minius_codex_lab AGENTS.md

Maintainer instructions

Purpose

minius_codex_lab is the public source and distribution repository for an independent legal-work workspace for Codex. It is not an official product, system, or position of the Ministry of Justice of Ukraine, OpenAI, or any government body. Never add real matters, client documents, operational memory, logs, generated legal work, personal data, restricted information, or secrets to this repository.

Repository map

• .agents/skills/ — canonical skill sources.
• .codex/agents/ — canonical role sources.
• workspace-template/ — runtime-only files and empty/seed data structures.
• tools/ — deterministic document tooling.
• scripts/ — validation, safety, and release tooling.
• tests/ — unit, lifecycle, safety, and reproducibility tests.
• docs/ — architecture, security, compatibility, development, and release policy.
• dist/ — generated release artifacts; never commit it.

The root AGENTS.md governs maintainers. The runtime instruction shipped to users is workspace-template/AGENTS.md; the release builder maps it to the root of the generated workspace.

Development commands

python3.11 -m venv .venv
source .venv/bin/activate
python -m pip install "ruff>=0.11,<1" "pytest>=8,<10"
python -m pip install -r workspace-template/requirements.txt

ruff check .
ruff format --check .
python3 scripts/validate_workspace.py --mode upstream
python3 scripts/check_repo_safety.py --profile upstream-public
python3 -m unittest discover -s tests -v
python3 -m unittest discover -s tools/verifiable_document/tests -v
python3 -m pytest
python3 scripts/build_release.py
python3 scripts/build_release.py --check-reproducibility

Change rules

• Keep skills canonical in .agents/skills/; every skill needs explicit trigger/non-trigger conditions, inputs, outputs, workflow, evidence and safety rules, stop conditions, definition of done, and a test fixture or documented acceptance test.
• Keep roles canonical in .codex/agents/; define a narrow responsibility, allowed and prohibited actions, write scope, and handoff format. Research and audit roles are read-only by default.
• Treat hooks, rules, configuration, manifests, and the release allowlist as security-sensitive. Review them before execution and test behavioral changes.
• Never edit generated ZIPs or checksums. Change source files and rebuild.
• A repeated procedure belongs in a skill; a durable cross-cutting rule belongs here; an independent specialization belongs in a role; a major architectural decision needs an ADR or RFC.
• Add a reproducing test or fixture before fixing a behavioral defect. Update documentation and CHANGELOG.md whenever behavior or a public contract changes.