CodeGuard AI

The Vision CodeGuard AI is an autonomous, multi-agent code health pipeline CLI. By distributing tasks among a collaborative swarm of 10 specialized AI agents, CodeGuard does not just find problems—it fixes bugs, writes tests, secures vulnerabilities, and documents code in a self-correcting feedback loop. How It Works: The 10-Agent Swarm CodeGuard orchestrates 10 specialized agents, each powered by a model optimized for that specific task: 1. CLI Agent (gemini-3-flash): Validates inputs and launches the pipeline. 2. Orchestrator (deepseek-r1): Manages agent handoffs and retry states. 3. Bug Scanner (deepseek-r1): Scans code for logic errors and edge cases. 4. Fix Generator (kimi-k2-7-code): Generates patches to resolve bugs. 5. Test Agent (nemotron-3-nano-30b-a3b): Writes unit tests (pytest) for fixes. 6. Vuln Scanner (gemini-3-flash): Scans against 530 security checks. 7. Remediator (kimi-k2-7-code): Auto-patches safe vulns; flags sensitive issues for humans. 8. Doc Writer (qwen-plus): Injects docstrings and generates markdown docs. 9. Monitor Agent (deepseek-r1): Injects runtime diagnostic hooks to log errors. 10. Validator (deepseek-r1 @ 0.1): Harsh quality gatekeeper verifying all outputs. Key Innovations • Independent Validator Loop: Worker agents are judged by a separate deepseek-r1 model. If rejected, feedback is injected for retries (up to 3 times) to ensure only clean code moves forward. • Dual Execution Modes: Supports Band Mode (persistent WebSockets via Band SDK) and Local/Offline Mode (deterministic execution with zero API credits). • Resilient Fallbacks: Auto-routes failed model calls to gemini-3-flash to prevent pipeline disruption. The Output CodeGuard outputs a structured directory containing step-by-step patched code, detailed JSON reports, validation logs, and auto-generated markdown documentation.