Plugins Quickstart

retrain now supports a unified plugin workflow across swappable parts:

1. create one file
2. set one TOML key
3. run training

60-second transform plugin

Generate a scaffold:

retrain init-plugin --kind transform --name my_transform --with-test

This creates plugins/my_transform.py and an optional smoke test.

Use it in TOML:

[algorithm]
transform_mode = "plugins.my_transform.my_transform"

[algorithm.transform_params]
scale = 1.0

Advantage plugin

retrain init-plugin --kind advantage --name my_adv

[algorithm]
advantage_mode = "plugins.my_adv.my_adv"

[algorithm.advantage_params]
scale = 2.0

Full algorithm plugin

If you want to override the whole advantage pipeline:

retrain init-plugin --kind algorithm --name my_algo

[algorithm]
algorithm_mode = "plugins.my_algo.my_algo"

[algorithm.params]
alpha = 0.1

When algorithm_mode is set, it takes priority over advantage_mode + transform_mode.

Discover what is available

retrain plugins
retrain plugins --json

Plugin search paths

Default search path:

[plugins]
search_paths = ["plugins"]
strict = true

strict = true fails fast on plugin load/shape errors.