row, then confirming the stop
Hovering a running row covers its animated state dot with a stop button; the dialog restates the command, kind, status, elapsed time, and job id before anything is cancelled.
The session header's background-job list becomes actionable for the human, not just readable:
Stop background job: <command>).Modal + Button composition the session rename and workspace delete dialogs use — restating the full command, the job kind, its status, how long it had been running, and its registry id before anything is cancelled.ctx.jobs, the row flips to stopping and then killed from the ordinary pushed frame, and the owning agent is told that a human stopped its job.Everything is localized (English and Simplified Chinese) and keyboard reachable: the stop button takes focus and reveals itself on :focus-visible, Escape closes the dialog, and prefers-reduced-motion disables the fades.
ctx.jobs already ran every background bash/pwsh/PTY/subagent job, and dsh-tool-jobs already exposed job_kill — to the model. The web list (dsh-client-ui-jobs) was a deliberately read-only projection, and its own README named the reason cancellation was deferred:
Rows are read-only — … Cancellation additionally owes a model-facing decision the seam does not answer:
kill()marks terminal delivery reported, so an interrupt written against the current contract would leave the model believing its job is still running.
That is the whole problem this plugin solves, and the design note 2026-08-08-web-background-job-display left stopping in the wire union precisely so this phase would not need a wire change. Two consequences shape the implementation:
JobRegistry.kill sets reported = true, and dsh-tool-jobs skips its completion notice for a reported job. So after a successful kill this plugin injects its own plugin-sourced notice (inject while the agent is busy, followup when it is idle and notice: wakeup), or the model would keep waiting on a job that no longer exists.kill on a terminal record also sets reported, which would swallow the completion notice the model is still owed. So the endpoint reports already-finished without calling kill whenever the job has left running.From npm — prebuilt, so the install needs no build step and no allowBuilds approval:
dsh plugin --profile web add @gorban/dsh-job-stop
From GitHub — no npm account involved:
dsh plugin --profile web add github:gorban/dsh-job-stop
From a release tarball — for a host that cannot reach npm or a GitHub source archive:
dsh plugin --profile web add https://github.com/gorban/dsh-job-stop/releases/latest/download/dsh-job-stop.tgz
Then restart dsh (a plugin row is mounted at profile load). The command adds the dependency and, because package.json declares dsh.bundle, lists the package in dsh.profile.bundles so its cordis.patch.yml row is applied.
Requires dsh 0.1.2-alpha.1 or newer, declared as engines.dsh so the Plugin Market's host-aware filter can read it: both the jobsBySession mirror this list renders from and the ctx.connection.fetch route seam the stop endpoint is claimed on first shipped in that release.
The shipped @deepseek-ai/dsh-client-ui-jobs row can stay enabled: this plugin registers its list under that entry's own slot cell id (job-list) at a lower priority, so the header keeps exactly one control and the read-only list is shadowed rather than duplicated. This plugin also renders correctly with that row disabled, because it reads the same jobsBySession mirror.
Host half (src/index.ts) claims one authenticated route on the Connection fetch registry:
ctx.connection.fetch.register({
path: '/api/dsh-job-stop/kill',
methods: ['POST'],
requestBody: 'buffered',
fetch: request => handleStop(ctx, config, request),
})
That registry is deliberate, and it is not the obvious choice. ctx.webServer.register({ kind: 'exact', path: '/api/...' }) — the pattern a plugin reaches for first — matches the web server's exact table before Connection's /api prefix route, so it runs ahead of requestRejection and therefore outside the host/Origin trust and browser-authentication fence. That is fine for a prompt toggle and wrong for a mutating kill, so this endpoint is claimed on the shared fetch handler instead, where the fence has already run.
