stepix

stepix makes the setup instructions in your README exact and runnable.

The problem

Research repos often ship their setup as a README ritual: install this, download that, preprocess, train. Those instructions are written for a person, not a computer. Steps are implied, paths and versions are left to the reader, and nothing says when a step is done or what it needs first. Every reader, human or agent, interprets them a little differently, so the same README gives different setups.

What stepix does

You keep the README. Under each instruction you add a small stepix block that defines the step precisely: its command, environment, inputs, outputs and what it depends on. uvx stepix run then executes the README as written, the same way every time. All you need is uv on Linux or macOS; there is nothing to install.

• Exact steps. Every step is fully defined; nothing is left to interpretation. Ensuring reproducibility and saving your time.
• Separate environments where needed. By default a step runs in the repo's environment. A step that needs different packages gets its own: for example, one step can use numpy<2 while the rest of the repo uses numpy>=2.
• Caching where you ask for it. A step with cache = true is skipped once it has succeeded, so a long download or training doesn't run twice. Every other step runs each time, and --no-cache runs them all.

Try it: this README is a pipeline

git clone https://github.com/stepankonev/stepix && cd stepix
uvx stepix run ./README.md  # downloads an image, then rotates it by a random angle
uvx stepix run ./README.md  # the download is cached; the image gets a new angle

It runs the two steps below; their code is in example/.

[params]
url = "https://raw.githubusercontent.com/scikit-image/scikit-image/v0.24.0/skimage/data/chelsea.png"
output = "out/rotated.png"

1. Download the image

example/download.py saves the image and ends with output(image_path=target), which is how the next step gets {steps.download.image_path}. With cache = true it downloads once, and later runs skip it.

[steps.download]
run = "python example/download.py {params.url} --out data"
cache = true
retries = 2

2. Rotate it

This step needs OpenCV, so it runs in its own environment, example/cv, pinned by its uv.lock. It isn't cached, so it runs every time, with a new angle.

[steps.rotate]
env = "example/cv"
run = "python example/rotate.py --in {steps.download.image_path} --out {params.output}"
requires = ["download"]

Steps and params

A ```stepix block is TOML, and its first line says what it holds: [steps.NAME] starts step NAME, and [params] holds values that commands use as {params.NAME} (one [params] per file).

In run, {steps.STEP.NAME} is a value from a step it requires. A step outputs one with output(NAME=value) in Python (from stepix import output), or from any language by printing ::stepix::output NAME=value. A step with cache = true is skipped when its last run succeeded, unless a param it uses changed, a step it requires ran, or you pass --no-cache. Every other step runs each time.

Commands

uvx stepix run                   # run README.md here (or: uvx stepix run path/to/file.md)
uvx stepix run --no-cache        # run cached steps too
uvx stepix run --step rotate     # only this step
uvx stepix run --from download   # this step and every step after it
uvx stepix validate              # check the README without running anything

--step and --from only choose steps; a step that is cached still skips unless you add --no-cache (uvx stepix run --step rotate --no-cache). Commands run from the file's folder, and state goes to .stepix/ next to it. Exit code 0 ok, 1 a step failed, 2 it can't start (the output says why), 130 interrupted.

With agents

stepix never prompts, so an agent runs it the way a person does. When a run starts, it prints where the logs go, .stepix/runs/<run>/logs/<step>.log; each log fills while its step runs, and .stepix/runs/<run>/manifest.json holds every step's state. To follow a long run, an agent reads those files.