dsh-reasoning-effort

Bring a Codex-style "model + reasoning effort" control directly into DeepSeek Harness.

English · Latest release · Report an issue

Switch models, drag the effort slider, and watch the eight-frame "big fat fish" speed up below the DSH input box. Levels come from the current model and stay in sync with the /model command.

• Follows model levels: auto-adapts level count, names, and order; rolls back on submission failure.
• Fits the DSH UI: supports light/dark themes; Simplified Chinese and English switch instantly with DSH language.
• Optional dynamic look: big fat fish enabled by default; plain button and system "reduce motion" setting supported.
• Custom model guidance: copy-pasteable config snippets, plus one-click full briefing for an Agent to investigate and fill in.

本次更新 · 安装与更新 · 版本支持 · 外观设置 · 常见问题

v0.7.2 Release Notes

• Added Copy to Agent: one-click compiles the current model's config location, level rules, and suggested snippets.
• Fixed communication issue in the custom model guidance panel; shows a prompt when diagnostics fail.
• Fixed old briefing being copied on model switch, and Chinese names getting mangled across HTTP chunks.
• Improved menu positioning and scrolling on narrow/short viewports.

The full record is in the v0.7.2 release notes and the CHANGELOG.

Version Support Policy

This plugin is only adapted, tested, and bug-fixed for relatively stable DSH RC releases. Alpha versions are not separately maintained — their client APIs, dependency structures, and plugin loading mechanisms change frequently; tracking multiple transitional versions raises maintenance cost and undermines reliability.

The current release is plugin v0.7.2, which the maintainer has verified hands-on in DSH 0.1.5-rc.2 (Web Profile) with no compatibility issues found. If you need to keep using an alpha, do your own temporary adaptations. An RC is still a release candidate, and this does not mean every past or future RC is automatically compatible.

Install & Update

1. Install a pinned release

Run in the terminal you use to start DSH:

dsh plugin --profile web add github:HanaAyane/dsh-reasoning-effort#v0.7.2
dsh --profile web --dump-config

Make sure the output contains name: dsh-reasoning-effort. Existing installations are updated with the same add command. For development, you can replace #v0.7.2 with #main, but the main branch may contain unreleased changes.

请为 DeepSeek Harness 的 web Profile 安装 dsh-reasoning-effort v0.7.2。
只执行下面两条命令，不要修改其他 Profile：
dsh plugin --profile web add github:HanaAyane/dsh-reasoning-effort#v0.7.2
dsh --profile web --dump-config
确认配置中出现 dsh-reasoning-effort 后告诉我结果。
不要关闭或重启正在运行的 DSH；提醒我手动重启 Web Host 并刷新页面。

2. Restart and refresh

The plugin is loaded when the Web Host starts. After installing, manually restart the DSH Web Host, then refresh the page.

3. Choose a model and effort

Open a session, click the model entry under the input. Drag or click the track; it snaps to the nearest level. Click the model row to expand the list.

Where levels come from

The slider reads the current model's reasoning.efforts from the DSH catalog. Level count, names, and order are set by the model and route — not a fixed three — and different endpoints aren't guaranteed to match.

Slider shows when a model publishes ≥2 levels; otherwise a hint appears. The plugin submits catalog level values; DSH validates and sends them, never bypassing model or deployment limits.

Declaring levels for any custom model

This section is vendor-agnostic and applies to any model you declare yourself in llm-pi-ai.

Why levels can't be read: the DSH model catalog only reports capabilities declared by the adapter. A model you declared yourself has no catalog entry, so unless you write out reasoningEfforts, the catalog will never have levels for it, and the slider will not appear.

How to fill it in: find the entry for that model in settings.yaml and add a reasoningEfforts — the keys are DSH levels, and the values are the notation the endpoint itself accepts; any level you don't list is treated as unsupported:

      models:
        - id: <你的模型 id>
          reasoningEfforts:
            low: "<端点接受的取值>"
            high: "<端点接受的取值>"

When you also need compat (a sibling of reasoningEfforts, only written when the endpoint requires it):

Without compat, the adapter auto-detects from the endpoint address: unknown addresses get standard OpenAI treatment, known vendor endpoints get their format applied automatically — so a wrong guess is worse than leaving it out.

