Skip to content

CLAUDE.md

Latest commit

 

History

History
42 lines (33 loc) · 1.82 KB

CLAUDE.md

File metadata and controls

42 lines (33 loc) · 1.82 KB
Raw

CLAUDE.md — operating rules for AI builders on this repo

Rules for Claude Code / any coding agent working in this repository.

Hard constraints (NEVER violate)

  1. core.py stays a single file. Do NOT split it into packages — the self-patching pipeline and the AST watchdog in entrypoint.sh depend on a single-file core.
  2. data/dna/SOUL.md is the identity anchor. Do NOT modify its content.
  3. SafetyManager, AsimovSafetyFilter, BarratSafetyProtocol, PolicyEngine, injection guards, and the leak filter accept hardening-only changes. Never weaken or bypass them.
  4. entrypoint.sh watchdog logic (AST check + rollback-to-last-good) — do not touch without explicit maintainer approval.
  5. Russian prompt strings inside core.py are runtime behaviour, not comments. Do not translate, rewrite, or remove them.

Workflow

  • Before committing any core.py change: python -c "import ast; ast.parse(open('core.py', encoding='utf-8').read())"
  • Run the smoke tests: pip install pytest && pytest tests/ -v
  • State files are written atomically only (_atomic_write_text, os.replace) — never add a direct write_text on JSON state.
  • After deploying: ./deploy.sh update and watch docker logs -f nagual for at least 2 minutes before calling it done.

Do NOT without explicit maintainer approval

  • Add new top-level classes or new _loop coroutines.
  • Add new LLM providers to UniversalLLMRouter.
  • Refactor the existing class hierarchy.
  • Rename Toltec concepts (assemblage point, recapitulation, tonal/nagual…).

Always welcome

  • New tests in tests/ (pytest, no heavy deps, extract units via AST — see tests/conftest.py).
  • CI improvements in .github/workflows/.
  • Documentation clarity, honest-status precision, comparison docs.
  • Closing gaps between README claims and code reality.