feat: real-time web dashboard for experiment monitoring

[Death-Incarnate]

What this adds

A single-file FastAPI dashboard (dashboard.py) that gives autoresearch a proper UI for monitoring and controlling experiments in real time.

Start it

cd autoresearch
uv run dashboard.py
# open http://localhost:7788

Set REPO_DIR env var to point at any autoresearch clone. Set PORT to change from 7788.

---

Tabs

Overview

• KPI cards: best val_bpb, baseline, improvement %, total/kept/discarded run counts
• val_bpb scatter chart colored by status (keep=green, discard=red, crash=gray)
• Status pie chart
• Live GPU stats: utilization %, VRAM used, temperature (polls nvidia-smi)

Live Training

• Step progress bar with % complete
• Loss + mfu% dual-axis chart, updates every 2 seconds via polling
• SSE log stream from run.log — color-coded (amber for step lines, green for summary blocks, red for errors)
• Autoscroll toggle

Experiments

• Full results table sorted newest-first, best run highlighted in gold
• Click any commit hash → inline diff of that experiment's train.py changes

Git History

• Last 40 commits with click-to-diff

Controls

• Launch form: run tag, model picker, max experiments
• Stop button for the running experiment
• Live hyperparameter table parsed directly from train.py

Settings

• Edit program.md in-browser with save
• Read-only train.py viewer

---

Implementation notes

• Zero frontend dependencies — all HTML/CSS/JS served inline from the single Python file
• FastAPI + uvicorn only (added to pyproject.toml)
• SSE endpoint for streaming log tail to the browser
• All file reads are relative to REPO_DIR so it works with any clone location

[Death-Incarnate]

As context for why this matters right now: Alexey Grigorev just wrote a breakdown of autoresearch that's getting traction — https://alexeyondata.substack.com/p/karpathys-autoresearch-went-viral

One of the main things people hit when they clone this repo and start running experiments is zero visibility into what's happening. The training loop is a black box unless you're tailing logs manually.

The dashboard fills that gap — live loss curve, GPU stats, log stream, experiment history — all without adding any dependencies to the core research loop. It's opt-in and read-only relative to the experiment state.