PromptRouter 🧠🚀

Stop burning API dollars on simple prompts. The ultra-intelligent, cost-slashing AI prompt router & autonomous coding agent for VS Code Copilot Chat — save up to 65% on your monthly AI bill with zero drop in answer quality.

💸 Why PromptRouter?

If you use GitHub Copilot Business or Enterprise, your prompts default to expensive heavyweight models (like Claude 3.5 Sonnet or GPT-4o) whether you're asking how to center a div or designing a distributed microservice.

PromptRouter changes the game. It acts as an autonomous neural traffic controller & direct coding agent inside VS Code:

• Simple question? ("How do I format a date in TS?") ⚡ Routed instantly to fast, ultra-cheap models (gpt-4o-mini).
• Hard architecture problem? ("Debug this race condition in Rust") 🚀 Escalated dynamically to elite reasoning engines (claude-3.5-sonnet, gpt-4o).
• Need code changes? 🛠️ Direct Agent Mode automatically generates full code implementations and writes them directly into your workspace files — no manual copy/pasting needed.

📊 Proven ROI: PromptRouter cuts total token spend by over 60% while preserving top-tier response accuracy.

🧠 The 6-Layer Intelligence Pipeline

PromptRouter evaluates every prompt in real-time across a 6-layer cognitive pipeline to compute a precision Complexity Score (0-100):

1. 🔍 Local Heuristic Classifier: Real-time regex & keyword scanning for technical depth, syntax density, and architectural complexity.
2. 🤖 Meta-Model Router: Queries a lightweight sub-agent (gpt-4o-mini) for an instantaneous difficulty score.
3. 📈 Conversation Escalation Engine: Detects thread depth and developer frustration signals ("still failing", "stack overflow") to auto-upgrade model tiers mid-session.
4. 💻 Language Context Sensor: Bumps complexity automatically for high-tier languages (e.g. C++, Rust, Assembly vs HTML/CSS).
5. 🔁 Adaptive Feedback Memory: Learns from your /feedback-up and /feedback-down ratings. If a cheaper model struggled with a specific topic, PromptRouter remembers and upgrades next time.
6. 🛡️ Budget Guardrail Enforcer: Tracks monthly token usage against your target budget and throttles model tiers before overages occur.

🔮 Live Visual Pipeline Header

Every response from @promptrouter includes a live visual breakdown of how your request was routed:

> 🔮 **PromptRouter Intelligence Pipeline**
> ├─ 🔍 **Local Heuristics:** Keyword & Code Analysis Active
> ├─ ⚡ **Meta-Router (gpt-4o-mini):** Complexity Rated **45/100**
> ├─ 📂 **File Context:** rag_agent.py (python)
> └─ 🎯 **Final Route:** **GPT-4o** *(Score: 65/100 | Tier: Advanced | Mode: Agent 🛠️)*

🛠️ Direct Workspace Agent Mode

When you ask PromptRouter to modify or create files (e.g., @promptrouter make changes in rag_agent.py and create a new agent):

1. Auto File Discovery: PromptRouter automatically locates rag_agent.py in your workspace and loads its code context.
2. Direct Code Mutation: PromptRouter generates the complete implementation and automatically writes the code directly into your workspace files via vscode.workspace.applyEdit.
3. Multi-File Updates: Handles multiple file edits in a single turn, opening updated files side-by-side in your editor.

📊 Detailed Analytics & Financial Dashboard (/stats)

Run @promptrouter /stats at any time to open your live financial dashboard:

• 💸 Unrouted Baseline Cost: What your prompts would have cost on default premium models ($5.00/1M).
• ⚡ Routed Actual Cost: What you actually paid using PromptRouter.
• 💰 Net Dollars Saved ($): Exact savings calculated across all sessions.
• 📈 Tier Progress Bars: Visual ASCII charts showing gpt-4o-mini vs gpt-4o/claude usage ratio.
• 🔋 Monthly Quota Health: Real-time tracking of your monthly token allowance.

🎮 Slash Commands

⚙️ Model Hierarchy & Configuration

Customize your routing hierarchy via VS Code Settings (Cmd/Ctrl + , -> search "PromptRouter"), models.config.json, or a workspace .promptrouter.json file:

🔒 Privacy & Security First

PromptRouter runs 100% locally in your VS Code editor. Your prompts, code snippets, and workspace files never touch any third-party telemetry servers.

Made with ❤️ by Madhav Anand.