SKILL.md
<!-- TOC: Quick Start | THE EXACT PROMPT | Risk Tiers | Workflow | References -->
SLB — Simultaneous Launch Button
Core Capability: Two-person rule for running potentially destructive commands from AI coding agents. When an agent wants to run something risky, SLB requires peer review and explicit approval before execution.
Why This Exists
Coding agents can get tunnel vision, hallucinate, or misunderstand context. A second reviewer (ideally with a different model/tooling) catches mistakes before they become irreversible.
Critical: Commands run in YOUR shell environment, not on a server. The daemon is a NOTARY (verifies approvals), not an executor.
Quick Start
# Install
curl -fsSL https://raw.githubusercontent.com/Dicklesworthstone/slb/main/scripts/install.sh | bash
# Initialize project
cd /path/to/project
slb init
# Start session
slb session start --agent "GreenLake" --program "claude-code" --model "opus"THE EXACT PROMPT — Basic Workflow
# 1. Run a dangerous command (blocks until approved)
slb run "rm -rf ./build" --reason "Clean build artifacts" --session-id <id>
# 2. Another agent reviews and approves
slb pending # See what's waiting
slb review <request-id> # View full details
slb approve <request-id> --session-id <reviewer-id> --comment "Looks safe"
# 3. Original command executes automatically after approvalRisk Tiers
| Tier | Approvals | Auto-approve | Examples |
|---|---|---|---|
| CRITICAL | 2+ | Never | rm -rf /, DROP DATABASE, terraform destroy, git push --force |
| DANGEROUS | 1 | Never | rm -rf ./build, git reset --hard, kubectl delete, DROP TABLE |
| CAUTION | 0 | After 30s | rm file.txt, git branch -d, npm uninstall |
| SAFE | 0 | Immediately | rm *.log, git stash, kubectl delete pod |
Essential Commands
| Category | Command | Description |
|---|---|---|
| Session | slb session start --agent <name> |
Start agent session |
| Session | slb session list |
Show active sessions |
| Request | slb run "<cmd>" --reason "..." |
Run dangerous command |
| Review | slb pending |
List pending requests |
| Review | slb approve <id> --session-id <id> |
Approve request |
| Review | slb reject <id> --reason "..." |
Reject request |
| Hook | slb hook install |
Install Claude Code hook |
| Pattern | slb patterns test "<cmd>" |
Check command tier |
Claude Code Hook
# Install hook
slb hook install
# Hook actions:
# - allow: Command proceeds (SAFE tier)
# - ask: User prompted (CAUTION tier)
# - block: Blocked, must use `slb request` (DANGEROUS/CRITICAL tier)Execution Verification (5 Gates)
| Gate | Check |
|---|---|
| 1. Status | Request must be APPROVED |
| 2. Expiry | Approval TTL must not have elapsed |
| 3. Hash | SHA-256 hash must match (tamper detection) |
| 4. Tier | Risk tier must still match |
| 5. First-Executor | Atomic claim prevents race conditions |
Emergency Override
For true emergencies, humans can bypass with extensive logging:
slb emergency-execute "rm -rf /tmp/broken" --reason "System emergency"References
| Topic | Reference |
|---|---|
| Full command reference | [COMMANDS.md](references/COMMANDS.md) |
| Pattern matching & tiers | [PATTERNS.md](references/PATTERNS.md) |
| Configuration | [CONFIG.md](references/CONFIG.md) |
| Security design | [SECURITY.md](references/SECURITY.md) |