pi-zenmux

ZenMux provider extension for Pi. Auto-discovers and registers all available models from the ZenMux API gateway, with proper DeepSeek thinking-mode compat.

Install

pi install npm:@conreo/pi-zenmux

Setup

Set your ZenMux API key:

export ZENMUX_API_KEY="sk-..."

Usage

Start pi — ZenMux models appear under the zenmux provider:

pi

Select a model with /model or use zenmux/<model-id> directly:

/model zenmux/deepseek/deepseek-v4-pro
/model zenmux/openai/gpt-4o

Commands

• /zenmux-refresh — Force refresh the model list from the ZenMux API

Supported Models

All models available through the ZenMux gateway are automatically discovered. This includes:

• DeepSeek — v4-pro, v4-flash, v3.2, chat, reasoner, r1 (with proper thinking-mode compat)
• OpenAI — gpt-4o, gpt-4o-mini, o3-mini, o4-mini, etc.
• Anthropic — claude-sonnet-4-20250514, claude-opus-4-20250514, etc.
• Google — gemini-2.5-pro, gemini-2.5-flash, etc.
• And more — all models registered at zenmux.ai

DeepSeek Thinking Mode

This extension registers DeepSeek models with the proper compat flags so Pi natively manages reasoning_content across multi-turn conversations with tool calls. No fetch-intercept hacks — Pi's built-in handling activates automatically.

How It Works

1. Startup: Reads cached model list from ~/.cache/pi-zenmux/models.json (24h TTL)
2. Cache hit: Registers models synchronously (zero latency)
3. Cache miss: Registers an empty provider, fetches models in the background, writes cache
4. Model registration: DeepSeek models get compat.requiresReasoningContentOnAssistantMessages and thinkingFormat: "deepseek" so Pi handles reasoning properly

Development

Test locally without publishing:

ZENMUX_API_KEY="sk-..." pi -e ./extensions/index.ts

License

MIT