⚡ Copilot Pulse — Intelligent Model Router for GitHub Copilot

Stop overpaying for simple prompts. Copilot Pulse analyzes every request and automatically routes it to the right model in your GitHub Copilot subscription — a lightweight model for trivial asks, a heavyweight model for hard problems — and auto-escalates to a stronger model when a cheaper one gets stuck.

One agent. Every tier. Zero manual model-switching.

📋 See ROUTING_IMPROVEMENTS.md for what changed in routing, cost-aware selection, metrics and tests.

Why Copilot Pulse?

GitHub Copilot gives you access to many models — from fast, cheap ones to powerful, premium ones. But you have to pick manually, and most people just leave it on one model: either overpaying premium credits for "rename this variable", or under-powering a "design a distributed system" request.

Copilot Pulse fixes that:

• 🎯 Automatic routing — scores each prompt's complexity (0–10) and picks the optimal tier.
• ⬆️ Auto-escalation — if a lighter model loops, errors, or gives up, Pulse retries on a stronger model automatically. You don't have to notice or intervene.
• 💰 Real savings — trivial prompts stay on cheap models, so premium credits are spent only where they matter.
• 📊 Live dashboard — see every routing decision, your savings rate, and credit usage in real time.

Features

🧠 Context-aware complexity scoring

Pulse doesn't just keyword-match. It weighs:

• Intent — architecture vs. debugging vs. lookup vs. boilerplate
• Structure — multi-part requirements, code blocks, stack traces, diffs
• Context — attached files, cross-file references, workspace size
• Conversation history — a follow-up like "now fix it" inherits the depth of the conversation instead of collapsing to the cheapest model
• Tool surface — agent turns are routed to models that reliably emit tool calls

Every decision comes with an explainable score breakdown in the dashboard.

⬆️ Automatic tier escalation

The headline feature. When the selected model can't finish the job, Pulse detects it and moves up a tier — Light → Medium → Heavy — carrying the full conversation forward. Triggers include:

• Repeating the same tool call without progress (stuck loops)
• Repeated tool failures
• Context-window overflow (auto-shrinks the request, then upgrades)
• Empty or malformed responses
• The model explicitly saying it can't do the task
• You replying "that's still wrong"

Escalation is capped per conversation so it never runs away on cost.

📊 Live routing dashboard

A dedicated side-panel view showing:

• Total queries routed & savings rate
• Premium credits used vs. saved
• Tier / model / task-type distribution
• A live log of recent routing decisions with the score breakdown behind each one

🛡️ Robust by design

• Streams never stop mid-answer — interrupted responses recover gracefully.
• Malformed or oversized prompts degrade cleanly instead of crashing.
• Long multi-step agent tasks run to completion (up to a configurable turn budget).

Getting started

1. Install Copilot Pulse (you'll need GitHub Copilot Chat installed and signed in).
2. Open the Chat view, choose the Copilot Pulse Agent chat mode.
3. Type @copilotpulse once at the start of your conversation, then chat normally.

That's it. Pulse routes across the models already in your Copilot subscription — no API key required. The @copilotpulse participant is sticky, so follow-up messages route automatically without re-typing it.

💡 Run Copilot Pulse: Show Supported Models from the Command Palette to see which of your subscription models landed in each tier.

Usage modes

The BYOK picker entries only appear once you've configured an endpoint — so you'll never see a model in the picker that can't answer.

Configuration

Commands

• Copilot Pulse: Open Dashboard — the live routing view.
• Copilot Pulse: Show Supported Models — your discovered models grouped into tiers.
• Copilot Pulse: Test Connection — verify a configured BYOK endpoint.

Which models does Pulse actually use?

Pulse adapts to your subscription — it does not ship a fixed model list.

At startup Pulse calls vscode.lm.selectChatModels({ vendor: "copilot" }) and partitions whatever models your account actually exposes into Light / Medium / Heavy bands by inferred strength. It never invents a model you don't have.

This means the catalog is environment-dependent. A session that exposes 14 models gets 14-way distribution; a session that exposes 5 gets 5-way distribution across the same three bands. Both are working as intended — the band a prompt is routed to is what Pulse controls; the concrete model is whichever member of that band your subscription provides.

If a prompt scores into a band where you have no model, ModelResolver maps it to the closest available model in the same or an adjacent band rather than failing or silently falling back to the most expensive option.

To see exactly what was discovered in your environment:

• run Copilot Pulse: Test Connection (reports the count and names), or
• open the dashboard's Models tab, which lists every discovered model with its inferred band and strength, or
• run Copilot Pulse: Open Dashboard and use ↻ Re-scan after changing plans.

Because tiering is derived from the live discovery result, savings percentages quoted in benchmarks are specific to the catalog available at that time.

Requirements

• VS Code 1.95.0 or newer.
• GitHub Copilot + GitHub Copilot Chat, signed in.
• BYOK model-picker entries additionally require the languageModelChatProvider proposed API (VS Code Insiders or an extension dev host).

Privacy

Copilot Pulse runs entirely locally. Prompt scoring and routing happen on your machine; your prompts are sent only to the model provider you're already using (your Copilot subscription, or your own configured endpoint). Pulse adds no telemetry and no third-party servers.

Known limitations

• The BYOK model-picker entries depend on a VS Code proposed API and won't appear on stable builds — use the @copilotpulse participant there instead.
• Model tiering is inferred from model names/families; brand-new models may need a release to be classified optimally.
• The set of routable models is whatever your Copilot subscription exposes at runtime (see Which models does Pulse actually use?) — it is not a fixed catalog, so routing granularity varies by account.

License

MIT © 2026 Copilot Pulse