Skip to content

CLAUDE.md

Latest commit

 

History

History
26 lines (20 loc) · 1.21 KB

CLAUDE.md

File metadata and controls

26 lines (20 loc) · 1.21 KB
Raw

CLAUDE.md — AI Warden

Zero-Trust Docker sandbox for AI coding agents. Public repo.

Session loop (do this every time)

  1. Start: read .ai/HANDOFF.md. Then verify it against reality (git status/fetch, latest tag, gh run list, docker info). Reality wins — fix the handoff if it is wrong.
  2. Work: top item in HANDOFF "Next steps" unless told otherwise.
  3. End: update .ai/HANDOFF.md, rewrite .ai/NEXT_PROMPT.md (bump version, cite the event behind every change, stay under 70 lines), add a line to the prompt changelog, commit and push.

Non-negotiables

  • Done means run. ./scripts/verify-isolation.sh must pass phases A–D. New behaviour gets a drill.
  • Recurring bug shape: a control that reports itself armed while enforcing nothing. Test the "silently did nothing" case of every guard you touch.
  • Security fixes ship in a tag, not just on main.
  • Commit messages via git commit -F <file>; explain why and what was run.
  • Docs state limitations plainly (docs/THREAT_MODEL.md §4).

The Gotchas table in .ai/HANDOFF.md lists the Windows/Git Bash traps that have cost the most time. Read it before scripting through Bash.