*Forge — pipeline canvas for the repo you already have open
Visual, block-based pipelines over your own Python functions. Every node
checkpoints its output with full provenance; change a parameter or edit a
function and only the affected subgraph reruns. Revert, and the old
checkpoint is still there — instant.
Quick start
Install the kernel into the Python environment your repo uses:
pip install starforge-kernel
# (or, from a Forge checkout: pip install -e <forge>/starforge)
Decorate a function and save:
from starforge import block
@block
def scale(data: dict, factor: float = 2.0) -> dict:
return {"values": [v * factor for v in data["values"]]}
Run “*Forge: New Pipeline”, drag blocks from the palette, wire them
up, press ▶ Run. Run again — instant. Edit the function — only it and
its descendants go stale.
What you get
Zero-template blocks — the function signature is the block: params
become typed input handles and inspector fields, T | None means optional,
tuple returns are multi-output. A CodeLens adds @block to any function.
Provenance-backed caching — Tier-2 staleness tracks your repo's import
graph; editing a helper marks its dependents stale. Configurable
(starforge.stalenessTier).
Figures inline — plt.show() renders on the node (thumbnail, carousel,
lightbox); plotly figures open live in an editor panel.
Progress + cancellation — starforge.progress(i, n, "label") drives a
live bar; Cancel kills the run process cleanly.
Comments, copy/paste, box-select, minimap navigation — the full canvas
experience, in your editor.
Team-shareable pipelines — .forge/pipelines/*.forge are plain JSON,
made to be committed; checkpoints and caches are auto-gitignored.
Requirements
Python ≥ 3.10 in the workspace with starforge-kernel installed
(starforge.pythonPath overrides interpreter discovery; the ms-python
extension's active interpreter is used when available).
Heavy libraries (pandas, matplotlib, plotly) are only ever loaded inside
run workers — the kernel itself is stdlib-only and idles at zero.
