pi-qwen-mode-proxy

Sampling mode proxy extension for pi that intercepts OpenAI-completions API requests to a llama.cpp server and injects mode-specific sampling parameters for Qwen models (tested with Qwen 3.6 27B). Parameters taken from the recommendation note on model page (https://huggingface.co/Qwen/Qwen3.6-27B).

Modes

| Parameter | 🧠 Thinking | 💻 Coding | 📝 Instruct | |-----------|:-----------:|:---------:|:-----------:| | temperature | 1.0 | 0.6 | 0.7 | | top_p | 0.95 | 0.95 | 0.80 | | top_k | 20 | 20 | 20 | | min_p | 0.0 | 0.0 | 0.0 | | presence_penalty | 0.0 | 0.0 | 1.5 | | repetition_penalty | 1.0 | 1.0 | 1.0 |

• Thinking — Creative, exploratory tasks. High temperature for diverse output.
• Coding — Precise, deterministic coding tasks. Lower temperature for consistent results.
• Instruct — Instruction-following with presence penalty to encourage topic variety.

Installation

npm

pi install npm:pi-qwen-mode-proxy

git

pi install git:github.com/YOUR_USERNAME/pi-qwen-mode-proxy

local

pi install /path/to/pi-qwen-mode-proxy

Usage

Requires a llama.cpp server serving a Qwen model, registered as the llamacpp provider in ~/.pi/agent/models.json:

{
  "providers": {
    "llamacpp": {
      "baseUrl": "http://10.10.10.11:8080/v1",
      "api": "openai-completions",
      "apiKey": "llamacpp",
      "compat": {
        "supportsDeveloperRole": true,
        "supportsReasoningEffort": true,
        "thinkingFormat": "qwen-chat-template"
      },
      "models": [
        {
          "id": "llamacpp",
          "name": "(Local AI)",
          "reasoning": true,
          "input": ["text"],
          "contextWindow": 131072,
          "maxTokens": 16384,
          "cost": { "input": 0, "output": 0, "cacheRead": 0, "cacheWrite": 0 }
        }
      ]
    }
  }
}

Commands

/mode              Show current mode and parameters
/mode thinking     Switch to thinking mode
/mode coding       Switch to coding mode
/mode instruct     Switch to instruct mode

The current mode is displayed in the status bar footer and persists across /reload via session storage.

How It Works

The extension hooks into pi's before_provider_request event, which fires after pi builds the OpenAI chat completions payload but before it's sent over the network. When the target model is llamacpp, the handler injects the six sampling parameters (temperature, top_p, top_k, min_p, presence_penalty, repetition_penalty) corresponding to the active mode.

No custom provider or streaming implementation is needed — the extension works as a lightweight interceptor on top of pi's built-in openai-completions provider.

Configuration

The model ID filter defaults to llamacpp. If your provider/model uses a different ID, edit extensions/index.ts and change the TARGET_MODEL constant.

License

MIT