Contributing to RaphaelLoop

Thank you for helping make long-running AI coding work safer and easier to verify.

Workflow

1. Fork the repository and create a focused branch.
2. Keep each Agent role responsible for one outcome. Update the Codex and Claude definitions together when a role contract changes.
3. Add or update deterministic checks for executable behavior. Runtime state changes require a failing transition test before implementation.
4. Run bash scripts/validate.sh before opening a pull request.
5. Describe the user problem, behavioral change, validation evidence, and remaining risk in the pull request.

Keep runtime files inside skills/raphael-loop/. Repository-facing README files, architecture material, tests, and contributor documentation must remain outside that directory so Agent Skills installers copy only the executable Skill payload. When behavior changes, update SKILL.md, the runtime reference, both README languages, architecture/security guidance, and the matching examples in the same pull request.

Use Conventional Commits, for example feat: add bounded recovery routing or fix: preserve agent backups during install.

Contribution boundaries

• Do not commit credentials, private logs, internal hosts, personal paths, active .raphael-loop/ state, or proprietary prompts.
• Do not weaken success predicates, quality thresholds, approval gates, or stop conditions to make a loop pass.
• Keep the kernel zero-dependency and local: no shell command strings, network clients, credential integrations, or automatic side effects. Any proposed expansion requires an explicit security design.
• Prefer small, reviewable changes over broad role or documentation rewrites.

By contributing, you confirm that you have the right to publish your contribution under the MIT License.