@xain_npm/dsh-client-ui-beep
DeepSeek Harness 的网页代理心跳声音化：代理工作时发出平静的 lub-dub 嗡鸣，流式输出时发出滴答声，会话等待你输入时发出提示音
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add @xain_npm/dsh-client-ui-beep@0.6.2Client Ui Beep 以 @xain_npm/dsh-client-ui-beep 发布，当前版本为 0.6.2。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。
English | 中文
dsh-beep — an agent-heartbeat sonification plugin for the DeepSeek Harness Web surface. It plays three procedural Web-Audio tones as a subtle, non-intrusive heartbeat that tells you what the agents on this page are doing without watching the screen:
Requires DSH ≥ 0.1.7. That release replaced the
settingsScopeservice (and thesettings.yamldocument behind it) with profile-backed configuration forms, where a plugin's settings ARE its cordisConfig. Versions 0.5.0+ use that seam (configFormsin the browser, a.volatile()Config schema on the host); 0.4.x and earlier fail to activate on 0.1.7+.The requirement is declared in
peerDependencies— the DSH packages this plugin actually reads at runtime, pluscordis— so an incompatible harness shows up as a peer warning at install time instead of a plugin that silently never activates.Peer ranges are deliberately wide:
>=0.1.7-rc.2 <0.3(since 0.6.2). DSH 0.2.x added a compatibility gate that skips an entire bundle when any single@deepseek-ai/*peer is unsatisfied — the only symptom is a red "Problem" badge on the Plugins page, so a too-narrow range turns into a plugin that silently does not load. The previous^0.1.7-rc.2form was exactly that trap: caret does not cross a minor on0.x, so 0.2.x failed it. The upper bound stops at<0.3on purpose — a 0.3 bump is where the plugin should be re-verified rather than assumed compatible. (The browser module-table seedsdsh-client-store,dsh-client-ui-slotsanddsh-client-ui-primitivesare deliberately not declared: they are injected by the web app at runtime and do not resolve from a profile'snode_modules.)⚠️ Semver prerelease caveat.
^0.1.7-rc.2matches0.1.7-rc.xand the final0.1.7, but not a future0.1.8-rc.x: semver only lets a prerelease in a range match versions with the samemajor.minor.patchtuple. This range needs widening whenever DSH publishes a new prerelease tuple.
This package is a standalone, publishable fork of the
ui-beepplugin from the deepseek-harness repository (MIT). It builds with its owntsconfig.json+tsdown.config.ts(the in-repo build uses the sharedclientBundlepreset; this copy reproduces the same output format standalone).
npm install @xain_npm/dsh-client-ui-beep
# or: pnpm add / yarn add
Then mount it in your harness cordis.yml (or a cordis.patch.yml overlay) as a
web client row:
- id: ui-beep
name: '@xain_npm/dsh-client-ui-beep'
config:
enabled: true # per-browser default (each browser keeps its own values)
masterVolume: 0.4 # master gain 0…2 (100% = Web Audio full scale)
npm install
npm run build # tsc + tsdown → lib/
npm test # vitest (66 specs)
npm login # your npm account
npm version minor # bump (see Versioning below)
npm publish # prepublishOnly runs build + test first
| Voice | When | Sound |
|---|---|---|
| hum | Any session is busy (working) — "Deep diving…" (model request in flight), tool execution (running code, reading files), reasoning — and no session awaits your input and the current session is not streaming visible output. Repeats as a calm heartbeat every 4 s; pauses while output streams (ticks take over) and while an interaction is pending (the chime already alerted you) | Soft low "lub-dub" heartbeat (~118 Hz + ~92 Hz), ~500 ms |
| tick | The current session streams visible output | Short high 2 kHz pop (~60 ms) |
| chime | Any session begins awaiting your input (approval, plan review, or question), or the last busy session goes idle (work finished — your turn again). An unanswered interaction re-chimes after 10 s, then every 30 s, until answered | Two-tone 880 Hz + 1320 Hz bell (~500 ms) |
The mapping is an AgentPulse heritage: the agent working → a low heartbeat hum, output activity → a light tick, the agent waiting on you (or done) → a clear chime.
The browser half reads two React-free observable faces:
ctx.uiSession.sessionStatus — the unified per-session status surface DSH 0.1.7 introduced (it replaced the separate sessions.list + uiSession.pendingInteractions pair). It is a map of SessionId → { running, pendingInteraction }, so one subscription drives both the heartbeat and the chime.
