Skip to main content

Dashboard

AlphaDiana includes a local React and FastAPI application for launching and monitoring evaluations without editing YAML by hand. It can browse saved runs, inspect task trajectories, compare runs, follow job logs, and manage supported ROCK sandbox workflows.

The dashboard is a local operations interface, not the static website homepage or the runner's status/dashboard.txt terminal-status artifact.

Install

From the repository root, activate the AlphaDiana environment and install the dashboard dependencies:

source scripts/activate.sh
pip install -e '.[dashboard]'

cd alphadiana/analysis/dashboard/frontend
npm install
cd ../../../..

Development mode starts Vite and FastAPI separately and reloads source changes. Production mode builds the frontend before serving it from FastAPI.

Start

cd alphadiana/analysis/dashboard
./run.sh

The launcher prints the selected URLs. It prefers ports 5173 (frontend) and 8000 (backend), then chooses the next available local port if either is busy.

To build and serve the frontend and API together:

cd alphadiana/analysis/dashboard
./run.sh --prod

Both modes bind the backend to 127.0.0.1. The application has no authentication layer, so do not expose it on a public interface or through an unauthenticated reverse proxy.

Data and runtime configuration

By default, the launcher reads and writes the checkout's results/ and configs/ directories. Override those locations before launch when needed:

export ALPHADIANA_RESULTS_DIR=./results
export ALPHADIANA_CONFIGS_DIR=./configs

The run list is rebuilt from saved result JSONL files. Job metadata and logs are persisted under the results directory. After a backend restart, job history is restored and formerly active jobs are marked interrupted; their old worker processes do not resume automatically. Use the Jobs page's resume flow to continue from the saved checkpoint.

The launcher loads checkout-local ROCK port settings when available. Sandbox deployment controls additionally require the ROCK SDK and healthy checkout-owned ROCK services; browsing DirectLLM results does not.

Pages

  • Runs lists saved evaluations and opens task-level details and trajectories.
  • Compare aligns selected runs for side-by-side inspection.
  • Jobs shows evaluations submitted through the dashboard, progress, and logs.
  • New Evaluation creates DirectLLM or supported OpenClaw jobs from a form.
  • Sandbox controls list, probe, and deploy supported ROCK sandboxes.

For result formats, lifecycle events, and logprob capture, see Observability & Proxies. For provider and ROCK setup, start with Installation.