Build & publish a tool

A tool is one Python file and a manifest. This page outlines the author’s loop; the full guide is in preparation.

Guide in preparation

Start from the scaffold

plugin_scaffold writes a starter tool that already passes every gate.
Write run(ds, params)

Read the dataset through ds, and return the outputs the manifest declares.
Declare the manifest

Each typed parameter becomes a control; requirements decide which datasets offer the tool.
Validate

plugin_validate runs the nine gates locally and names the one that failed.

Nine gates
Pack

plugin_pack writes a single .pvtool bundle.
Publish Coming soon

The store runs the gates again and lists the tool only if every required gate passes.

Nine gates, again
Install

The viewer runs the gates once more before it installs.

Nine gates, once more

Write one with a coding agent

A guide written for coding agents carries the contract the gates check.

What each gate checks, in the order they run.

#	Gate	What it checks
1	bundle	The archive’s layout, size and file types.
2	manifest	Every field of `plugin.json`, defaults included.
3	entry	`tool.py` imports and defines `run(ds, params)`.
4	imports advisory	Reports network, subprocess, file-write, dynamic-code and environment use, with file and line. Advisory: it never blocks.
5	fixture	Runs on a synthetic dataset at its own defaults, within its time and memory limits.
6	result	Returns the declared outputs and no others, each with values to draw.
7	determinism	Two runs with the same parameters give the same result.
8	edge	Runs with every parameter at its minimum, then at its maximum, without crashing.
9	no_mutation	Writes nothing into the dataset.

A tool is arbitrary Python; read the safety note before you install one.