Combswell — Reasoning That Keeps Its Thread

Guarded models for VS Code.

Combswell is a standalone VS Code language-model provider that routes model requests through the Combswell router service and guards the exposed reasoning/thinking stream. When that stream stutters — a repeated normalized token window or a high duplicate-token ratio — the request is cancelled and one replacement request is issued with a concise corrective instruction.

It appears as a first-class, selectable provider in VS Code's native model picker. It is not a generic HTTP proxy and not a custom chat participant.

Requirements

• VS Code 1.104.0 or newer (the provider API was finalized in 1.104).
• A Combswell API key.
• GitHub Copilot on an individual plan (models contributed through this API are currently only available to individual Copilot plans).

Getting started

1. Install the extension from the Marketplace.
2. Run Configure API Key from the Command Palette and paste your Combswell API key. It is stored in VS Code SecretStorage and is never written to settings or logs.
3. Open the Chat view, choose Manage Models…, select Combswell, and enable the models you want.

Run Configure API Key again to replace the key, or Clear API Key to remove it.

Settings

• combswell.models — model ids to expose in the model picker.
• combswell.gatewayUrl — base URL of the Combswell router service. Leave empty to use the COMBSWELL_GATEWAY_URL environment variable, then the default.
• combswell.requestReasoning — request exposed reasoning from models that advertise support. Hidden chain-of-thought is never accessible.
• combswell.reasoningEffort — reasoning effort requested from models that support it.

Privacy

• The API key lives in SecretStorage and is never logged.
• Authorization headers are never logged.
• Response buffers are bounded, and request/response contents are released when a request completes.
• No OpenAI-compatible HTTP server is exposed and Copilot's network traffic is not intercepted.

License

MIT. See LICENSE.