pi-extension-raptor-mini

Adds GitHub Copilot's Raptor mini model (oswe-vscode-prime) to pi.

Why

pi's model catalog is baked in at build time. Raptor mini went to public preview on 2025-11-10, after the catalog was last regenerated, so pi --list-models doesn't show it. This extension adds it at runtime.

When pi-ai picks it up upstream, this extension becomes a no-op (it detects the duplicate, warns, and exits).

Install

pi install npm:pi-extension-raptor-mini

The npm: prefix is required. pi install pi-extension-raptor-mini is interpreted as a local filesystem path, not an npm package.

Alternatives:

pi install git:github.com/WSeubring/pi-extension-raptor-mini   # from git
pi -e /path/to/pi-extension-raptor-mini                        # local checkout, no install

Use

pi --list-models | grep oswe
# github-copilot  oswe-vscode-prime  264K  64K  no  no

pi --provider github-copilot --model oswe-vscode-prime -p "hello"
# or match by name substring:
pi --model raptor -p "hello"

You need to be logged into Copilot in pi (/login github-copilot inside the interactive UI, or pi prompts on first use). Any Copilot plan works — Free, Pro, Pro+, Business, Enterprise.

Model facts

| Field | Value | Source | |---|---|---| | Canonical id | oswe-vscode-prime | zed#49514 | | Name | Raptor mini (Preview) | GitHub Changelog 2025-11-10 | | Base | Code-tuned GPT-5-mini | GitHub Changelog | | Context | ~264,000 tokens | lobehub model-bank (GitHub does not publish officially) | | Max output | ~64,000 tokens | Community estimate — no authoritative source | | Tool calls | Supported | GitHub Changelog |

Heads-up

pi's github-copilot provider talks to Copilot's API with VS Code's own integration headers (Copilot-Integration-Id: vscode-chat, etc.). That's how Copilot models — including Raptor mini — are reachable at all. This behavior predates this extension and applies to every request pi's Copilot provider already makes; installing this extension doesn't change the request shape, just adds one model id to the catalog.

GitHub's AUP prohibits misrepresenting client identity; in practice, enforcement has targeted automation/volume abuse rather than client-identity alone. If you want zero ambiguity, use pi's anthropic / openai / google providers with your own API keys instead.

License

MIT