The handler then:
{ sessionId, jobId } and rejects malformed bodies (400);ctx.agents.get(...) and refuses when there is none (409) — ids are predictable, so authorization, not secrecy, is the boundary, and the registry fences kill by that Agent;ctx.jobs.get(...), never ctx.jobs.read(...): read consumes the job's single output cursor and would silently take bytes the model's job_output will never see;already-finished for a job that has left running, otherwise calls ctx.jobs.kill(...) and, on requested, delivers the notice.Browser half (src/client/) shadows the built-in list entry and adds the hover affordance. It talks to the registry only through that one POST; job state still arrives through the pushed jobsBySession mirror, so no polling or optimistic row state is involved.
- id: job-stop
name: "@gorban/dsh-job-stop"
config:
notice: wakeup # or: quiet
maxConsecutiveWakes: 3 # turns this plugin may open per agent between its own inputs
| Option | Default | Meaning |
|---|---|---|
notice | wakeup | wakeup opens a turn on an idle owning agent so it learns immediately (one model request, the same default dsh-tool-jobs ships for its own completion notices). quiet leaves the account pending until something else wakes the agent. A busy agent is injected either way. |
maxConsecutiveWakes | 3 | How many turns this plugin may open on one agent between that agent's own human inputs. Mirrors dsh-tool-jobs' bound of the same name and for the same reason: a burst of stops should not spend a model request each. Past the budget the account is injected instead, so the model still learns — it just is not woken for it. |
Every dependency is provided by the dsh installation this plugin runs inside, and is resolved from the host's module graph — the plugin pins itself to no harness build:
| At runtime | Used for |
|---|---|
ctx.jobs, ctx.agents | the registry and the owner fence |
ctx.connection.fetch | the authenticated route (@deepseek-ai/dsh-client-connection) |
@deepseek-ai/dsh-jobs, dsh-session, dsh-llm, schemastery | job/session ids, the identified notice message, config schema |
@deepseek-ai/dsh-client-ui-primitives, dsh-client-ui-slots, dsh-client-locale, dsh-client-ui-conversation, React | the list, the dialog, and the slot contract |
These are deliberately not declared as peerDependencies. Declaring them makes a standalone pnpm install try to fetch harness packages from npm, and at least one published range currently resolves to a build whose tree depends on @deepseek-ai/dsh-type-meta, which is not on the public registry. The plugin therefore installs with no network resolution of harness code at all.
pnpm install
pnpm build # host half -> lib/index.js, browser half -> lib/client.js
pnpm test # builds, then runs the host-half behavior suite (node --test)
Both halves are built by tsdown through build/tsdown.client.ts, which emits the browser bundle as a window.__ModuleLoader__.load({ id, factory }) closure whose id is this package's published name — the browser module system rejects a bundle that registers any other id.
npm version patch # or minor/major; commits and tags
git push --follow-tags
Then publish a GitHub Release for the tag. That one action attaches the prebuilt dsh-job-stop.tgz asset (which the tarball install route above points at) and, once NPM_PUBLISH_ENABLED is set, publishes to npm. The release notes are what the Plugin Market shows for an update, so write them for users rather than as a commit log. .github/workflows/release.yml documents the npm trusted-publisher setup that has to exist once.
The client half is not typechecked by the build (the published rc packages on npm predate the slot contract and the jobsBySession mirror this plugin targets). To typecheck against a real checkout instead:
DSH_CHECKOUT=/path/to/deepseek-harness node scripts/typecheck-harness.mjs
That project spans the plugin and the harness sources and fails only on errors inside this repo's src/; the harness's own guest diagnostics are reported as a note. It is worth running after touching the slot registration or the host route, because those are the two places where a stale published type would hide a real mistake.
running rows offer the stop affordance. A stopping job is already ending; offering to stop it again would be a second request for the same outcome.bash that is the complete command (args.command, untruncated), which is why the dialog is worth opening. A producer that publishes a truncated label gives the dialog only that.409). The kill is fenced by that exact owner instance, and guessing one would be an authorization hole.run_in_background card from the transcript; the card was never live-updating, and the list is where the outcome is legible.job-list slot entry, this plugin's entry stops replacing it and both lists would render. The fix is a one-line id change here.stopped by the user from the background-job list), kept stable so it greps in logs and in a job's terminal detail; it is not user-facing copy.MIT