add windsurf as a Provider

[Coral5644]

Hi, I hope you can add windsurf as a Provider. I have tested it for you and it works. However, your code is confusing and I can't incorporate it. Here are my experiences that might be helpful to you. Thank you for your hard work.

Details

1. Token Storage Architecture

Windsurf stores authentication sessions using the standard Chromium encrypted storage pattern:

Component	Location	Purpose
Master key	%APPDATA%/Windsurf/Local State → os_crypt.encrypted_key	Base64-encoded DPAPI blob containing AES-256 key
Encrypted sessions	%APPDATA%/Windsurf/User/globalStorage/state.vscdb	SQLite DB with secret:// prefixed entries, AES-256-GCM encrypted

2. Extraction Chain

Step 1: Read Local State, decode the base64 encrypted_key, strip the 5-byte DPAPI prefix.

Step 2: Call CryptUnprotectData (Windows DPAPI) — succeeds for any process running as the same user, no admin required, no prompt shown.

Step 3: Use the resulting 32-byte AES-256 key to decrypt sessions from state.vscdb. The encrypted values use Chromium's v10 format: 3-byte prefix + 12-byte nonce + ciphertext + 16-byte GCM auth tag.

Result: Plaintext accessToken in the format sk-ws-01-... (103 characters), along with account metadata.

3. Verified API Access with Extracted Token

The extracted token grants full access to Codeium's backend services:

Direct API calls to server.codeium.com (Connect RPC + Protobuf):

RPC Method	Service	Result
MigrateApiKey	SeatManagementService	Server returns full token confirmation
GetPrimaryApiKeyForDevsOnly	SeatManagementService	Returns internal UUID
GetUserNotifications	SeatManagementService	Server accepts token
BatchRecordAnalyticsEvents	ProductAnalyticsService	Successfully sends events as the user
GetProfileData	SeatManagementService	Server accepts token

Via Language Server proxy (localhost Connect RPC):

RPC Method	Result
GetUserStatus	Returns ~31KB of user data (name, email, UUID, subscription info, available models)
InitializeCascadePanelState	Session initialization succeeds
StartCascade + SendUserCascadeMessage	Full AI inference — streaming responses from Claude, GPT, Gemini models

4. Full AI Inference Exploitation

The most critical finding: an attacker can start the bundled language server binary with the stolen token and make unlimited AI inference requests through the Cascade API. The flow:

Attacker process
  → Extract token from disk (DPAPI + AES-GCM)
  → Spawn language_server binary with --stdin_initial_metadata
  → Write Metadata protobuf (containing stolen api_key) to stdin
  → Call InitializeCascadePanelState → GetUserStatus → UpdatePanelStateWithUserStatus
  → StartCascade → StreamCascadeReactiveUpdates + SendUserCascadeMessage
  → Receive full AI model responses (streaming)

This was verified to work with multiple models including Claude Opus 4.6, Claude Sonnet 4.6, and others available on the user's subscription.