To verify: save, open the model menu — slider appears = success; if it errors, check the table above row by row. Built-in knowledge entries are shortcuts, not requirements.

Level guidance (custom providers)

Levels for DSH's built-in routes come from the pi-ai catalog, and the plugin is strictly read-only and never modifies it. Guidance is only provided for models you declared yourself under llm-pi-ai in settings.yaml:

1. Open the model menu. If the current model is one you declared yourself and the catalog can't read its levels (or the declaration disagrees with the knowledge base), View level declaration guidance will appear in the menu;
2. The panel shows suggested levels (the model's recorded levels when the knowledge base has an entry, otherwise a generic template), a complete entry YAML you can copy directly (including the - id: line, with your original fields such as name automatically preserved), and the path to settings.yaml;
3. After copying, use it to replace wholesale the corresponding - id: entry in settings.yaml (don't create a second llm-pi-ai: root) and save. DSH loads it automatically; if it doesn't take effect, restart the Web Host and refresh the page.

Models not in the knowledge base get a generic, freely editable template. For situations like "the endpoint rejects the request because of the developer role," the panel shows a warning and the corresponding compat switch (e.g., it suggests supportsDeveloperRole: false).

If you don't want to fill things in yourself, or it still errors out after you do, click Copy to Agent next to the panel: it assembles the current model's symptoms and locations (route, model id, settings.yaml path, entry line, levels read from the catalog, knowledge-base suggestion, endpoint hints), your task, the complete level declaration rules, and a starter snippet into a single briefing on the clipboard. Paste it to any coding agent, and it can read settings.yaml, check the endpoint documentation, write the configuration, and explain why.

dsh-reasoning-effort:
  entries:
    - id: my-model
      provider: "*"          # provider 路由名，* 通配
      model: "my-model-id"   # 模型 id，* 通配
      note: 说明文字
      efforts:               # 档位名 → 端点实际接受的取值
        low: "low"
        high: "high"
        max: "max"
      # compat:              # 只在端点需要固定格式时才写
      #   thinkingFormat: "qwen"
      #   supportsReasoningEffort: false

Big Fat Fish Slider

Eight-frame running character enabled by default. To use a plain white button:

1. Open Settings → General.
2. Find Big fat fish slider under "Appearance".
3. Turn the toggle off, then go back to the model entry.

Only the button look changes; snapping, keyboard, glow, and model selection are unaffected. Animation speeds up while dragging; with "reduce motion" it stays on a stable frame.

The reasoning effort slider master switch can temporarily disable the enhanced control. No uninstall needed — the native DSH selector resumes. Both toggles live only in the current browser.

FAQ

No slider after install

Check each item:

1. Use dsh --version to confirm the version actually running; the combination currently verified by maintainers is DSH 0.1.5-rc.2 + plugin v0.7.2.
2. DSH Web Host has been restarted after installation.
3. Settings → General Settings → Reasoning Effort Slider is enabled.
4. The current model exposes at least two reasoning effort levels in the DSH model catalog (for models with no declaration, see the next item), and the deployment has not disabled thinking.

Model has no levels

First check the View Level Declaration Guide in the model menu. If manual configuration is needed, fill in the reasoningEfforts and compat fields for the corresponding model in settings.yaml based on the current model and endpoint documentation; do not copy levels or context parameters from other models directly.

KB is reference only; actual values depend on endpoint. If not effective, restart Web Host and refresh.

RC issues, how to report

Please include the DSH version, plugin version, client type (Web or desktop wrapper), reproduction steps, and relevant console errors in an Issue. Redact tokens and credentials before reporting.

How To Confirm The Plugin Is Loaded

Run:

dsh --profile web --dump-config

The configuration should include name: dsh-reasoning-effort.

How To Uninstall

dsh plugin --profile web remove dsh-reasoning-effort

Restart DSH Web Host after uninstall; native model selector auto-restores.

Development & Build

pnpm install
pnpm run check
pnpm pack

The development environment uses Node.js 22.19+ (also satisfying the target DSH requirements) and pnpm@11.7.0. pnpm run check performs TypeScript and internationalization validation, and rebuilds the Host entry, browser module, and type declarations. For the full interaction and color conventions, see design/visual-spec.md; security issues should be reported per SECURITY.md.

License

MIT © HanaAyane