dsh-activity-line

English · 中文

Source of truth: this plugin lives in NeoWangKing/neo-dsh. is a mirror published for discovery and for dsh plugin add github:…; issues and PRs belong on the monorepo. scripts/mirror-plugin.sh keeps the mirror in step.

A DeepSeek Harness web plugin: one line under the composer that says what the running turn is actually doing.

Install (into your own dsh)

# 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:/

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:

desktop app: the ⟳ button (or restart the app)
CLI: restart the dsh web process

Uninstall:

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.

Why it exists

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:

the model is thinking;
a bash command has been running for three minutes;
an approval is waiting for you;
a failed request is being retried.

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.

States (first match wins, top to bottom)

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.

Data sources (subscriptions from the standard session kit, never DOM scraping)

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.

Tests

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.

Files

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

Install

README