# dsh-audio-rail

[中文](README.md) | **English**

A DeepSeek Harness (DSH) plugin that makes the quick-jump anchors (turn rail) on the right edge of the conversation page dance with whatever music your system is playing — like a spectrum analyzer rotated 90°.

## Demo

![Demo: the turn-rail anchors pulsing with the music](docs/demo1.gif)

GitHub strips hand-written `<video>` tags, so the demo ships as an auto-looping inline GIF; the full video with audio is at [demo1.mp4](docs/demo1.mp4).

## How it works

```
┌─ bin/AudioRailCapture.exe ───┐  stdout JSON   ┌─ index.js (host) ─┐   SSE    ┌─ client.js (browser) ──┐
│ WASAPI loopback capture of   │ ─────────────▶ │ /audio-rail/events │ ───────▶ │ one band per anchor;   │
│ the default render endpoint  │                │ /audio-rail/status │          │ ::before scaleX pulses │
│ FFT → 12 log bands @ 30 fps  │                └────────────────────┘          └────────────────────────┘
└──────────────────────────────┘
```

- **Capture**: `src/AudioRailCapture.cs` (compiled with csc.exe, single zero-dependency binary) captures the system mix from the default render endpoint in loopback mode (WASAPI shared mode + `AUDCLNT_STREAMFLAGS_LOOPBACK`), then reduces it via FFT to 12 log-spaced bands (45 Hz–16 kHz). Normalization is relative to a ~1 s sliding-average loudness reference (-9/+17 dB window): sustained content rests at midline, transients punch upward, quiet passages dip — no peak-follower saturation. Frames stream as JSON lines at ~30 fps and decay to zero in silence.
- **Host**: `index.js` registers `/audio-rail/events` (SSE) and `/audio-rail/status`. The capture process runs only while at least one page is subscribed, stops 15 s after the last client disconnects, and restarts with exponential backoff on crashes (exit code 3 = device hot-plug, recoverable).
- **Browser**: `client.js` injects one CSS rule turning each mark bar's (`.…_mark::before`) `transform` into `translateY(-50%) scaleX(var(--dsh-audio-rail-scale, 1))`; an EventSource feeds band frames into a requestAnimationFrame loop with attack/release smoothing per anchor. Bass sits on the bottom anchor. Honors `prefers-reduced-motion` (the plugin stays fully inert). The host's own active/preview/busy mark styles are untouched — the pulse is a pure transform overlay.

## Install

After cloning, build the capture helper first (`build.cmd` invokes the csc.exe that ships with .NET Framework on every Windows machine — nothing to install):

```
build.cmd
```

Then ask an agent in a DeepSeek Harness session to install the bundle, or manually:

```
dsh plugin --profile web add link:<path-to-this-clone>
```

The browser client hot-loads after installation — no page refresh needed.

## Notes

- On some Windows/runtime combinations `IMMDevice::Activate` returns `E_NOINTERFACE` for `IID_IAudioClient`; the helper therefore probes `IAudioClient3 → IAudioClient2 → IAudioClient1`, and every COM call below the device goes through raw vtable delegates (sidestepping a runtime marshaling bug that throws `InvalidCastException` on `Activate`).
- High-rate devices (e.g. 192 kHz) automatically use a larger FFT to keep the low bands resolved.
- Uninstalling or disabling the plugin closes the SSE stream and the capture process exits — nothing stays resident.
- Windows only (WASAPI is the Windows audio stack).

## Files

| File | Role |
| --- | --- |
| `index.js` | Host half: routes + capture-process lifecycle |
| `client.js` | Browser half: style injection + SSE + rAF animation |
| `src/AudioRailCapture.cs` | WASAPI loopback capture + FFT (C# 5, builds with csc.exe) |
| `build.cmd` | One-command helper build script (uses the system csc.exe) |
| `bin/AudioRailCapture.exe` | Helper build output (**not distributed** with the repo — generated by `build.cmd`, git-ignored) |
| `docs/demo1.gif` | Demo animation (inline in the README) |
| `docs/demo1.mp4` | Full demo video |
| `cordis.patch.yml` | Bundle patch inserting the `ui-dsh-audio-rail` row |

## License

[MIT](LICENSE)