stepyard 0.1.2

Ready-to-run automations with a Python escape hatch

Stepyard

YAML pipelines. Python plugins. Runs on your machine - no server to set up, no cloud account needed.

pip install stepyard
stepyard init my-automations && cd my-automations
stepyard run hello

---

Quick example

Build a container, smoke-test it, summarise the result with an LLM, and post to Slack. One YAML file, no glue scripts.

name: deploy
description: Build the image, smoke-test it, and post an AI summary to Slack.

steps:
  - id: build
    uses: shell.run
    with:
      command: docker build -t myapp:${{ env.GIT_SHA }} .

  - id: smoke_test
    uses: http.request
    with:
      method: GET
      url: https://staging.myapp.com/healthz

  - id: summary
    uses: llm.generate            # built-in - reads OPENAI_API_KEY from the env
    with:
      model: gpt-4o-mini
      prompt: |
        Write a one-line Slack message about this deploy.
        Build exit code: ${{ steps.build.output.code }}
        Health check HTTP status: ${{ steps.smoke_test.output.status }}

  - id: notify
    uses: http.request
    with:
      method: POST
      url: ${{ env.SLACK_WEBHOOK }}
      json_body:
        text: ${{ steps.summary.output.output }}

Run it:

GIT_SHA=$(git rev-parse --short HEAD) stepyard run deploy

✓  build         12.4s
✓  smoke_test     0.3s
✓  summary        0.9s
✓  notify         0.2s

Flow completed in 13.8s

Reading step outputs. Each node has a documented output shape, referenced as ${{ steps.<id>.output... }}:

Node	output shape
shell.run	{ stdout, stderr, code }
http.request	{ status, headers, body }
llm.generate	{ output, usage, model, provider } (use ${{ steps.<id>.output.output }} for text)

Full reference: docs/nodes/builtin.md.

---

Why Stepyard?

• Flows are YAML files in your repo. Steps, conditions, loops, and retries are plain keys. Version-control them alongside your code; no proprietary DSL to learn.
• Extend with plain Python. One @node decorator turns any function into a reusable step. Inputs are type-validated automatically; plugin dependencies run in an isolated virtualenv, so they never clash with Stepyard's own.
• Nothing leaves your machine. State is stored in a local SQLite database. Data only goes out if a step in your flow explicitly sends it.
• Every run is its own process. Each flow executes in a dedicated OS subprocess, so a crash, timeout, or sys.exit in one run cannot take down the scheduler or a sibling run. See Execution model.
• Built-in scheduler, no hosted service. Add a trigger: block with cron, interval, or startup, run stepyard service start, and flows execute on schedule without a control plane.

---

Install

# from PyPI
pip install stepyard

# or, for development, with uv
git clone https://github.com/rorlikowski/stepyard && cd stepyard
uv pip install -e ".[dev,docs]"
uv run stepyard doctor      # verify the install

Requires Python 3.10+. Works on macOS, Linux, and Windows (WSL).

---

The 60-second tour

stepyard init my-automations   # scaffold flows/ + .gitignore + .stepyard/
cd my-automations

stepyard run hello             # run a flow now (in its own subprocess)
stepyard status               # see every flow and its latest run
stepyard show <run-id>        # drill into the steps of one run
stepyard logs <run-id>        # stream the captured logs
stepyard validate --all       # type-check your YAML without running it

Schedule it instead of running by hand - add a trigger and start the daemon:

name: nightly_backup
trigger:
  uses: cron
  with:
    schedule: "0 3 * * *"      # every day at 03:00

steps:
  - id: dump
    uses: shell.run
    with:
      command: pg_dump ${{ env.DATABASE_URL }} | gzip > /tmp/backup.sql.gz

stepyard service start          # run the scheduler in the background
stepyard service status

---

Execution model

Stepyard is deliberately process-isolated:

1. stepyard run <flow> spawns a fresh subprocess (python -m stepyard.engine.runner) for that single run. Its stdout/stderr are captured to .stepyard/logs/.
2. The scheduler daemon (stepyard service start) runs separately. It evaluates triggers, enqueues runs in SQLite, and a worker spawns one subprocess per run (bounded by STEPYARD_MAX_CONCURRENT_FLOWS, default 4).
3. Inside a run, steps execute sequentially. Built-in nodes run in-process; plugin nodes installed into an isolated virtualenv run in a second short-lived subprocess that talks JSON over stdin/stdout - so a plugin's dependencies can never clash with Stepyard's own.

The practical upshot: one misbehaving flow or plugin cannot corrupt the scheduler or sibling runs.

---

Write your own step

# my_plugin/nodes.py
from stepyard.sdk import node, NodeResult, NodeStatus

@node(name="math.add")
def add(a: int, b: int) -> int:
    return a + b

@node(name="files.archive")
def archive(path: str) -> NodeResult:
    if not path:
        return NodeResult(status=NodeStatus.FAILED, error="path is required")
    # ... do the work ...
    return NodeResult(status=NodeStatus.SUCCESS, output={"archived": path})

Register it via an entry point and it becomes available as uses: math.add in every flow. See docs/plugins/creating.md.

---

Documentation

Section	What's inside
Getting Started	Install, quickstart, two tutorials
Core Concepts	Flows, expressions, control flow, triggers, errors
How-to Guides	Scheduling, secrets, approvals, debugging
Built-in Nodes	Every node that ships with Stepyard
Plugin Development	Write, test, and publish plugins
CLI Reference	Every command and flag

The full documentation is published at rorlikowski.github.io/stepyard.

Browse the docs locally with uv run mkdocs serve.

---

License

MIT - see LICENSE.