dsh-activity-line
DeepSeek Harness 网页插件：编辑框下方显示实时活动状态，说明当前轮次实际在做什么（工具运行中 / 等待你的操作 / 正在重试 / 模型思考中）。
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:NeoWangKing/dsh-activity-line#853000db17512ddfdee34abb0b4a6aaa7961862eActivity Line 以 dsh-activity-line 发布，当前版本为 1.0.0。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。
English · 中文
Source of truth: this plugin lives in NeoWangKing/neo-dsh. https://github.com/NeoWangKing/dsh-activity-line is a mirror published for discovery and for
dsh plugin add github:…; issues and PRs belong on the monorepo.scripts/mirror-plugin.shkeeps the mirror in step.
A DeepSeek Harness web plugin: one line under the composer that says what the running turn is actually doing.
# from npm
dsh plugin --profile web add dsh-activity-line
# from a local checkout
dsh plugin --profile web add link:/path/to/activity-line
# from git
dsh plugin --profile web add github:<user>/<repo>
The plugin is installed into the profile, not an agent preset, so a preset manager cannot overwrite it.
Restart the host afterwards — the host process caches the assembled plugin bundles, so reloading the page is not enough:
dsh web processUninstall:
dsh plugin --profile web remove dsh-activity-line
Requirements: dsh ≥ 0.1.5 on the web profile (it needs the
conversation.composer.dock slot and the useSession / useChat /
useProjection session kit). The plugin has no runtime dependencies of its own
and writes no configuration; all you get is the line.
The shipped turn status is the static label "Deep diving..." hardcoded in
ChatView, with a clock that only appears after 15 seconds. So these look
identical:
bash command has been running for three minutes;Measured in real sessions on this machine: single steps of 193 s / 203 s were each
one bash call; one 418.6 s step produced no output at all and was cancelled
by hand; some sessions carry five llm/retry events. This line names the phase.
| Shown | Meaning | Source |
|---|---|---|
♻️ 请求重试中（第 N 次） | a retry is in flight — the honest explanation for a long silence | useProjection('llmRetry') |
🛠 bash: pnpm run build 运行中 1m23s | a tool is running (oldest call wins; 另有 N 个 when several run, and the clock is that tool's own) | chat snapshot legacy.runningCalls |
✍️ 输出中 42s | the assistant is streaming | chat snapshot legacy.partial |
🧠 等待模型响应 42s | the turn runs but nothing has arrived yet | legacy.turnTimings + a 1 s clock |
| (nothing) | no turn is running | — |
The texts are Chinese today; the line registers a locale namespace, so the labels follow the app's language as the dictionary grows.
The slot hands over point-in-time props, so every value has to be subscribed:
| Source | What it gives |
|---|---|
useSession | session status snapshot: running, sessionId |
useChat | the chat snapshot's legacy slice: runningCalls (name, raw arguments, logged time), partial, turnTimings; a build without useChat falls back to useConversation(c => c.views.get('chat')) |
useProjection('llmRetry') | the retry projection { [provider|policy]: { retry, retryId } }, cleared on every step/start and turn/end; without that projection it falls back to sessions.binding(id).session.projections.faceOf('llmRetry') |
Placement: conversation.composer.dock — the band under the composer card, the
same list slot the shipped "N turns M steps · tok/s" stats line occupies. This is
a second entry (order: 5); nothing is replaced.
node test/activity.test.mjs
Covers phase priority (retry > tool > stream > wait), multi-tool selection and counting, the tool's own clock, the degraded case with no chat snapshot, duration formatting, argument previews (first line / collapsed whitespace / truncation / non-JSON / arrays), the projection face shapes, and both turn-anchor sources.
| File | Role |
|---|---|
index.js | host half: a no-op placeholder (this is a client-only plugin) |
client.js | client half: reads the snapshots and the projection, renders the line |
cordis.patch.yml | bundle patch: inserts this plugin's row |
test/activity.test.mjs | unit tests for the pure helpers above |