9Router for GitHub Copilot

Connect 9Router and multiple OpenAI-compatible inference backends to GitHub Copilot Chat and inline code completions — with multi-provider profile management and resilience layers for self-hosted models.

Quick Start

1. Install from the VS Code Marketplace.
2. Run 9Router: Manage Providers (Cmd+Shift+P / Ctrl+Shift+P) or click the 9Router status bar item to add your first inference provider.
3. In Copilot Chat, open the model picker → Manage Models → enable the models you want to use.
4. Select a model and start chatting.

Multi-Provider Support

Connect and use multiple inference backends simultaneously (e.g. local Ollama, a remote vLLM cluster, and cloud 9Router).

• Profile Isolation: Each provider profile maintains its own endpoint URL, API key, custom HTTP headers, and enabled/disabled state.
• Secure Storage: API keys and sensitive custom headers are stored in VS Code SecretStorage, not in plaintext workspace settings.
• Smart Model Namespacing:
  • Single Active Provider: Models expose raw IDs (e.g. qwen2.5-coder-32b) for seamless backward compatibility.
  • Multiple Active Providers: Models are automatically namespaced as <profile-id>/<model-id> (e.g. ollama/qwen2.5-coder-32b, vllm/qwen2.5-coder-32b) to prevent collisions, with provider details displayed directly in the model picker.
• Unified Provider Registration: Contributed directly to VS Code's language model chat provider registry with provider management accessible from VS Code's Manage Language Models UI.
• Quick Toggle: Enable or disable any profile to show or hide its models without deleting the configuration.
• Automatic Migration: Upgrades from single-provider versions migrate existing endpoints, keys, and headers into a Default profile automatically.

Inline Code Completions (Ghost Text)

9Router can provide inline code completions directly from your inference backend using Fill-in-the-Middle (FIM):

1. Set 9router-for-github-copilot.enableInlineCompletion to true.
2. Run 9Router: Select Inline Completion Model to pick a FIM-capable base model (e.g. Qwen2.5-Coder base).
3. Optional: Route inline completions to a specific provider via 9router-for-github-copilot.inlineCompletionProvider (leave blank to use the first active provider).

Status Bar & Observability

The status bar shows a connection icon and host. Its icon and hover card use the same snapshot, so failed model fetches stay visible. Disabled profiles do not affect overall connection health.

Hover over the status bar to see provider states and model counts, model capabilities, session token totals, the last request's usage, and quick actions for Refresh, Configure, Test Connection, and Output Log.

Why This Over Native BYOK?

VS Code's built-in BYOK works for standard models. 9Router adds a resilience layer for open-source and self-hosted models:

• Tool calls fail with malformed JSON? Repairs truncated arguments and injects missing required schema fields.
• Reasoning models leak <think> tags? Extracts thinking blocks and routes them into Copilot's collapsible thinking UI.
• Context overflow errors? Auto-detects real server limits, retries automatically, and budgets prompt tokens.
• Model outputs tool calls as plain text? Lowers agent temperature and stabilizes schema enforcement.

Inference stays on your server. No per-token platform surcharges. Does not consume Copilot premium quota.

Supported Backends

9Router, vLLM, Ollama, llama.cpp / llama-server, LM Studio, LocalAI, LiteLLM, or any OpenAI-compatible /v1 endpoint.

Commands

Key Settings

Provider endpoints, API keys, and custom headers are managed per profile via 9Router: Manage Providers. Workspace and user settings govern global runtime behavior:

Troubleshooting

Models not appearing in Copilot?

• Verify your endpoint responds to curl <server-url>/models or <server-url>/v1/models (trailing /v1 is handled automatically).
• Run 9Router: Test Server Connection to check profile health.
• Ensure the profile is enabled in 9Router: Manage Providers.

Tool calls output as text instead of executing?

• Ensure agentTemperature is set to 0.
• Disable parallelToolCalling if the model does not reliably handle parallel calls.
• If using vLLM, ensure --enable-auto-tool-choice and a compatible chat template are enabled.

Context overflow errors?

• Set explicit limits in modelContextWindows for your model or family wildcard (e.g. {"qwen*": 32768}).
• The extension automatically learns corrected context limits on HTTP 400 context errors and retries once.

Models not in the Agents window?

• The Agents window runs in a separate process. Add this setting and reload VS Code:

"extensions.supportAgentsWindow": {
  "hotrungnhan.9router-for-github-copilot": true
}

Utility Tasks (Titles, Commit Messages)

Route Copilot utility tasks (chat titles, commit messages) to your local or self-hosted models:

• Open VS Code Settings → set chat.utilityModel and chat.utilitySmallModel to a 9Router model.

Privacy

Inference prompts and code context are transmitted directly to your configured provider endpoints only. API keys and headers are stored in VS Code SecretStorage.

Development

Requires VS Code 1.144 or newer for the current VS Code API types.

Use Bun for dependencies, scripts, tests, and bundling. Install dependencies with bun install --frozen-lockfile. Run bun run lint, bun run test:compile, and bun test; bun test --coverage adds coverage reporting. Tests run directly from TypeScript with bun:test. Build the extension with bun run build:prod or create a VSIX with bun run package. The generated bundle targets Node because it runs inside VS Code’s extension host; development commands run with Bun.

• api/: HTTP transport and wire types.
• chat/: message conversion, token budgets, tool arguments, and response streaming.
• models/: model metadata, reasoning picker schemas, and profile presentation.
• profiles/: stored profiles, legacy migration, and model routing.
• provider/: configuration and per-profile execution pipelines.
• commands/: VS Code command and input flows.
• status/: connection summaries, usage statistics, and snapshot rendering.
• completions/ and discovery/: inline completion logic and backend metadata discovery.

Credits

This project is a fork of arbs-io/github-copilot-llm-gateway. Thanks to the original authors for the base extension.

License

MIT — see LICENSE.