| name | safety-rules |
| description | Multi-agent coordination rules for fleet-generator. Prevent overwrites, conflicts, and lane crossings. Always active. |
| when_to_use | Always active. Trigger phrases: safety rules, lane discipline, don't overwrite, conflict, in another agent's lane, mission scope is a boundary. |
| status | active |
| tags | ["skill"] |
| updated | 2026-05-07T00:00:00.000Z |
Safety Rules
These rules apply at all times to every agent on fleet-generator. They
exist because previous violations broke things; each rule is the
shape of a past incident.
1. SPEC-BEFORE-REALITY
Grep symbols, read files, list directories before writing claims
about existing code. Memory is not a source of truth.
Before any agent claims "the type is X" or "the field is Y," verify
or stay silent. No "I'm pretty sure it's…"
2. LOOK-BEFORE-WRITE
List the target directory as step zero of any write. Catch
collisions and stale state before they ship.
3. WRITE-SURFACE
Each agent has a defined write surface (its lane). The orchestrator
writes mission files. The worker writes code. Read-only agents write
nothing. Crossing surfaces requires explicit human approval.
4. NO-OVERWRITE
If a file was recently modified by another agent (check
git log --oneline -5 <file>), verify before changing it. Two
agents should not modify the same file simultaneously.
5. MISSION-SCOPE-IS-A-BOUNDARY
Only modify files listed in the mission. Everything else is
off-limits, even if "while you're in there" looks tempting.
6. ASK-WHEN-CROSSING-LANES
If your task requires work outside your role, stop and ask the
human. Don't quietly extend your lane.
7. REPORT-CONFLICTS-IMMEDIATELY
If you find a conflict with another agent's work, stop and report it
rather than guessing how to resolve it.
How agents use these rules
When an agent is about to violate one of these rules, the right move
is stop and ask the human. Safety rules are observed-failure
rules — they exist because someone violated them and broke
something. Don't reinvent the failure.
