Slb

SLB - Simultaneous Launch Button

A CLI that implements a 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 approval before execution.

Why Use SLB

Coding agents can hallucinate or misunderstand context. A second reviewer catches mistakes before they become irreversible. Essential for multi-agent workflows where a bad command could destroy work.

Risk Tiers

Quick Start

Initialize in a Project

cd /path/to/project
slb init

Creates .slb/ directory with:

• state.db - SQLite database
• config.toml - Configuration
• pending/ - Pending request files

Start a Session

slb session start --agent "GreenLake" --program "claude-code" --model "opus"
# Returns session_id and session_key

Run a Dangerous Command

# Blocks until approved
slb run "rm -rf ./build" --reason "Clean build artifacts" --session-id <id>

Review and Approve

# See pending requests
slb pending

# View request details
slb review <request-id>

# Approve the request
slb approve <request-id> --session-id <reviewer-id> --comment "Looks safe"

# Or reject
slb reject <request-id> --session-id <reviewer-id> --reason "Too risky"

Commands

Session Management

# Start a session
slb session start --agent "AgentName" --program "codex" --model "gpt-5"

# List sessions
slb session list

# End a session
slb session end <session-id>

Running Commands

# Run with reason
slb run "git push --force" --reason "Force push after rebase" --session-id <id>

# Run with timeout override
slb run "terraform destroy" --reason "Tear down test env" --timeout 300s

# Dry run (classify only, don't execute)
slb run "rm -rf /tmp/test" --dry-run

Reviewing

# List pending requests
slb pending

# List all requests
slb requests

# View specific request
slb review <request-id>

# Approve
slb approve <request-id> --session-id <id>

# Reject
slb reject <request-id> --session-id <id> --reason "Reason"

TUI Dashboard

# Interactive review dashboard
slb tui

# Features:
# - See all pending requests
# - Review details
# - Approve/reject inline

Configuration

.slb/config.toml:

[risk]
# Override risk levels for specific patterns
[[risk.overrides]]
pattern = "rm -rf ./node_modules"
tier = "safe"
reason = "Node modules are regenerated"

[timeouts]
caution_auto_approve = "30s"
request_expiry = "1h"

[notifications]
agent_mail = true

Integration with Agent Mail

SLB can notify reviewers via MCP Agent Mail:

slb init --with-agent-mail

# Reviewers get inbox messages for pending requests

Audit Trail

All requests, reviews, and executions are logged:

# View audit log
slb audit

# Export audit log
slb audit --export audit.json

Status and Health

# Check SLB status
slb status

# Check system health
slb health

Example Multi-Agent Workflow

# Agent 1 wants to run dangerous command
slb run "git push origin main --force" \
  --reason "Rebase complete, need to force push" \
  --session-id agent1-session

# Agent 2 reviews
slb pending
slb review REQ-123
slb approve REQ-123 --session-id agent2-session --comment "Verified rebase is clean"

# Original command executes automatically

Command Classification

SLB automatically classifies commands based on:

• Command patterns (rm -rf, DROP, destroy, etc.)
• Arguments and flags
• Working directory context
• Historical data

Override with explicit tier:

slb run "custom-script.sh" --tier dangerous --reason "Modifies prod data"