VS Code Copilot Bridge (Desktop, Inference-only)

Local OpenAI-compatible HTTP facade to GitHub Copilot via the VS Code Language Model API.

• Endpoints (local-only, default bind 127.0.0.1):

  • POST /v1/chat/completions (SSE streaming; use "stream": false for non-streaming)
  • GET /v1/models (dynamic listing of available Copilot models)
  • GET /healthz (ok/unavailable + vscode.version)

• Copilot pipe: vscode.lm.selectChatModels({ vendor: "copilot", family: "gpt-4o" }) → model.sendRequest(messages)

• Prompt normalization: last system message + last N user/assistant turns (default 3) rendered as: [SYSTEM] … [DIALOG] user: … assistant: …

• Security: loopback-only binding by default; optional Authorization: Bearer .

Install, Build, and Run

Prerequisites:

• VS Code Desktop (stable), GitHub Copilot signed in

• Node.js 18+ (recommended)

• npm

• Auto-recovery: the bridge re-requests Copilot access on each chat request if missing; no restart required after signing in. /healthz will best-effort recheck only when bridge.verbose is true.

Steps:

1. Install deps and compile: npm install npm run compile

2. Launch in VS Code (recommended for debugging):

   • Open this folder in VS Code
   • Press F5 to run the extension in a new Extension Development Host

3. Enable the bridge:

   • Command Palette → “Copilot Bridge: Enable”
   • Or set in settings: bridge.enabled = true
   • “Copilot Bridge: Status” shows the bound address/port and token requirement

Optional: Packaging a VSIX

• You can package with vsce (not included): npm i -g @vscode/vsce vsce package
• Then install the generated .vsix via “Extensions: Install from VSIX…”

Enabling the Language Model API (if required)

If your VS Code build requires enabling proposed APIs for the Language Model API, start with:

Models and Selection

• The bridge lists available GitHub Copilot chat models using the VS Code Language Model API. Example: curl http://127.0.0.1:<port>/v1/models → { "data": [ { "id": "gpt-4o-copilot", ... }, ... ] }

• To target a specific model for inference, set the "model" field in your POST body. The bridge accepts:

  • IDs returned by /v1/models (e.g., "gpt-4o-copilot")
  • A Copilot family name (e.g., "gpt-4o")
  • "copilot" to allow default selection

Examples: curl -N -H "Content-Type: application/json"
-d '{"model":"gpt-4o-copilot","messages":[{"role":"user","content":"hello"}]}'
http://127.0.0.1:<port>/v1/chat/completions

curl -N -H "Content-Type: application/json"
-d '{"model":"gpt-4o","messages":[{"role":"user","content":"hello"}]}'
http://127.0.0.1:<port>/v1/chat/completions

• If a requested model is unavailable, the bridge returns: 404 with { "error": { "code": "model_not_found", "reason": "not_found" } }.
• Stable VS Code: code --enable-proposed-api thinkability.copilot-bridge
• VS Code Insiders or Extension Development Host (F5) also works.

When the API is not available, the Output (“Copilot Bridge”) will show: “VS Code Language Model API not available; update VS Code or enable proposed API.”

Troubleshooting

• /healthz shows copilot: "unavailable" with a reason:
  • missing_language_model_api: Language Model API not available
  • copilot_model_unavailable: No Copilot models selectable
  • consent_required: User consent/sign-in required for Copilot models
  • rate_limited: Provider throttling
  • not_found: Requested model not found
  • copilot_unavailable: Other provider errors
• POST /v1/chat/completions returns 503 with the same reason codes.

Configuration (bridge.*)

• bridge.enabled (boolean; default false): auto-start on VS Code startup
• bridge.host (string; default "127.0.0.1"): bind address (keep on loopback)
• bridge.port (number; default 0): 0 = ephemeral port
• bridge.token (string; default ""): optional bearer token; empty disables auth
• bridge.historyWindow (number; default 3): number of user/assistant turns to keep
• bridge.maxConcurrent (number; default 1): max concurrent chat requests; excess → 429

Viewing logs

To see verbose logs:

1. Enable: Settings → search “Copilot Bridge” → enable “bridge.verbose”
2. Open: View → Output → select “Copilot Bridge” in the dropdown
3. Trigger a request (e.g., curl /v1/chat/completions). You’ll see:
   • HTTP request lines (method/path)
   • Model selection attempts (“Copilot model selected.”)
   • SSE lifecycle (“SSE start …”, “SSE end …”)
   • Health checks (best-effort model check when verbose is on)
   • API diagnostics (e.g., missing Language Model API)

• bridge.verbose (boolean; default false): verbose logs to “Copilot Bridge” output channel

Manual Testing (curl)

Replace with the port shown in “Copilot Bridge: Status”.

Health: curl http://127.0.0.1:/healthz

Models: curl http://127.0.0.1:/v1/models

Streaming completion: curl -N -H "Content-Type: application/json"
-d '{"model":"gpt-4o-copilot","stream":true,"messages":[{"role":"user","content":"hello"}]}'
http://127.0.0.1:/v1/chat/completions

Non-stream: curl -H "Content-Type: application/json"
-d '{"model":"gpt-4o-copilot","stream":false,"messages":[{"role":"user","content":"hello"}]}'
http://127.0.0.1:/v1/chat/completions

Auth (when bridge.token is set):

• Missing/incorrect Authorization: Bearer → 401

Copilot unavailable:

• Sign out of GitHub Copilot; /healthz shows unavailable; POST returns 503 envelope

Concurrency:

• Fire 2+ concurrent requests; above bridge.maxConcurrent → 429 rate_limit_error

Notes

• Desktop-only, in-process. No VS Code Server dependency.
• Single-user, local loopback. Do not expose to remote interfaces.
• Non-goals: tools/function calling emulation, workspace file I/O, multi-tenant proxying.

Development

• Build: npm run compile
• Watch: npm run watch
• Main: src/extension.ts
• Note: Previously used Chat API shims are no longer needed; the bridge now uses the Language Model API.