dsh-captain-call

When you run multi-agent workloads in AgentTeams, the tedious part is rarely the tasks themselves — it is figuring out who finished, whether they hit the deadline, and what the quality looks like, all buried in a status panel you have to refresh by hand. dsh-captain-call turns that chore into a phone call: the moment a teammate completes a task, the desktop companion GINKA fires up a full-screen WeChat-style incoming-call UI with a blurred avatar backdrop, a ringing tone, and a green answer button. Tap to connect and a Chinese voice report reads out exactly who it is, which task wrapped up, whether it was on time, and how many minutes it took.

The speech stack is fully open-source and local-first. Kokoro-82M-zh (82 M parameters, ONNX inference, 100+ Chinese male/female voices, works offline) is the default engine, with automatic fallback to Microsoft neural TTS and then the browser system voice, so audio always comes through. On the input side, native browser SpeechRecognition powers the tap-to-talk reply button; the transcript is merged with the teammate persona and real task state for an instant response, and unsupported browsers simply degrade to typed chat without breaking the flow.

It is not one-way. The GINKA widget in the bottom-right corner opens into a full contact list where the captain can dial any teammate on demand, following a call-progress handshake. Call history, mute toggle, and a deduplication baseline are built in, so stale notifications never replay after a restart, and uninstalling restores every DOM node, CSS rule, timer, and audio stream cleanly.

Ideal for anyone working inside DeepSeek Harness or AgentTeams who likes WeChat-style voice-call interactions, wants an offline-capable open-source TTS pipeline, and enjoys turning task delivery into a small, audible, on-screen ritual. Personal local use is fully supported; avatar and ringtone assets belong to their respective owners and should be replaced before any public distribution.