repocraft

Deterministic Python generation from validated input.

repocraft has four paths:

• a short request such as triangle selects a built-in recipe and writes one .py file plus one unittest file;
• a small YAML spec becomes a complete Python project with packaging metadata, a src/ layout, tests, lint and type-check configuration, CI, CodeQL, Dependabot, and a JUnit gate that rejects skipped tests;
• an optional ask path where a model only proposes a strict {recipe, parameters} intent, which repocraft validates before writing;
• an optional craft path where a model writes the snippet and repocraft verifies it through deterministic gates before anything is written.

The first two paths are deterministic: the same input and the same repocraft version always produce byte-identical output. GitHub repository creation is a roadmap item.

Status

Version 0.4.1 moves the determinism matrix to v0.3.0, makes the cross-OS gate mandatory, covers recipe rendering and the locale/timezone axes, and gates PyPI publishing on a matching tag plus a passing test suite. No runtime behavior changes.

Version 0.4.0 adds the craft path: a model writes a free-form snippet and repocraft verifies it — python syntax, that the generated tests import the module, ruff when it is installed, and the generated tests themselves — before the confirmation prompt and the no-overwrite rules apply.

Version 0.3.0 adds the optional AI intent layer and three new recipes:

• repocraft ask sends a free-form request to an OpenAI-compatible endpoint and turns the reply into a strict {recipe, parameters} intent;
• the model never writes code or files; the intent is validated against the recipe catalog with the same rules as make;
• configuration through OPENAI_API_KEY, OPENAI_BASE_URL, REPOCRAFT_MODEL, or --endpoint and --model;
• new recipes: fizzbuzz, pascal, and password.

Version 0.2.0 added the deterministic snippet mode:

• strict JSON Schema validation for specs and recipes, plus duplicate-key rejection in YAML;
• four built-in recipes: fibonacci, factorial, prime_sieve, and triangle;
• deterministic English matching with case, accent, and suffix normalization;
• confirmation before writing, --dry-run, --param, and refusal to overwrite;
• golden snapshots and smoke tests for every generated snippet.

The project mode from 0.1.0 is unchanged: library and CLI scaffolds with golden-tested templates, CI, CodeQL, Dependabot, and coverage/skip gates.

The deterministic commands (validate, render, recipes, make) have no model or network dependency; only ask uses the network.

Requirements

• Python 3.10 or newer.

Installation

python -m pip install repocraft

Or from this checkout:

python -m pip install .

Snippet mode

List the closed recipe catalog:

repocraft recipes

Ask for a snippet. The command shows the selected recipe, its parameters, and the target files, then asks for confirmation:

repocraft make "triangle"       # triangle.py and test_triangle.py
repocraft make "factorial"
repocraft make "prime numbers"

Preview without writing anything:

repocraft make "fibonacci" --dry-run

Override a declared integer parameter. Free-text numbers are not interpreted:

repocraft make "fibonacci" --param count=20 --yes

Rules:

• only the built-in catalog is matched; unknown or ambiguous requests stop without writing anything;
• requests containing standalone digits are refused; use --param instead;
• parameters are bounded integers and are validated before rendering;
• existing target files are never overwritten;
• --yes skips the confirmation prompt;
• generated tests use the standard library: python test_<recipe>.py.

The matcher normalizes case and accents and matches word prefixes, so Triangles and draw a triangle both select the triangle recipe. It never calls a model.

Ask mode (optional)

ask is the only command that calls a model. It turns a free-form request into a validated intent and then reuses the snippet path end to end:

export OPENAI_API_KEY=sk-...            # required
repocraft ask "a tall triangle"          # proposes triangle with its default height
repocraft ask "first 20 fibonacci numbers" --yes

Rules:

• the model replies with JSON only: a recipe id and integer parameters, or null with a reason when nothing fits;
• repocraft validates the reply against the catalog; unknown recipes, invalid parameters, and malformed replies are rejected;
• the confirmation prompt, --dry-run, --param, and the refusal to overwrite behave exactly as in make;
• --endpoint selects any OpenAI-compatible base URL; --model or REPOCRAFT_MODEL selects the model;
• without OPENAI_API_KEY in the environment the command refuses to run.

Craft mode (optional)

craft is the path where the model writes the code and repocraft stays in control of quality. The model proposes a module, its source, and a unittest file; repocraft then runs deterministic gates on the proposal before anything is written:

export OPENAI_API_KEY=sk-...            # required
repocraft craft "binary search"          # proposes, verifies, asks, writes

Gates, in order:

1. python syntax of the code and the test;
2. the test must import the proposed module by name;
3. ruff on the generated files, when ruff is installed; fixable findings are auto-fixed first and the check is marked accordingly;
4. the generated unittest file must run and pass at least one test;
5. the confirmation prompt shows the code and every check that passed.

