One-Line Positioning

Replaces the official effort selector in DeepSeek Harness's model menu with a Claude Code-style infinite-drag Effort slider panel; also enables any third-party custom models to have truly effective thinking intensity levels.

Core Capabilities

• Intercept the "Reasoning Level" row in the official model menu and pop up a Claude Code-style Effort slider panel (infinite drag, snap on release, WebGL flame follow)
• Automatically inject generic 5-level scale (OFF / Low / Medium / High / Ultracode) for third-party models that haven't declared reasoning metadata, making the official level list and this panel selectable and request validation passing
• Automatically complement line-level dictionary and protocol dialect (compat) for pi-ai custom models, translating levels to real online fields like reasoning_effort / thinking, with changes taking effect hot without restart
• After panel closes, the level value in the menu row and the "Model · Level" entry button above the input box continue to display the current level's corresponding color and glow
• Respect user's existing reasoningEfforts declarations (set to false or custom dictionary), never override
• Provide 8 kinds of line dialects (OpenAI effort / DeepSeek / OpenRouter / Together / Z.AI / Qwen / string-thinking / ant-ling), and support overriding default dialect by route

Technical Implementation

• Language: TypeScript + React 18 (browser side uses JSX/TSX), host side pure TS
• Key Dependencies: @deepseek-ai/cordis (host context), @deepseek-ai/dsh-settings (settings segment registration), @deepseek-ai/schemastery (configuration Schema validation), react-dom/client (panel rendering)
• Architecture Pattern: DSH dual-half external plugin — host side lib/index.js listens to llm/adapters-updated / settings/updated events for metadata wrapping and pi-ai line-level supply; browser side lib/client.js performs click interception on the "Reasoning Level" row in model menu and renders the panel with React 18; styles are injected by lightningcss inlining effort.module.css hash as <style data-plugin>
• Entry Files: src/index.ts (host) / src/client/index.ts (browser)

Use Cases

• Configured multiple third-party provider models in DSH, but the official level selector doesn't work or won't take effect on most models
• Accustomed to Claude Code's infinite drag panel with flame effects and want the same interaction to adjust levels in DSH
• Want reasoning_effort field to actually reach the request body under different OpenAI-compatible endpoints (different line dialects)

Prerequisites & Compatibility

Installation

dsh plugin --profile web add dsh-effort-slider

Configuration Options

Configuration is written to the effort-slider segment under DSH's ~/.dsh/settings.yaml and can be applied hot:

FAQ

Q: Do I need to restart DSH after installation?

A: Yes. Host-side patch must be injected at restart; after installation, restart DeepSeek Harness once, and press Ctrl+F5 in the web page to reload the browser-side script.

Q: What if I don't see the "Reasoning Level" row in the model menu?

A: If the model itself has no reasoning metadata and the host-side metadata supply didn't take effect, the "Reasoning Level" row won't appear. First confirm DSH has been restarted, and check if effort-slider.enabled is set to true in ~/.dsh/settings.yaml; if still ineffective, check the DevTools console for [effort-slider] intercept row: ... to determine if the interception layer is working.

Q: If I only write reasoningEffort in the panel, will it change my selected model?

A: No. The panel only calls selectModel({ reasoningEffort }) to write the level, keeping the model selection unchanged; models that don't support multi-level reasoning will display "Current model does not provide multi-level reasoning levels".

Q: Can third-party models really make thinking intensity take effect?

A: Yes. The host side does two things simultaneously: injects generic 5-level metadata for adapters that haven't declared reasoning; complements line-level dictionary and compat dialect for pi-ai custom models, which takes effect hot, then pi-ai translates the levels to real online fields like reasoning_effort / thinking / OpenRouter reasoning.effort.

Q: Can I configure different line dialects for different providers?

A: Yes. Specify the dialect for that route in effort-slider.routes.<route> (such as deepseek / openrouter / qwen etc.); when not written, it falls back to defaultDialect.

Q: How do I uninstall?

A: Execute dsh plugin --profile web remove dsh-effort-slider, and if you previously manually modified cordis.patch.yml, also delete the ui-effort-slider line, then restart DSH.

Q: Does this plugin read my conversation content or make network requests?

A: No. Host side doesn't read any user files, only writes its own settings segment; browser side only interacts with the local DSH's /api RPC endpoint, never reads credentials or conversation history.

Difficulty Level

Beginner — only one installation command + one DSH restart needed; if the default 5 levels work, no configuration needed at all; only need to configure defaultDialect by route when specific OpenAI-compatible endpoint fields are incorrect.

Known Issues & Limitations

• When the current model itself doesn't provide multi-level reasoning, the panel will display "Current model does not provide multi-level reasoning levels", cannot adjust (README.md:60)
• When the host's llm or settings service is not ready, the entire supply will be turned off and log effort-slider: llm/settings service unavailable, at which point the panel will get empty directory/fallback scale (src/index.ts:76-79)
• May conflict with skin plugins that also intercept the "Reasoning Level" menu (like dsh-ui-web's aurora skin), potentially showing two panels simultaneously; need to disable its interception in the other skin to avoid conflicts (README.md:165)
• pi-ai settings segment registration happens after its adapter, and the registration process doesn't trigger settings/updated event; host side uses up to 30 seconds of ready retry as fallback; very short cold starts before this time may need to wait a few seconds (src/index.ts:82-120)

Click the Effort row (second row of the official model menu) instead of the built-in level list: an Effort slider panel pops up — OFF/MAX scale ticks, Low/Medium/High/Ultracode status shown live, and after the panel closes the effort value stays color-coded on the menu row and on the model-seat trigger button.

