dsh-plugin-liang-calibrator

滑动变祖器 — the liang-intensity-calibrator as the DeepSeek Harness model + thinking-effort selector.

Clicking the composer's model seat opens the 31-level calibrator directly. The six
stages map 1:1 onto the DeepSeek model catalog's combinations (2 models × 3
thinking levels = 6 = the six stages):

With any other catalog the slider splits the 0–30 track proportionally across
every provider/model/effort combination, and the stage follows the selected
combo. A small Model row beneath the slider still drills into the plain
model list.

Install

dsh plugin --profile web add dsh-plugin-liang-calibrator

then register the entry in your profile's cordis.patch.yml
($DSH_HOME/profiles/web/cordis.patch.yml):

- insert:
    - id: liang-calibrator
      name: dsh-plugin-liang-calibrator

and restart dsh web.

How it works

• Host half (lib/index.js): serves the 31 portrait keyframes under
  /liang-assets/frames/ through the profile's webServer service — no CDN,
  no patched static server.
• Browser half (lib/client.js): a standard dsh.client bundle that
  registers the conversation.input.model slot at priority −1. Slot
  rendering is priority-based ("lowest renders"), so the calibrator shadows the
  stock selector without touching it — uninstalling the plugin restores the
  original UI.
• The portrait is drawn from per-level keyframes rather than a scrubbed video:
  the stock static server has no HTTP Range support, so media-element seeking
  silently fails; image keyframes work everywhere.
• Requires @deepseek-ai/dsh-client-ui-model-selection (ships with the default
  web profile) for the shared model directory service.

Uninstall

dsh plugin --profile web remove dsh-plugin-liang-calibrator

(and drop the entry from cordis.patch.yml).

Development

Regenerate the browser bundle from an installed DSH checkout:

python3 scripts/assemble-client.py /path/to/node_modules/@deepseek-ai [region.js]

Portraits

lib/assets/frames/frame-00.webp … frame-30.webp derive from the
liang-intensity-calibrator
project's public/frames. Reuse or redistribution requires confirming you hold
the relevant portrait and asset rights (see that project's README).

License

MIT. The calibrator concept and portrait assets belong to their original
authors — see above.