Rules:

• --endpoint, --model, REPOCRAFT_MODEL, and OPENAI_BASE_URL configure the endpoint exactly as in ask;
• the module name must be a lowercase snake_case identifier;
• only the standard library is allowed in the prompt, but the gates verify what actually runs;
• --dry-run lists the planned files after verification; the no-overwrite rule applies;
• without OPENAI_API_KEY in the environment the command refuses to run.

Unlike the deterministic paths, the written content depends on the model; repocraft guarantees only that every gate passed and that you saw the code before it was written.

Project mode

Write a spec:

schema_version: 1
project:
  name: example-cli
  package: example_cli
  description: An example command-line application.
  type: cli
  license: Apache-2.0
  author: Your Name
  version: 0.1.0
  python: ">=3.10"
  cli:
    command: example-cli
  repository:
    owner: your-user
    name: example-cli
    private: true

Validate it:

repocraft validate spec.yaml

Plan the output without writing anything:

repocraft render spec.yaml example-cli --dry-run

Render it:

repocraft render spec.yaml example-cli

render refuses an output directory that already contains files. It never overwrites or merges into existing work.

Spec fields

Field	Required	Meaning
schema_version	yes	Must be 1.
project.name	yes	Distribution name, lowercase letters, digits, and hyphens.
project.package	yes	Import package name, lowercase letters, digits, and underscores.
project.description	yes	One-line description, 10 to 200 characters.
project.type	yes	library or cli.
project.license	yes	MIT or Apache-2.0.
project.author	yes	Copyright holder and package author.
project.version	no	Semantic version, default 0.1.0.
project.python	no	>=3.10 through >=3.14, default >=3.10.
project.cli.command	for cli	Console script name. Forbidden for library.
project.repository	no	owner, name, and private (default true).

Determinism contract

For fixed input and a fixed repocraft version:

• output paths and file contents are byte-identical;
• recipe output depends only on the recipe version and validated parameters;
• recipe matching and error ordering are deterministic;
• entries are rendered in sorted order;
• files use LF line endings and a trailing newline;
• no timestamps, random identifiers, model output, or environment values are embedded;
• the generated MIT copyright line carries the author but no year.

The golden tests in tests/golden/ enforce the contract file by file, including the built-in recipe snapshots.

The ask command adds a model, but the contract still holds: the model only proposes an intent, and a confirmed intent renders exactly like make. The craft command goes further — the model writes the code — so its output is not part of the contract; its gates, the confirmation prompt, and the no-overwrite rule are.

Development

python -m venv .venv
. .venv/bin/activate  # Windows: .venv\Scripts\activate
python -m pip install -e ".[dev]"
python -m ruff check .
python -m mypy
python -m pytest --cov=repocraft --cov-branch --cov-fail-under=100

The smoke tests install the rendered fixtures into temporary target directories with --no-deps --no-build-isolation, then run the generated projects' own quality gates. Hatchling is a dev dependency for exactly that reason.

CI runs the suite on Python 3.10 through 3.14 and rejects skipped tests through scripts/check_junit.py.

Roadmap

1. Interactive wizard for people who do not want to write project YAML.
2. Free-form natural language to a full project spec through an OpenAI-compatible endpoint, with a confirmation step before any file is written.
3. GitHub repository creation and push, private by default.
4. MCP server exposing the same validated core.

Related projects

Part of the Furox scientific-computing portfolio: independently versioned projects that pair agent-facing interfaces (MCP servers and Agent Skills) with deterministic, fail-closed verification.

• scientific-computing-system — zero-dependency Python numerics with run manifests.
• scientific-computing-system-2.0 — NumPy/SciPy/pandas build with replayable fit records.
• scientific-toolkit — MCP gateway over nine research tools.
• axiomize — unit-checked modeling engine with SBML, CellML and Modelica export.
• axiomize-quantum-skills-2.0 — modeling engine plus quantum-inspired reasoning bundle.
• quantum-reasoning-skill — agent skill that keeps competing hypotheses alive.
• plan-auditor — evidence-based completion gate for coding agents.
• nomosguard — deterministic security decision layer with a hash-chained ledger.
• eq-layer — interaction-policy layer that decides how an agent speaks.
• literature-evidence-mcp — literature MCP server with a hash-chained evidence ledger.

License

MIT. See LICENSE.