Features · Compatibility · Install / Uninstall · Quick Start · Configuration · Permissions & Data · How It Works · Troubleshooting · Development

Demo screenshots (menu-row coloring / model-seat trigger / Effort panel — effort level is color-coded):

---

🔧 v0.2.5 — Fixed the real landing bug in universal effort provisioning: explicit models arrays are now replaced as a whole (dsh-settings path ops cannot traverse array nodes, otherwise models is corrupted, the schema rejects the write, and provisioning fails silently).

🔧 v0.2.4 — Fixed universal effort provisioning never landing when the pi-ai settings section registers late: provisioning now retries until the section is ready (registration happens after the adapter and does not emit settings/updated).

🆕 v0.2.3 — The model-seat trigger button (「model · effort」 above the composer) also color-codes its effort name, persistent after the menu closes.

🔧 v0.2.2 — Fixed menu-row coloring breaking on whole-tree menu remounts: painting now does a full scan and re-paints on any DOM change (throttled), and infers the color from the effort text when the panel has not reported one yet.

🆕 v0.2.1 — After the panel closes, the effort value on the model menu's Effort row keeps its level color (OFF rose-gray / Low amber / Medium blue / High purple / Ultracode bright purple with glow).

✨ v0.2.0 — Any custom third-party model/provider gets working thinking-effort control (adapter metadata provisioning + pi-ai wire-level provisioning, hot-applied).

🎛️ v0.1.0 — Initial release: intercepts the official Effort menu and shows a Claude Code–style Effort slider panel.

---

Features

Only reasoningEffort is written — model selection is untouched. Models with no multi-level reasoning show 「当前模型不提供多档推理等级」 (no multi-level reasoning available).

Universal effort provisioning (v0.2.0)

Any custom third-party model/provider gets thinking-effort control that actually works on the wire:

• Adapter metadata provisioning (host): models without reasoning metadata get the universal 5-level scale injected (off/low/medium/high/max → OFF/Low/Medium/High/Ultracode), so the official model menu and this panel are selectable and request validation passes;
• Wire-level provisioning (host, pi-ai): for llm-pi-ai models missing reasoningEfforts, the dictionary and compat wire dialect are patched in (hot-applied, no restart); pi-ai translates the level into real wire fields (reasoning_effort / thinking / OpenRouter reasoning.effort, etc.);
• Client fallback scale: if the directory returns no reasoning metadata, the panel still opens with the universal 5-level scale;
• Existing user declarations (reasoningEfforts: false or a custom dictionary) are always respected and never overwritten.

Supported wire dialects for custom endpoints (set effort-slider.defaultDialect or routes.<route>):

Compatibility

Install / Uninstall

# Install (recommended, same bundle mechanism as dsh-navbar)
dsh plugin --profile web add github:2768651338/dsh-effort-slider#main

# Build locally and install (clone this repository)
pnpm build
dsh plugin --profile web add file:./dsh-effort-slider

After installing, restart DeepSeek Harness and press Ctrl+F5 once in the web page. The lib/ artifacts are committed, so GitHub installs need no local build.

Quick Start

1. Install the plugin, restart DSH, press Ctrl+F5.
2. Open the model menu above the composer → click Effort → the Effort slider panel appears.
3. Drag to choose a level; it snaps on release. After the panel closes, the effort value keeps its color on the menu row.
4. For custom models with no thinking effort, the panel opens with the universal 5-level scale; pi-ai models get their wire dictionaries patched on the host side (hot-applied).

A [effort-slider] intercept row: ... line in DevTools Console means interception is working.

Configuration

effort-slider:
  enabled: true          # provisioning master switch
  defaultDialect: effort # global default wire dialect
  routes:
    my-gateway: deepseek # per-route override

Permissions & Data

How It Works

The browser half follows the official external-plugin convention: classic script + window.__ModuleLoader__.load factory; react / react-dom/client / react/jsx-runtime are platform externals; effort.module.css is hashed and inlined by lightningcss, injected as <style data-plugin> when the factory runs.

Troubleshooting

Project Structure

src/
  index.ts                  host half: universal effort provisioning (adapter metadata + pi-ai wire + settings section)
  effort-core.ts             pure logic: dialect → wire mapping, provisioning patches (unit-tested)
  client/
    index.ts                browser half: intercept the Effort row + panel anchor + menu-row coloring
    css-modules.d.ts
    effort/
      EffortPanel.tsx       Effort panel (levels / ticks / slider / glow)
      useWebglFire.ts       WebGL2 three-pass fire loop (spring follow + idle sleep)
      shaders.ts             vertex / ignite / blur / composite shaders
      effortColors.ts        level → color mapping (panel and menu row share it)
      effort.module.css      panel styles (inlined via lightningcss)
cordis.patch.yml           bundle patch (inserts the ui-effort-slider row)
lib/                        built artifacts (client.js ships with sourcemap)
test/                      host.spec.mjs unit tests + host-apply.spec.mjs + client.smoke.mjs

Development

pnpm install
pnpm build   # tsdown → lib/index.js (host half) + lib/client.js (browser half)
pnpm test    # host unit tests + apply integration test + jsdom smoke (intercept/render/snap/fallback/coloring/teardown)

Contributing. Fork → change → pnpm build → run pnpm test → open a PR against main. Small fixes (docs, tests) are welcome without prior discussion; report issues with the DSH version and the exact error.

License & Security

License: BSD-3-Clause — see LICENSE. The UI implementation references the aurora skin of the community dsh-ui-web project (BSD-3-Clause); full upstream notices and license texts are in THIRD_PARTY_NOTICES.md.

Security: this plugin reads no credentials and sends nothing over the network (only talks to the local DSH). To report a security issue privately, use GitHub's Report a vulnerability on the Security tab — do not open a public issue with exploit details.

---