Skip to content

RunForge Handbook

Welcome to the RunForge Handbook — your guide to deterministic, contract-driven ML training directly in VS Code.

RunForge is a VS Code extension for push-button machine learning training. It provides deterministic, seeded runs with full provenance — every model can be traced back to the exact code, data, and configuration that produced it. No cloud dependencies, no hidden databases, no magic.

  1. Determinism — every run is seeded. Same preset + same seed + same data = same model
  2. Provenance — every run.json includes Git commit SHA, Python path, and extension version
  3. Auditability — artifacts are standard formats (JSON, joblib) saved to disk
  • Getting Started — Install, configure, and run your first training session
  • Reference — Presets, run lifecycle, artifacts, interpretability, and settings
  • Cancel and Recovery — Cancel an active training run, recover an orphaned index, and grant workspace trust
  • RunForge: Cancel Active Training — fire VS Code’s CancellationToken; Python gets a 5s grace window to write a durable .cancelled marker before SIGKILL
  • RunForge: Recover Index — walk .ml/runs/, re-append missing entries to .ml/outputs/index.json, return a structured RecoveryReport