GitHub Copilot Model Advisor
Know which GitHub Copilot Chat model to use right now, without leaving VS Code.
You notice Copilot feels slow. Is it the model, the provider, or just your current task? Without leaving the editor you have no signal — until now.
GitHub Copilot Model Advisor answers that question in seconds: it checks which models are enabled for your account, reads public provider health feeds without sending any prompts, and optionally measures first-token latency when you want a live speed signal.
Website: cperezsx.github.io/GitHubCopilotModelAdvisor
Contents
Why It Exists
The everyday scenario: you are mid-task, Copilot feels sluggish, and you do not know whether to switch models or wait. Opening a provider dashboard breaks your flow. Asking a colleague is not always an option.
Model Advisor keeps that signal inside VS Code. Open the sidebar, run a health check, and in a few seconds you see which of your enabled models is healthy and fast right now. The default check never sends a prompt to any model and uses zero GitHub Copilot tokens. Benchmarking is opt-in, always asks for confirmation, and saves timestamped latency history in a global cache shared across your VS Code workspaces.
What Is Real and What Is a Heuristic
Being honest about what the extension can and cannot know:
Reliable signals
| What |
Why it is reliable |
| Model list |
Read directly from vscode.lm.selectChatModels({ vendor: "copilot" }) — the official VS Code API. Always reflects your account, organization, and session. |
| Provider health and incidents |
Read from the same public status feeds the providers themselves maintain (OpenAI, Anthropic, Google, Moonshot AI, GitHub). When an incident is declared there, it appears here. |
| First-token latency (benchmark) |
Measured by sending a tiny real prompt and cancelling the stream on the first chunk. As honest as latency measurement gets. |
Heuristics and known gaps
| What |
Why it is approximate |
| Recommendation score |
Starts at 100, subtracts points for slow tokens, degraded providers, and active incidents. Useful orientation — not a guarantee about your specific prompt or task. |
| Google / Gemini status |
Google Cloud's incident feed is generic. The extension flags incidents that mention Gemini, Vertex AI, or generative AI — it will miss silent degradation that goes undeclared. |
| Benchmark vs real-world latency |
A hi prompt measuring 300 ms does not predict a 4 000-token context. It tells you the model is alive and responding, not how it will perform under load. |
| Silent saturation |
If a provider's servers are stressed but no incident has been declared, the extension cannot detect it. Neither can any external tool. |
The Auto entry (GitHub's routing alias) is excluded from detection, scoring, and benchmarks: because GitHub picks the underlying model per request, latency and provider health cannot be attributed to it.
Quickstart
- Install GitHub Copilot Chat and sign in to GitHub in VS Code.
- Install GitHub Copilot Model Advisor.
- Open the Model Advisor icon in the Activity Bar.
- Click Health check.
- Use the recommended model at the top, or pick from the light / medium / complex task suggestions.
When raw speed matters, click Benchmark on one model or Benchmark all. The extension shows a confirmation dialog before sending any prompt. If a previous benchmark exists, you can reuse cached results instead of spending more tokens. Cached latency shows its age, stale state, recent median, and trend.
How It Works
Models are discovered through the official VS Code language model API, so the list always matches what your account can actually use. Each model starts at a score of 100 and loses points based on three signals:
1. Latency (only when benchmarked)
| First-token time |
Penalty |
| Under 800 ms |
0 |
| 800 – 1 500 ms |
−10 |
| 1 500 – 3 000 ms |
−30 |
| Over 3 000 ms |
−50 |
| Timeout |
−90 |
| Error |
−80 |
| Not benchmarked |
−10 |
2. Provider health
| Status |
Penalty |
| Operational |
0 |
| Degraded performance |
−20 |
| Partial outage |
−40 |
| Major outage |
−80 |
| Unknown |
−5 |
3. Active incidents add −30 on top. Ties break by latency.
Health signals come from public status feeds:
| Provider |
Source |
| OpenAI |
status.openai.com |
| Anthropic |
status.claude.com |
| Google |
status.cloud.google.com |
| Moonshot AI (Kimi) |
status.moonshot.cn |
| Microsoft (MAI) |
githubstatus.com — Microsoft AI publishes no dedicated public feed, so MAI models track their delivery route |
| xAI (Grok) |
githubstatus.com — status.x.ai has no machine-readable feed, so Grok models track their delivery route |
| GitHub Copilot |
githubstatus.com |
Commands & Settings
| Command |
Purpose |
Open Advisor |
Opens the sidebar. |
Check Health Now |
Token-free detection, health check, and scoring. |
Benchmark Latency (Uses GitHub Copilot Tokens) |
Confirms, then runs the live latency benchmark. |
Copy Diagnostics |
Copies a support-friendly report with model scores, confidence, health, and cache timestamps. |
Open Settings |
Opens extension settings. |
| Setting |
Default |
Purpose |
githubCopilotModelAdvisor.autoCheckOnStartup |
false |
Run a token-free check when VS Code starts. |
githubCopilotModelAdvisor.showInStatusBar |
true |
Show the status bar item. |
githubCopilotModelAdvisor.testPrompt |
hi |
Prompt used only for explicit latency benchmarks. |
githubCopilotModelAdvisor.benchmarkCacheTtlMinutes |
120 |
Minutes before cached latency is marked stale. Use 0 to never mark it stale automatically. |
githubCopilotModelAdvisor.allowBenchmarkAll |
true |
Allow benchmarking all enabled models in one action. Disable to keep only per-model benchmarks. |
Limits
- Model switching: VS Code does not expose an API for an extension to set the active Copilot Chat model. The extension advises; you select.
- Silent saturation: Undeclared provider degradation is invisible to any external tool, including this one.
- Benchmark ≠ production latency: The tiny benchmark prompt measures responsiveness, not throughput under your real workload.
- Latency cache: Cached benchmark results are shared across VS Code workspaces for this extension. The UI shows when the cached benchmark was measured, whether it is stale, median recent latency, and trend so you can decide whether to rerun it.
- Confidence: Recommendations expose a confidence label. Live benchmark plus healthy provider is strongest; stale cache or degraded provider lowers confidence.
- Google precision: Only declared Google Cloud incidents mentioning Gemini, Vertex AI, or generative AI are flagged.
- Microsoft (MAI) and xAI (Grok) health: Neither publishes a machine-readable public status feed, so their model health follows the GitHub Copilot service status.
- Tokens: Only explicit benchmarks use GitHub Copilot tokens. Health checks never do.
Development
npm install
npm run check # type-check only
npm run compile
npm run package # build .vsix
Press F5 to launch an Extension Development Host. Run node scripts/webview-preview.js to render the sidebar with sample data in out/webview-preview.html for visual inspection and screenshots. Design notes and internal specs live in docs/internal/.
Support
Report bugs via GitHub Issues. Include your VS Code version, OS, extension version, and whether GitHub Copilot Chat is installed and signed in. Never include secrets, private prompts, or provider tokens. See SUPPORT.md and SECURITY.md.
Authors
Carlos Perez — GitHub: @cperezsx · LinkedIn: cperezsx
Jose Miguel Durá — GitHub: @JMDura · LinkedIn: Jose Miguel Durá Sirvent
License
MIT. See LICENSE.
