PactKit

PactKit is a Python toolkit designed for spec-driven agentic development and governance of AI coding assistants. —Prompt, AI, Code, and Truth. This contract ensures that instructions are separated from state, creativity is distinct from logic, and data is the definitive source of truth, with code serving as the executor of all deterministic operations.

The toolkit structures development workflows by assigning dedicated agents to each phase of the Plan-Act-Check-Done (PDCA) lifecycle. Each of the nine agents has strictly defined capabilities and limitations, such as the system architect writing specifications without coding, the developer writing code and tests but not modifying specs, the QA engineer running tests but not altering code, and the repo maintainer committing changes without writing features. This agent-based architecture is designed to eliminate overlap and ambiguity in responsibilities, supporting a transparent and auditable development process.

Test-driven development (TDD) is a core principle in PactKit, with the TDD loop operating only on tests created within the current story. Pre-existing tests are protected; if any fail during development, the agent halts and reports the issue rather than making silent modifications. This approach reinforces safety and reliability in the development pipeline. The toolkit also features a regression gate, ensuring that only successful test passes allow progression to the next phase.

PactKit is distributed as a Python package, installable via pip, and is available under the MIT License. It supports integration with four IDEs and provides a set of 11 commands and 10 skills to facilitate its workflows. Performance optimizations, such as lazy rule loading, have reduced token usage per turn, enhancing efficiency during agent operation. The tool is open source, with documentation and code accessible on GitHub and PyPI.