dsh-record

dsh-record

A DeepSeek Harness (dsh) plugin that puts a microphone button in the composer: click to record, click again to stop, and the recording is attached to your current prompt as a conversation file attachment. The output format — MP3 (default), WAV, or WebM — is chosen on the plugin's settings page.

Chinese

Features

• Composer microphone button — registered at the conversation.input.right slot, before the submit action. No floating widgets; coexists with other composer controls.
• Format chosen on the settings page — the plugin declares a Loader Config with a single format field (mp3 / wav / webm, default mp3). Pick it in Settings → Plugins → dsh-record; the browser half reads the saved choice over GET /dsh-record/config at startup (falling back to mp3 if the route is unreachable).
• Language-aware tooltip — hovering the mic shows 录音 on a Chinese interface and Record on an English one. The button registers zh/en dictionaries with the host locale service, so the tooltip, aria-label and status notices follow the interface language and update when you switch it.
• Real-time recording — browser MediaRecorder, webm/opus preferred with capability-based fallback. An elapsed timer and a live file-name preview (rec-YYYYMMDD-HHMMSS-<n>s.mp3) show while recording.
• Three output formats (transcoded in the browser after you press stop):
  • MP3 — decoded from the native container and encoded via lamejs (128 kbps CBR). No server-side dependency.
  • WAV — lossless 16-bit PCM, decoded and re-wrapped.
  • WebM — the browser's native Opus container, no transcoding; smallest file, fastest attach.
• Prompt attachment — the file is registered through the host draft-attachment channel (conversation.createDrafts + inputActions.addAttachments) and shows as a composer file card with upload progress, riding the normal upload queue with your prompt.
• Graceful fallback — if MP3/WAV transcoding fails (e.g. OfflineAudioContext unavailable), the raw WebM is attached instead, so a recording is never lost.
• Cancel discipline — Escape, the ✕ button, or leaving the tab discards the recording; recordings under 0.5 s are dropped; a stop during an in-flight submit releases the draft cleanly.

The plugin records, optionally transcodes, and attaches the file to your prompt. It also writes recordings to the configured directory when a save folder is set. It performs no speech recognition — pair it with a speech-to-text plugin such as dsh-stt.

Install

From the marketplace (recommended)

Install it from the plugin marketplace. The listing points at the prebuilt tarball attached to the latest release, so nothing is compiled on your machine and no build toolchain is needed.

From source

The lib/ build output is not committed to this repository (it is gitignored). A fresh clone therefore has no main entry, and dsh plugin add file:<clone> fails with a missing lib/index.js. Build before installing:

git clone https://github.com/most-silence/dsh-record
cd dsh-record
npm install
npm run build          # required: produces lib/index.js + lib/client.js
dsh plugin --profile <your-profile> add file:$PWD

Prefer to skip the build? Download dsh-record.tgz from the latest release and install that instead — it already contains lib/:

dsh plugin --profile <your-profile> add file:dsh-record.tgz

Restart the dsh host after installing, then refresh the web page.

Configuration

Settings → Plugins → dsh-record (open the plugin's own detail page) exposes:

Field	Values	Meaning
format	mp3 (default) · wav · webm	Output container. MP3 and WAV are transcoded in the browser when you stop; WebM is attached as captured. Takes effect on the next page load.
dirMode	workspace (default) · custom	Where recordings are written: the current workspace root, or a fixed folder.
customDir	absolute path	The folder used when dirMode is custom. Recordings are written here through a same-origin-guarded host route (file name sanitized, .. traversal rejected, 64 MB cap).

Changing format or the directory takes effect on the next page load, since the browser half reads the configuration once at startup.

Architecture

browser half (lib/client.js)                    host half (lib/index.js)
├─ mic button @ conversation.input.right        ├─ Config schema (volatile)
├─ settings card @ plugins.bundle.config        │    { format, dirMode, customDir }
├─ zh/en dictionaries → locale service          ├─ GET  /dsh-record/config → serves config
├─ MediaRecorder → webm/opus                    └─ POST /dsh-record/save   → writes the file
├─ decode → lamejs (mp3) / PCM wrap (wav)
└─ attach via createDrafts + addAttachments

Three hardening rules keep the plugin from ever blocking profile startup — all learned from real crashes:

1. @breezystack/lamejs is bundled inline. The DSH module table is frozen and does not contain lamejs, so leaving it as a build-time external makes require("@breezystack/lamejs") miss the table and abort the whole web boot. tsdown forces it inline (noExternal).
2. The host Config schema uses z.union([...]), never z.enum([...]), and schemastery stays external. The host's schemastery version has no enum method, and a schema built from a bundled private copy is incompatible with the host validator — either one throws during import and fails the dsh-record row as required, taking down the profile.
3. The schema is marked .volatile(). The host's settings projection only surfaces volatile fields; a plain schema renders no form at all, which is why an earlier configuration page appeared empty.

Development

npm install
npm run build     # tsdown: lib/index.js (ESM) + lib/client.js (CJS, ModuleLoader-wrapped)
npm test          # 21 node:test cases: factory (9) + encoders (5) + settings (7)

Tests run with Node 24 (native TypeScript type stripping); no browser needed.

Known limitations

• The attachment channel relies on host client contracts verified on dsh 0.2.1-rc.2 (conversation.input.right slot, createDrafts / addAttachments). Upstream changes may degrade to an inline error notice.
• MP3/WAV transcoding runs after you press stop (not during capture), so very long recordings add a brief encode pause.
• Recording requires a microphone-permission grant; the page must be served over HTTPS or localhost.
• The output format and directory live in the host profile config, not the browser — they are read once per page load, so a change applies after a refresh.
• The custom directory is typed by hand; an in-app folder-picker button is planned for the next minor version.

License

MIT