Pi DeepSeek Tuner

中文版

DeepSeek model tuner extension for Pi. Intercepts and optimizes API payload interactions with DeepSeek models (both official and third-party compatible endpoints) by intercepting requests at the provider gateway level, saving costs and ensuring high availability.

Key Mechanics

1. Reasoning History Stripping (Cost-Saving)

   • The Problem: Resending previous chain-of-thought (reasoning_content or <think> tags) back to DeepSeek models in multi-turn conversations acts as billable input tokens, drastically blowing up API costs.
   • The Solution: Intercepts before_provider_request payload and deletes the reasoning cache from assistant history, saving up to 80% on API fees.

2. Strict Tool-Call Pairing (Sanitization)

   • The Problem: Dangling or unanswered tool_calls generated by the model due to connection issues or session interrupts cause DeepSeek API to throw a 400 Bad Request validation error.
   • The Solution: Validates the conversation history, backfills placeholder results for unanswered calls, and discards orphan tool responses to strictly respect the API schema.

3. Auto Thinking Mode Toggle

   • Automatically injects { thinking: { type: "enabled" } } for DeepSeek R1 models to ensure reasoning chunks are streamed successfully even through non-official proxy endpoints.

Project Structure

pi-deepseek-tuner/
├── index.ts        # Core hook logic (TypeScript)
├── package.json    # Extension manifest with peerDependencies for Pi
├── tsconfig.json   # TypeScript configurations targeting NodeES
└── biome.json

Installation

pi install npm:pi-deepseek-tuner

Testing

Run the native TypeScript unit test suite (Node.js native test runner):

npm run test

Development

We align with the code quality and security standards of the primary pi-main workspace:

1. Linter & Formatter: Biome is configured to auto-format and check codebase. We use tab indentation with a width of 3 characters.
2. Commands:
   • Check format and Lint: npm run check
   • Run tests: npm test

License

MIT