code-safe
code-safe enforces careful, step-by-step analysis before any code changes — for critical paths, uncertain changes, or complex refactoring.
Core principles — non-negotiable
1. Understand first, code later
- Analyse until you fully understand. No partial understanding proceeds.
- Never assume. Verify with code reads.
- Unclear → STOP and ask. Don't guess.
2. Mandatory pre-coding analysis — all three blocks
| Block | Content |
|---|---|
| What you understood | Problem restated in your own words + why this needs fixing (with evidence) |
| Where the issue is | EXACT file(s) and line(s) — no vague references. What is currently wrong/missing (cite code) |
| How you propose to correct it | Solution approach + BEFORE/AFTER code blocks (concrete, not pseudocode) + WHY this is correct |
3. Explicit permission — zero tolerance
- Writing code without "go ahead" / "implement it" = violation
- Change ONLY the parts approved. Nothing adjacent. Nothing "while I'm here"
- Additional issues found mid-work → STOP, present new analysis, wait for new permission
4. No business logic assumptions — hard ban
- Never add business rules based on guesses
- Never add conditional logic without explicit approval
- "Probably needs a check" → STOP. Ask Bruno.
5. Scope control — hard ban
- Approved scope is the ceiling, not the floor
- Discover related issue → STOP. Don't "fix while here". Report and wait.
Workflow
User request → analysis → present findings → Q&A → agreement → authorization → implementation → summary
Discovery mid-implementation — hard stop
Additional issue discovered mid-work triggers MANDATORY HALT:
- STOP at that exact point. Don't finish the "current" edit
- NEVER "fix while you're there". That's the failure this mode exists to prevent
- Report findings. Wait for new authorization. New issue = new analysis cycle
One change at a time
- Complete one change fully before proposing the next
- Each change has its own analysis → permission → implementation cycle
Input handling
- With argument → Code-Safe applies ONLY to those files. Read them first, then wait for task description.
- No argument → Code-Safe is global for the session until released.
Related
- brainstorm — even safer, no code at all, just exploration
- fix — disciplined fix flow with dependency mapping (similar gates)
- confirm — verbal alignment without any tool calls at all