@senomas/proxy-provider

A pi-coding-agent extension that registers a deepseek provider via a configurable AI proxy.

What it does

This package extends pi with a "deepseek" provider (always registered). On every session:

1. Discovers .pi-project.json: walks up from the current directory; if not found, checks the main git worktree root (so linked worktrees work); if still not found, fails with a descriptive error
2. Reads the project, proxy, and optional providers fields from that file
3. Fetches model definitions from ${proxy}/static/providers/deepseek.json and registers the default "deepseek" provider
4. If a providers map is present, fetches and registers each additional provider
5. All providers get the proxy as baseUrl and the project name as the AI-Proxy-Project header

Prerequisites

• Node.js 18+ (for global fetch)
• pi-coding-agent installed (npm install -g @earendil-works/pi-coding-agent)

Installation

pi install npm:@senomas/proxy-provider

Configuration

Create a .pi-project.json file in your project root:

{
  "project": "my-project",
  "proxy": "https://proxy.example.com"
}

| Field | Type | Description | |-------|------|-------------| | project | string | Project name sent as the AI-Proxy-Project header | | proxy | string | Base URL of the AI proxy server | | providers | object (optional) | Map of provider names→config filenames served by the proxy at {proxy}/static/providers/ |

Multiple providers

You can register additional providers by adding a providers map. Each entry registers a provider with the same baseUrl and AI-Proxy-Project header as the default "deepseek".

{
  "project": "my-project",
  "proxy": "https://proxy.example.com",
  "providers": {
    "deepseek-work": "deepseek.json",
    "deepseek-chat": "deepseek-chat.json"
  }
}

Behavior details:

| Aspect | Detail | |--------|--------| | Default | "deepseek" from deepseek.json is always registered | | Override | If you include a key named "deepseek" in providers, it overrides the default (last registerProvider wins) | | Empty {} | Treated as "no additional providers" — default "deepseek" is still registered | | Config files | Each value must be a non-empty JSON filename served by the proxy under {proxy}/static/providers/ | | Validation | Keys and values must be non-empty strings. Invalid entries throw a descriptive error at startup | | Fail-fast | If fetching any config file fails, the entire extension throws — no partial registration |

Config file content — each referenced JSON file contains the provider's model definitions, which are merged with baseUrl and headers.AI-Proxy-Project before registration. The exact format depends on your proxy server's provider schema (typically a list of models with their capabilities).

Server-side dependency

This extension requires a running AI proxy server that exposes provider configuration files under /static/providers/.
Refer to the ai-team-proxy project for the reference proxy implementation and deployment instructions.

Verification

After installation, list available models to confirm the provider is registered:

pi --list-models | grep deepseek

You should see the deepseek models exposed by your proxy.

How it works

The extension exports an async default function that:

1. Finds .pi-project.json in two steps:
   • Walks up from the current directory (process.cwd()) looking for .pi-project.json
   • If not found, resolves the main git worktree root via git rev-parse --path-format=absolute --git-common-dir (falling back to parsing git worktree list --porcelain for git < 2.31) and checks <root>/.pi-project.json — so a config kept only in the main worktree is found from any linked worktree. Outside a git repo this step is skipped, and if no config exists anywhere the extension fails with a descriptive error.
2. Parses the file with fs.readFileSync/JSON.parse and validates the project, proxy, and optional providers fields
3. Calls fetch() to retrieve the default "deepseek" provider metadata from the proxy
4. If a providers map is present, fetches and registers each additional provider in the same way
5. Calls pi.registerProvider(name, config) for each provider with the fetched models, proxy baseUrl, and AI-Proxy-Project header

Testing

npm install
npm test

Runs the unit/integration tests (Node's built-in test runner via tsx) against test/*.test.ts.

License

MIT