helixir CLAUDE.md

Agent Behavioral Rules

Non-Negotiable Rules

These rules are enforced by hooks and cannot be overridden by any agent instruction:

• NEVER use --no-verify to skip git hooks — hooks are safety gates, not obstacles
• NEVER commit secrets, credentials, API keys, or tokens to any file
• NEVER force-push to main, master, staging, or production branches
• NEVER push without all applicable quality gates passing (run pnpm preflight or equivalent)
• NEVER install packages without verifying they exist in the npm registry first

Verification Requirements

• Read files before editing them — understand existing code before modifying
• Verify package existence before installing: npm view <package> or check npmjs.com
• Confirm current state before claiming status — check git, files, build output
• Check tool availability before assuming it is installed

Attribution

Do NOT include AI attribution in commits, PR bodies, code comments, or any content. When block_ai_attribution is enabled in .rea/policy.yaml, the commit-msg hook REJECTS commits containing structural AI attribution (Co-Authored-By with AI names, 'Generated with [Tool]' footers, etc.). The attribution-advisory hook also blocks gh pr create/edit and git commit commands with attribution. You must remove all attribution markers before committing — the hooks will NOT silently fix them.

Human-in-the-Loop Escalation

When you encounter an unexpected blocker, ambiguous requirement, or situation not covered by the current context:

1. STOP — do not invent a workaround or make assumptions
2. Describe the situation clearly: what you tried, what failed, what you need
3. Wait for explicit human instruction before proceeding

The cost of pausing is always lower than the cost of acting incorrectly.

Policy File

Read .rea/policy.yaml at the start of every session to confirm: