🧠 JADIT Merge Minds — Devpost Submission

💡 Inspiration

During hackathons under tight deadlines, a classic bottleneck inevitably emerges: the late-integration bottleneck.

Traditionally, developers work in absolute isolation on their individual local terminals. There is no real-time, cross-workspace visibility. True collaboration is only attempted at the very end of the cycle when developers push their changes and try to merge them. This "last-mile integration" is notoriously painful:

Overlapping Merge Conflicts: Multiple developers unknowingly modify the same blocks of code, resulting in complex, manual Git triage.
Brittle Builds: Stale files, broken import dependencies, and inconsistent configurations compile perfectly fine locally but completely crash upon integration.
Communication Overhead: Valuable time is wasted back-channeling over Slack or Discord to manually coordinate who owns which file.

We wanted to build a unified workspace where multiple developers can code simultaneously on their own terms, while a centralized, real-time AI brain manages the project flow, detects collisions, auto-handles file locks, and fixes broken builds instantly.

🚀 What it does

JADIT Merge Minds is an intelligent, collaborative development ecosystem that eliminates merge conflicts before they happen. It syncs developer workspaces in real-time and introduces an autonomous Multi-Agent Council that acts as an omniscient project co-pilot.

The platform visualizes the entire codebase as a dynamic dependency map, highlighting potential structural collisions as you type. If two developers open overlapping files, the system initiates asynchronous handshakes and intelligent file locking. To remove human bias from project evaluation, JADIT continuously evaluates repository health, calculating a mathematically rigorous Release Readiness Score ($R$) based on open issues, merge requests, active editor collisions, file locks, and live CI/CD pipeline statuses—giving teams an objective, dynamic metric of how close they are to a stable release.

🛠️ How we built it

We combined robust real-time synchronization frameworks with an autonomous multi-agent registry:

Real-Time State Orchestration: We utilized Yjs and WebSockets to establish Conflict-Free Replicated Data Types (CRDTs). This instantly synchronizes cursor locations, live file modifications, and lock telemetry across all connected developer environments.
Interactive Workspace Map: We engineered an interactive, auto-laying SVG node graph that translates raw file structures and dependencies into a live visual representation. As files are edited or locked, the map renders live warning indicators (orange/red nodes) to warn developers away from high-risk areas.
The Multi-Agent Council: We implemented a structured, multi-agent model consisting of five specialized personas running on Google Cloud's Vertex AI (Gemini 2.5 Flash):
- Orchestrator: Monitors workspace state and runs active conflict analysis loops.
- Reviewer: Audits code diffs and syntax consistency in real-time.
- Security Analyst: Scans for exposed secrets or vulnerabilities pre-commit.
- Planner: Tracks task alignment with overall project requirements.
- DevOps Agent: Interfaces directly with our GitLab MCP tools to monitor pipelines and environment deployments.
Dual-Engine Resiliency: To prevent API outages from breaking a critical development sprint, we engineered a custom routing model. While it defaults to Gemini 2.5 Flash + GitLab MCP commands, it silently falls back to offline, rule-based local agents if it encounters rate limits or network drops.

⚠️ Challenges we ran into

Real-time Collision Telemetry: Catching conflict risks before a developer hits save or push required us to dynamically monitor active cursor paths. We resolved this by tying cursor coordinates and active file views directly to the WebSocket channel, feeding them straight into the AI Orchestrator’s telemetry analysis loop.
API Rate Limits & Reliability: While testing complex multi-agent queries, we repeatedly hit Vertex AI's free-tier rate limits (429 Resource Exhausted). This inspired us to build our dual-engine fallback router, ensuring premium UX reliability by instantly switching to local static rules during rate caps without breaking the developer's focus.
Complex File Lock Handshakes: Managing file permissions in a fast-paced environment without causing deadlocks or frustrating the team was tough