ML training with forensic certainty.

Cancel In-Progress Training

New "RunForge: Cancel Active Training" command. Or hit the cancel button on the live progress notification. 5s graceful SIGTERM window, then SIGKILL. Cancelled runs land a .cancelled marker so the run picker can classify them.

Recover Index

New "RunForge: Recover Index" command walks .ml/runs/ and re-appends any run missing from index.json. Idempotent. Useful after a crashed write or a workspace move.

Workspace Trust Guard

Python subprocess spawn now requires vscode.workspace.isTrusted. Untrusted workspaces get an actionable SafeError pointing at the Manage Workspace Trust UI.

Live Progress Notifications

Training surfaces per-epoch progress through VS Code’s native progress notification API, with a built-in cancel button.

Hardened CSV Errors

Non-comma delimiters, non-UTF-8 encodings, all-NaN labels, single-column CSVs, and header-only CSVs each surface specific, actionable diagnostics instead of opaque pandas tracebacks.

Custom ESLint Rules

The architectural doctrines in docs/CONTRACTS.md are now mechanized as ESLint rules. No canonical-value literal duplication. No shadow types in consumer modules. Catches what humans miss.

Doctrine Documentation

docs/CONTRACTS.md codifies six architectural rules plus seven operational patterns surfaced by five waves of structured audit. Non-negotiable for any cross-domain work.

Deterministic by Design

Every run is seeded. Re-run the same preset with the same seed on the same data — you get the exact same model. No randomness outside explicitly seeded behavior.

3 Models + Profiles

Logistic Regression, Random Forest, Linear SVC. Combine with fast, thorough, or default training profiles. Hyperparameters are explicit, recorded, and never guessed.

Full Interpretability

Feature importance (RandomForest), linear coefficients (Logistic/SVC), model-aware metrics, and a unified interpretability index — all saved as versioned JSON artifacts.

1. Point at your dataset

# CSV with a "label" column
RUNFORGE_DATASET=/path/to/data.csv

# Or set in VS Code settings:
"runforge.datasetPath": "/path/to/data.csv"

2. Choose a model

// .vscode/settings.json
{
  "runforge.modelFamily": "random_forest",
  "runforge.profile": "thorough"
}

3. Train

Ctrl+Shift+P → RunForge: Train (Standard)

# Artifacts land in:
.ml/runs//
  ├── run.json
  ├── metrics.v1.json
  └── artifacts/model.pkl

4. Inspect results

Ctrl+Shift+P → RunForge: Browse Runs
# Select run → Open Run Summary

# Or view interpretability:
→ View Latest Feature Importance
→ View Latest Interpretability Index