Decider toolkit

The decider toolkit (package decikit) builds decision pipelines from plain Python functions, run over polars frames or single records. A function's arguments are the columns it reads and its name is the column it writes. Steps compose with flow, dag, branch and loop; decision trees, decision tables and scorecards load from JSON as steps. Tunable values are param()s, retuned per call through a params document without rebuilding or recompiling. The same pipeline runs interpreted, or compiled with numba (per step or fused), and can be paused and stepped through in a debug session.

Install

pip install decikit                        # library and CLI
pip install "decikit[serve-starlette]"     # HTTP serving with uvicorn + starlette
pip install "decikit[serve-sanic]"         # or sanic
pip install "decikit[mcp]"                 # `decikit mcp` server for agents (fastmcp)
pip install "decikit[prove]"               # symbolic proofs (z3-solver)

From a checkout: uv sync, then uv run pytest.

Example

import polars as pl
from decikit import FusedEngine, flow, param

def ratio(income: float, debt: float) -> float:
    return debt / income

def affordable(ratio: float, limit: float = param(0.4)) -> bool:
    return ratio <= limit

pipeline = flow(ratio, affordable, name="afford")
df = pl.DataFrame({"income": [1000.0, 500.0], "debt": [200.0, 400.0]})
pipeline.run(df)                                                   # income, debt, affordable
pipeline.run(df, params={"afford": {"affordable": {"limit": 0.9}}})
FusedEngine().bind(pipeline).run({"income": 1000.0, "debt": 200.0})  # one record

CLI

decikit guide                 # print the getting-started guide
decikit template NAME [DIR]   # write a starter project
decikit build [VERSION]       # stage and warm a config version (fills the numba cache)
decikit serve                 # POST /invocations, GET /ping

Settings come from DECIKIT_* environment variables; see decikit --help.

Claude Code plugin

This repository is a Claude Code plugin marketplace. In Claude Code:

/plugin marketplace add capitec/dsp-decision-engine
/plugin install decikit@decikit

You get the decikit skill (a guide to the project layout, best practices, pipelines, checks and the config documents) and the decikit mcp server, which reads pipelines, edits trees, tables, scorecards and params with validation, and talks to the VS Code extension. The server runs through uv in the directory Claude Code starts in, so start it from your project root. Setup and troubleshooting: setup-project.md.

Documentation

• Start with decikit guide (the same text as decikit/GUIDE.md): concepts, runnable examples, project layout and common mistakes. docs/ is the docs site.
• docs/architecture/: how it works (steps and IR, engine, configurable steps, serving, lab); published on the docs site.
• docs/design_decisions/: why it is built this way: decisions, experiments and known limitations.
• The package docstrings (help(decikit)) are the user documentation.

Contributing

See CONTRIBUTING.md. Run uv run pytest before opening a pull request.

License

MIT. See LICENSE.