EvoLLM Dashboard

Real-time monitoring and exploration dashboard for LLM-driven code evolution frameworks.

Built for OpenEvolve and ShinkaEvolve, with a plugin system for custom adapters.

Features

• Multi-framework support — auto-discovers OpenEvolve checkpoints and ShinkaEvolve databases; extensible via entry-point plugins
• Live experiment tracking — file-system watching (watchdog) and SQLite polling for real-time updates over WebSocket
• Program browser — paginated, filterable table with code viewer (Monaco), diff viewer, and per-program metrics
• Genealogy tree — interactive lineage visualization with golden-path highlighting from initial seed to best program
• Island topology — island state overview, migration event timeline, and per-island score histories
• Fitness analytics — best/mean score curves, score distributions, improvement rates, programs-per-minute
• LLM analytics — cost tracking, model usage breakdown, posterior evolution over time, patch-type distribution
• Embedding explorer — 2D/3D scatter plots of program embeddings with cosine similarity heatmap
• Conversation history — LLM prompt/response pairs with code diffs and improvement deltas
• Experiment comparison — side-by-side metrics for multiple runs
• Dark theme — full dark UI with framework-specific color badges

Quick Start

# Clone
git clone https://github.com/NitroxHead/evollm-dashboard.git
cd evollm-dashboard

# Install backend
pip install -e .

# Install frontend
cd frontend && npm install && cd ..

# Run (two terminals)
python3 -m uvicorn backend.main:app --host 127.0.0.1 --port 8001
cd frontend && npx vite

Open http://localhost:5173. The dashboard scans its own directory for experiments on startup and every 30 seconds.

Project Structure

evollm-dashboard/
├── backend/
│   ├── main.py              # FastAPI app, CORS, lifespan
│   ├── config.py            # Scan intervals, thresholds, ports
│   ├── adapters/
│   │   ├── base.py          # FrameworkAdapter ABC
│   │   ├── shinka_adapter.py
│   │   ├── openevolve_adapter.py
│   │   └── registry.py      # Plugin registry + entry-point loader
│   ├── models/
│   │   ├── unified.py       # Pydantic models (programs, experiments, metrics)
│   │   └── api.py           # Response schemas
│   ├── services/
│   │   ├── experiment_manager.py
│   │   └── change_detection.py
│   └── api/                  # REST + WebSocket endpoints
├── frontend/
│   └── src/
│       ├── pages/            # 12 route pages
│       ├── components/       # Visualizations, layout, programs
│       ├── hooks/            # React Query + WebSocket hooks
│       ├── stores/           # Zustand stores
│       └── lib/              # API client, utilities
└── configs/projects/         # Built-in framework configs

Tech Stack

Layer	Stack
Backend	FastAPI, uvicorn, Pydantic, watchdog
Frontend	React 18, Vite, Tailwind CSS v4
Visualizations	Recharts, D3.js, Monaco Editor
State	Zustand, TanStack React Query
Transport	REST + WebSocket

Plugin System

Custom framework adapters can be installed as Python packages using the evollm.projects entry point group.

# pyproject.toml of your adapter package
[project.entry-points."evollm.projects"]
my_framework = "my_package.config"

Your config module exports a PROJECT_CONFIG dict:

PROJECT_CONFIG = {
    "name": "my_framework",
    "adapter_class": MyAdapter,        # subclass of FrameworkAdapter
    "detect": detect_fn,               # (path: str) -> bool
    "glob_patterns": ["**/data.json"],  # discovery patterns
    "resolve_experiment_path": resolve_fn,  # match -> experiment root
    "display_name": "My Framework",
    "badge_color": "#f59e0b",
    "badge_bg": "rgba(245, 158, 11, 0.15)",
    "change_detection": "poll",        # "poll" or "watchdog"
}

Install with pip install -e ./my-adapter and the dashboard auto-discovers it on startup.

API

All endpoints are under /api:

Endpoint	Description
GET /api/health	Status and experiment count
GET /api/frameworks	Registered framework metadata
GET /api/experiments	List all experiments
GET /api/experiments/{id}	Experiment details
GET /api/experiments/{id}/programs	Paginated program list
GET /api/experiments/{id}/programs/{pid}	Single program
GET /api/experiments/{id}/search?q=	Code search
GET /api/experiments/{id}/metrics	Aggregated metrics
GET /api/experiments/{id}/islands	Island states + migrations
GET /api/experiments/{id}/lineage	Genealogy tree
GET /api/experiments/{id}/conversations	LLM conversations
GET /api/experiments/{id}/analytics	Cost and model analytics
WS /ws/{id}	Real-time updates

Configuration

Key settings in backend/config.py:

Setting	Default	Description
SCAN_INTERVAL	30s	How often to re-scan for experiments
STATUS_RUNNING_THRESHOLD	60s	Modified < 60s ago = running
STATUS_PAUSED_THRESHOLD	600s	Modified < 10min ago = paused
SQLITE_POLL_INTERVAL	2s	SQLite change detection polling
DEFAULT_PAGE_SIZE	50	Default pagination size

The backend port defaults to 8001 (set via DASHBOARD_PORT env var). The frontend Vite dev server proxies /api and /ws to the backend.

License

MIT