dsh-hardware-buddy
A self-contained deepseek-harness host (Node.js) plugin that mirrors agent state onto a CodeBuddy / StickS3 hardware pet over USB CDC, and routes the device's physical A/B approval buttons back into the dsh approval/request waterfall.
- 📟 State mirror —
agent/status, session/*, session/event are aggregated into a compact JSON heartbeat (hard 900-byte cap) pushed to the device every 3s (not per-event).
- ✅ Physical approval — dangerous tools prompt on the device; A = allow-once, B = reject. Excluded tools (e.g.
MCP__danger_*) fall through to the dsh Web UI instead.
- 🎙️ Voice feedback — device speaks the cheerleader clips (approve / deny / error / idle / boot) when approvals resolve.
- 🔌 Plug & play — auto-discovers the StickS3 by USB VID
0x303A (optionally PID), with a cu.usbmodem* / ttyACM* name fallback. Tolerates unplug/replug (5s reconnect poll) and never crashes the plugin when the device is absent.
This plugin only runs on the host (Node.js) side. It deliberately does not declare dsh.client and must not be bundled into the browser client (it depends on the native serialport module).
Requirements
- Host: macOS or Linux, Node.js
^22.19.0 || >=24.0.0, dsh CLI (@deepseek-ai/dsh). Windows is not supported.
- Device: an M5Stack StickS3 flashed with the DSH firmware (BLE compiled out + USB CDC single channel) from the DeepseekHarnessBuddy repo (
firmware/). Without the DSH firmware build the USB protocol does not match.
Install (from GitHub Releases)
dsh plugin --profile <name> add \
https://github.com/baidang201/DeepseekHarnessBuddy/releases/download/v0.1.0/dsh-hardware-buddy-dsh-hardware-buddy-0.1.0.tgz
The package ships a dsh.bundle manifest (cordis.patch.yml) that dsh applies automatically. Pin to a specific release version by changing v0.1.0 in the URL (a new version bumps the tarball name too).
If you prefer to manage the entry by hand, add it to your profile patch (e.g. $DSH_HOME/profiles/<name>/cordis.patch.yml — this is also the file dsh's HMR watches):
- insert:
- id: hardware-buddy
name: '@dsh-hardware-buddy/dsh-hardware-buddy'
config:
port: null # null = auto-discover
vendorId: '0x303A'
productId: null
approvalTimeout: 30000 # ms before an unanswered prompt auto-cancels
heartbeatIntervalMs: 3000 # full-snapshot interval (<< device 30s window)
dangerousTools: # regex list routed to the hardware screen
- '^bash$' # defaults finalized from the real 25-tool dump
- '^write$' # (all lowercase — see probe-report.md)
- '^edit$'
- '^str_replace_editor$'
excludedTools:
- '^MCP__danger_.*' # regex list routed to the Web UI approval
celebrateThreshold: 50000
entriesLimit: 5
logLevel: info
Config changes in cordis.patch.yml hot-reload via dsh's HMR — no restart needed.
Usage
- Plug the StickS3 into USB. On first boot after a cold start, pick the device up once (see "Screen stays frozen" below) so the display activates.
- Run dsh (web or headless). The plugin auto-connects and the pet shows the agent state.
- When a dangerous tool needs approval: the device beeps, shows the tool + args on the approval screen, and waits.
- A = allow once · B = reject · 30s no response = auto-reject (safe default).
- Approval results are relayed back into dsh's
approval/request waterfall; the model continues or stops accordingly.
Screen stays frozen after boot? (not a bug)
After a cold boot, if the device has never been picked up, the firmware waits for an unambiguous orientation before it starts rendering (an upstream CodeBuddy anti-speculation design): a device lying flat gives ambiguous IMU readings, so the screen keeps the boot frame while beeps and buttons keep working. Pick the device up once (hold it upright or sideways) — the orientation is then remembered for the rest of the boot, and the screen renders normally even when you put it back down flat.
Safety model (important)
The plugin follows the Cordis waterfall rule strictly: not calling next() vetoes the entire chain, including the Web UI. Therefore the plugin only takes over an approval when (a) the tool is not on the excludedTools list and (b) the device is currently online. In every other case — device offline, excluded tool, a missing agent/callId on the exec, or any internal error — it calls next(), delegating to the Web UI / default policy.
The plugin never produces an unavailable outcome and never denies a tool for its own reasons. A timeout or a runtime abort resolves to cancelled (which dsh maps to a deny with a "cancelled" reason) — that is dsh's policy, not the plugin refusing.
Troubleshooting
| Symptom | Fix |
|---|
No StickS3 CDC port found | Replug USB; on Linux run sudo usermod -aG dialout $USER and re-login; check pio device list |
| Device never wakes / no heartbeat | Confirm the DSH firmware (not the stock CodeBuddy BLE build) is flashed |
| Want to see what the plugin is doing | Start dsh with HB_DEBUG=1 — prints serial discovery, heartbeats, approvals on stderr |
| Approval screen disappears in 3s | Update the plugin (old builds cleared the prompt on the next heartbeat) |
Development
npm install
npm test # vitest unit tests (protocol / bridge / approval)
npm run lint # oxlint
npx tsc -p tsconfig.json --noEmit # type check
npm run build # emit dist/
The package is fully self-contained: tsconfig.json does not extend any external config and npm install works without a workspace/pnpm setup. Unit tests do not require a real dsh runtime or a connected device.
Probe plugin
probe.ts is a standalone companion entry (run inside a real dsh profile) that dumps the live event payloads, the real tool-name list (so dangerousTools defaults can be finalized), and the measured VID/PID. It is excluded from the build and from oxlint src/.