running is the busy signal: it stays true for a whole turn, from prompt admission through tool execution and reasoning. While any session is running, the low hum repeats on a fixed interval (default 4 s) — first beat immediate when work begins (or on page load, if a session is already mid-work). The hum pauses only while the current session is actively streaming visible output (ticks take over — "streaming" means text growth within the last 1.5 s) and while any session holds a pending interaction (the chime already alerted you). It resumes ~1.5 s after output stops — even when the agent keeps working (a tool call after a message) — and when the interaction clears. It stops entirely when the last running session goes idle. This is level-based: it is the "something is working" indicator. The tone itself is a soft low lub-dub heartbeat (two gentle sine swells around 90–120 Hz with slow attacks) — reassuring, not urgent.pendingInteraction drives the chime: a 0→1 appearance for a session chimes. Decisions are edges, never levels — a session that stays pending does not re-chime on every refresh. An interaction that stays unanswered re-chimes after 10 s and then every 30 s until it is answered or the session disappears; an interaction already pending when the page loads starts the same reminder ladder (no immediate chime). The end-of-work case is the mirror edge: when the last running session goes idle, a chime tells you the agent is done.ctx.uiConversation.binding(id).snapshot, an ObservableSnapshot<ConversationSnapshot>): its notifier fires on every assembled frame; visible output text growth in the chat target's live partial fires the tick at the render cadence (the audio engine's debounce adds a hard floor). The current session id comes from the renderer scope adapter — ctx.uiSession.adapter.current.getSnapshot().key (DSH 0.1.7 moved it there; the list state no longer carries current).All tones are synthesized in code with linear fade in/out envelopes — no asset files, no clicks or pops on rapid state flips. Each voice is debounced to a 50 ms minimum interval.
Browsers block audio until a user gesture, so the engine arms on the first pointerdown/keydown anywhere on the page and is a silent no-op before that. It never throws or spams the console when audio is unavailable. A Preview click also arms the engine directly (the click is itself a gesture).
The row config: IS the plugin's settings document — DSH 0.1.7 stores a
plugin's settings as its own cordis Config, every field declared .volatile()
so a value written from the Settings page is adopted live (no restart). All fields
are optional:
- id: ui-beep
name: '@xain_npm/dsh-client-ui-beep'
config:
enabled: true # DEFAULT for a browser that has never chosen (see below)
masterVolume: 0.4 # per-browser default; master gain 0…2
tickVolume: 1 # per-browser default; streaming-output gain 0…2
humVolume: 1 # per-browser default; working heartbeat gain 0…2
chimeVolume: 1 # per-browser default; awaiting-input / done gain 0…2
tickPath: '' # optional absolute path to a custom audio file
humPath: '' # (omit or leave empty to use the built-in tone)
chimePath: ''
enabled and the four volume fields are not shared switches — they are
defaults for a browser that has never chosen. Once you touch the switch or a
slider in a given browser, that browser stores its own values in
localStorage and the Host value stops mattering for that field:
Settings page switch + sliders ┐
├─→ localStorage (per browser) → each device
composer speaker button ───┘ independent, surfaces never disagree
↑
Host config.enabled / *Volume = initial defaults only
custom audio PATHS stay on the Host (the files live on the server machine)
Why: the tone comes out of this browser, so this browser decides — and a
phone speaker is not a desktop speaker. It also means a change applies
immediately with no round trip to the Host, which is what made these controls
unreliable on a phone: the write carries an optimistic revision, a stale one is
rejected, and the old code discarded that failure silently (so the control looked
dead). Mute your phone and your desktop keeps chiming; both surfaces in one
browser always show the same state.
Each field is folded independently, so changing only a volume leaves the switch following the Host default. Choices persist across reloads and restarts. In a private window (or with storage blocked) they degrade to session-only rather than failing.
The Settings → 提示音 (Sound) page owns the same values: an enable switch, a master volume, and one volume per voice (streaming tick / working hum / awaiting-input chime), each 0–200 % with a preview button. 100 % is Web Audio's nominal full scale; the stretch past it is the user's own headroom — the plugin caps nothing, so anyone who raises a slider decides for themselves how loud the beeps are (values above full scale may clip). The defaults stay conservative so a first-time user is not startled.
Internal cadence is not configurable. The heartbeat period (4 s), the re-chime delays (10 s / 30 s) and the streaming-pause window (1.5 s) are defaults inside the watcher; they were row-config knobs before 0.5.0 and are deliberately no longer part of the Config (they are tuning, not user settings).
A mute toggle also sits in the composer's right tool row, beside the model selector: a speaker button that mutes/unmutes beeps in this browser. It reads and writes the same local document as the Settings page toggle, so the two can never disagree (speaker with an X when muted); muting stops a looping hum immediately, and unmuting while an agent is busy plays a beat at once (no waiting for the next heartbeat interval).
Each voice can play a user-supplied audio file instead of the built-in
tone. The Settings page shows a Choose audio button per voice; picking one
opens a whole-filesystem file browser (rooted at / or the drive root) that
lists ordinary user folders and audio files (.mp3, .wav, .ogg, .flac,
.m4a, .aac, .opus, .webm) — hidden (dotfile) entries and system
directories are skipped. The chosen absolute path is stored in the plugin
Config (a plain string field) — the file is never uploaded or copied, and it can
be moved/replaced on disk freely. Playback semantics:
The Host half serves the file through two loopback, browser-authenticated
routes (GET /ui-beep/audio/:voice, GET /ui-beep/browse) — the path is read
from the plugin Config, never from the request URL, so the routes cannot be
pointed at arbitrary files. The browser fetches and decodes each file once, then
caches the decoded buffer.
localStorage), so each
device controls its own beeps and the Settings page toggle and composer speaker
can never disagree. The Host enabled field became only the default for a
browser that has never chosen. Also: previewing auditions while muted.ctx.configForms; the watcher moved to
uiSession.sessionStatus.None. The package is a browser-side read-only sonification of already-logged session facts (running/busy, streaming output, pending interactions); it plays audio and registers nothing model-facing. The model's own view of its work stays with the tools and host services that produce those facts.
None; the package never assembles or sends provider requests.