dsh-reasoning-effort

English | 中文

>  > 

A DSH plugin that enables reasoning effort selection in the official model selector. Works with any provider and any model. Install with one dsh plugin command.

---

English Documentation

Method 1: Let the Agent install (recommended)

If your current Agent can execute terminal commands, send the following text verbatim to it:

Could you help me install the reasoning-effort plugin for the DeepSeek Harness "web" Profile?

First, check whether it's already there:
dsh --profile web --dump-config

If dsh-reasoning-effort already appears in the output, just tell me "already installed" and stop.
If not, install it (only modify the web Profile, leave others alone):
dsh plugin --profile web add github:byxumi/dsh-reasoning-effort#main

NOTE: If pnpm prints "Ignored build scripts" (or "Ignored build scripts:
dsh-reasoning-effort ... Run pnpm approve-builds"), it means pnpm blocked the
auto-patch hook. Open:
C:\Users\<username>\.dsh\profiles\web\pnpm-workspace.yaml
and change the allowBuilds line for dsh-reasoning-effort that says
"set this to true or false" to true (keep indentation unchanged), then
re-run the install command above. This only allows the plugin's auto-patch
hook — no other code is executed.

After installation, verify again:
dsh --profile web --dump-config

Tell me the result — whether dsh-reasoning-effort now appears in the config.
Do NOT shut down or restart the running DSH process; remind me to manually restart the DSH Web Host afterwards.

The Agent should check first, then install if needed, and report the final status.

After installation, manually restart the DSH Web Host for the plugin to take effect.

> About auto-patching: the repo's package.json has prepare/postinstall hooks > that automatically run node install.js --off during installation, which patches > all pi-ai adapter copies (profiles, npx cache, npm global, desktop runtime) with > the compat fields — preventing 400 developer is not one of [...] errors. > pnpm blocks the hook on first install and requires you to approve it in > pnpm-workspace.yaml allowBuilds (the prompt above covers this step). > Even if you skip the approval, the plugin self-heals at load time > (but requires an extra restart).

#### Prompt for Agent (uninstall)

To uninstall, send the following:

Could you help me fully uninstall the reasoning-effort plugin from DeepSeek Harness?

Follow these steps:
1. Find or download the uninstall script (uninstall.js):
   - If uninstall.js already exists in the current directory or ~/.dsh/, skip the download.
   - Otherwise download it with curl:
     curl -fsSL -o uninstall.js https://raw.githubusercontent.com/byxumi/dsh-reasoning-effort/main/uninstall.js
   - If curl is unavailable (legacy Windows), use PowerShell:
     Invoke-WebRequest -Uri https://raw.githubusercontent.com/byxumi/dsh-reasoning-effort/main/uninstall.js -OutFile uninstall.js
2. Run (cleans pi-ai patches, settings.yaml declarations, and Profile config — all in one step):
   node uninstall.js --all
3. If node is also unavailable, just tell me "Node.js missing" — don't force-install anything.

Tell me the result and remind me to restart the DSH Web Host.

#### What the Agent receives

After the plugin is installed and DSH restarts, the Agent automatically gains:

1. reasoning-effort skill registration — the model can invoke it to view usage instructions 2. System prompt injection — a Reasoning Effort section is added to every conversation's system prompt

After installation, ask the Agent in a new session: "Can you control reasoning effort?" — it will confirm it knows the feature and guide you.

Features

The selected level is sent as reasoning_effort in API requests.

What the plugin does

1. Registers the reasoning-effort skill — callable by the model/user to view instructions 2. Key point: when mounted as a profile bundle, combined with reasoningEfforts declarations in settings.yaml, the official model selector automatically shows the Effort entry

Configuration template

llm-pi-ai:
  providers:
    your-gateway:
      api: openai-completions
      baseURL: https://your-gateway/v1
      models:
        - id: your-model
          name: your-model
          compat:
            supportsReasoningEffort: true
            supportsDeveloperRole: false
            supportsStore: false
            maxTokensField: max_tokens
          reasoningEfforts:
            off: null
            low: low
            medium: medium
            high: high
            xhigh: xhigh
            max: max

> supportsDeveloperRole: false prevents pi-ai from sending system as developer role. > supportsStore: false / maxTokensField: max_tokens keep the request body compatible with OpenAI-style gateways.

Uninstall

dsh plugin --profile web remove dsh-reasoning-effort
# or via script:
node uninstall.js --all

Troubleshooting

---

License

[MIT](LICENSE)

Community

• [Contributing](CONTRIBUTING.md)
• [Code of Conduct](CODE_OF_CONDUCT.md)
• [Security](SECURITY.md)