AI Dev Council

Multi-perspective AI analysis for technical decisions in VSCode.

What It Does

Type @council <your question> in VSCode chat to get:

• Analysis from 3 different AI perspectives (pragmatic, security, DevEx)
• Debate round: Agents critique each other before final recommendation
• Confidence levels from each agent
• Key considerations and disagreements
• Clarifying questions to ask yourself

The council remembers your full conversation history, so follow-up questions work naturally.

Requirements

• VSCode 1.95.0 or later
• GitHub Copilot subscription (free or Pro)

Note: Council uses your existing Copilot subscription — no additional API keys or costs. Currently optimized for fast, low-cost models during this testing phase, so feel free to experiment freely.

Installation

Install from the VSCode Marketplace:

1. Open VSCode
2. Go to Extensions (Ctrl+Shift+X / Cmd+Shift+X)
3. Search for "AI Dev Council"
4. Click Install

Usage

1. Open VSCode chat (Ctrl+Shift+I or Cmd+Shift+I)
2. Type: @council Should I use Redis or Postgres for session storage?
3. Wait ~10 seconds for full debate response
4. Read synthesis and individual agent views
5. Ask follow-ups - context is maintained automatically

Examples

Technical decisions:

@council Should I use microservices or a monolith for my startup?
@council Is Redis or PostgreSQL better for session storage with 10k users?
@council Should I add TypeScript to this JavaScript project now?

Follow-up questions (conversational):

@council Should I use Redis for caching?
@council What about operational costs?
@council How does this scale to 100k users?

How It Works

Two-Round Debate Architecture

Round 1 — Independent Analysis:

• Alex (Pragmatic): Prevents over-engineering, questions assumptions
• Jordan (Security/Scale): Finds risks, bottlenecks, failure scenarios
• Sam (DevEx): Debuggability, maintainability, cognitive load

Round 2 — Critique: Each agent sees the others' responses and can:

• Change their recommendation
• Disagree with specific points
• Report confidence level (Low/Medium/High)

Synthesis: Fourth model merges all perspectives into one actionable recommendation, noting where agents changed minds or still disagree.

Gate Layer

Vague questions (like "Should I use MySQL?") will ask for more context before engaging the full council — saving time and giving you better answers.

Logging (Opt-in)

Enable logging to save all council interactions for analysis:

1. Open Settings (Cmd+,)
2. Search "council"
3. Check ✓ Enable Logging

Logs saved to ~/.council/logs/ in JSONL format with:

• Query and response
• All agent perspectives (initial + critique)
• Session tracking
• Timestamps and duration

Development

Want to contribute or run from source?

Building from Source

git clone https://github.com/sathvikc/ai-dev-council.git
cd ai-dev-council
npm install
npm run compile

Press F5 in VSCode to launch Extension Development Host.

Project Structure

src/
├── extension.ts          # Chat participant registration
├── council.ts           # Orchestration (2-round debate + synthesis)
├── gate.ts              # Query classification (READY/NEEDS_CONTEXT/OFF_TOPIC)
├── logger.ts            # Opt-in JSONL logging to ~/.council/logs/
├── agents/prompts.ts    # Agent and critique prompts
├── api/vscode-lm.ts     # VSCode Language Model API wrapper
├── config.ts            # Model IDs and temperatures
└── types.ts             # TypeScript interfaces

License

MIT