Skip to content

Your first reviewed flow

The normal kazeflow workflow is short:

  1. Write ordinary Python functions.
  2. Inspect the selected work.
  3. Explicitly decide to execute it.
  4. Review the structured result.

1. Install

kazeflow supports Python 3.10 through 3.13. Check the interpreter, then create an isolated environment:

python3 --version
python3 -m venv .venv
source .venv/bin/activate

py --version
py -m venv .venv
.\.venv\Scripts\Activate.ps1

Install and verify the core. It has no required third-party runtime dependency:

python -m pip install kazeflow
kazeflow --help

Install Rich only when you want the optional progress display. Keep the quotes in shells such as zsh:

python -m pip install "kazeflow[tui]"

2. Declare a flow

Create daily.py:

daily.py
from kazeflow import Flow, asset


@asset
def fetch() -> str:
    return "report input"


@asset
def publish(fetch: str) -> None:
    print(fetch)


flow = Flow(["publish"])

A parameter named fetch declares that publish depends on the fetch asset. The module-level flow chooses publish as the default target. Asset bodies remain ordinary functions that can be unit-tested directly.

3. Inspect without invoking assets

$ kazeflow plan daily.py
Plan: publish
2 assets · no partition selection · default concurrency
Graph:
  fetch
  fetch --> publish *

The * marks the selected target. Planning validates the dependency closure and configuration without invoking either decorated asset body.

4. Approve and run

$ kazeflow run daily.py
Planned run:
Plan: publish
2 assets · no partition selection · default concurrency
Graph:
  fetch
  fetch --> publish *
Proceed? [y/N] y
Run result:
- run_id: <run-id>
- status: success
- duration: <duration>
Tasks:
- ✓ fetch (success; <duration>)
- ✓ publish (success; <duration>)

Run IDs and durations vary. In CI or a pipe, make the same decision explicitly:

kazeflow run daily.py --yes

Use --format json when another program needs one portable result document rather than human-oriented text. Portable failure messages and tracebacks can contain application values, so JSON is structurally limited but not automatically redacted.

5. Review the before and after

Value Purpose
FlowPlan Targets, dependencies, partition selection, and normalized configuration before execution.
RunResult Terminal flow, task, and attempt statuses, timings, in-memory outputs, and failure metadata.
Logs and events Optional progress and diagnostic observations; they replace neither the plan nor the result.

The same model is available from Python:

from kazeflow import run

plan = flow.plan({"max_concurrency": 2})
for task in plan.tasks:
    print(task.name, task.dependencies, task.partition_keys)

result = run(["publish"], {"max_concurrency": 2})
print(result.status, result.duration)

Loading is still Python execution

Module top-level code, imports, and an explicitly selected factory can run while the CLI loads an entry. Plan review is not a sandbox or an automatic approval.

Next steps