Slb

What it does

The Slb skill implements a two-person rule for AI coding agents to prevent potentially destructive commands from being executed without peer review and approval. It acts as a “Simultaneous Launch Button” (SLB), requiring a second reviewer to explicitly approve actions like deleting files, pushing code changes, or managing Kubernetes resources. This system is designed for multi-agent workflows where automated processes could cause significant damage if not carefully controlled. Importantly, commands execute within the user's existing shell environment, inheriting configurations like AWS credentials and virtual environments.

When to use it

• Running potentially destructive commands: Before executing rm -rf, git push --force, or similar high-risk operations with an AI coding agent.
• Multi-agent workflows: In scenarios where multiple agents are running in parallel, minimizing the risk of a single faulty command impacting the entire system.
• Protecting critical infrastructure: When automated actions could affect data, code repositories, or cloud resources.
• Ensuring accountability: To create an audit trail and ensure that someone reviews potentially harmful commands before they are executed.

Key capabilities

• Tiered Risk Assessment: Commands are categorized by risk level (Critical, Dangerous, Caution, Safe) determining the number of approvals required.
• Peer Review Workflow: Facilitates a clear process for agents to request approval and reviewers to approve or deny commands.
• Client-Side Execution: Commands run within the user's existing shell environment, inheriting configured credentials and settings.
• Session Management: Allows tracking and managing active agent sessions with features like resuming after crashes and cleaning up stale sessions.

Example prompts

These are examples of how a user might interact with an AI agent equipped with Slb:

1. "Agent, please clean the build artifacts using rm -rf ./build." (The agent would then use slb run "rm -rf ./build" --reason "Clean build artifacts").
2. "Show me any pending commands requiring approval." (The reviewer would use slb pending).
3. “Approve request ID <request-id> with the comment 'Looks safe'." (The reviewer would use slb approve <request-id> --session-id <reviewer-id> --comment "Looks safe").

Tips & gotchas

• Initialization Required: The skill requires project initialization (slb init) to create necessary configuration files.
• Client-Side Dependency: Slb runs within your shell environment; it does not execute commands on a remote server. Ensure you have the correct environment configured.
• Reviewer Role: A second agent or human reviewer is essential for the approval process.