LangGraph for Spack

A LangGraph-based prototype for generating Spack package recipes from a source repository URL.

Workflow Diagram

View the workflow diagram

Specifications Document

View the specifications document

Overview

The project analyses a software repository, extracts its package metadata, resolves its source archive, and generates an initial Spack package.py recipe.

The current implementation is deterministic and rule-based. LLM-assisted recipe modelling, validation, and repair will be added in later development phases.

Current workflow

The prototype currently:

• clones the provided source repository;
• inspects the repository files;
• detects the project type and build system;
• extracts package metadata;
• identifies package dependencies;
• resolves a source archive;
• generates a Spack package recipe;
• writes the generated files to the outputs/ directory.

Project structure

• graph.py — defines the LangGraph workflow and transitions.
• nodes.py — contains the workflow node implementations.
• state.py — defines the shared state used by the graph.
• tool.py — provides the command-line entry point.
• .gitignore — excludes generated files, virtual environments, caches, and local workspace files.

Usage

Activate the Python virtual environment:

source .venv/bin/activate

Run the tool with a repository URL:

python tool.py <repository-url>

Example:

python tool.py https://github.com/deisa-project/deisa-dask

Generated files

Generated files are written under:

outputs/<spack-package-name>/

The output currently includes:

package.py
metadata.json

The outputs/ directory is intentionally excluded from Git because it contains generated files.

Current limitations

• The workflow currently focuses on Python projects.
• Recipe generation is rule-based.
• Generated recipes may still require human review.
• Advanced validation against Spack is not yet fully integrated.
• LLM-assisted reasoning and confidence-based decisions are not yet implemented.

Planned development

Future work includes:

• structured recipe modelling;
• LLM-assisted metadata interpretation;
• confidence scoring and provenance tracking;
• automatic Spack validation;
• recipe repair based on validation errors;
• explicit human-review branches;
• support for additional project types and build systems.

Project status

Prototype under active development at Maison de la Simulation, CEA Paris-Saclay.