coding-agent-backup_skill

Overview

This skill delegates coding tasks to Codex, Claude Code, or Pi agents via a background bash-driven process. It focuses on interactive terminal workflows, long-running automation, and safe review/refactor patterns while enforcing PTY usage and workspace isolation. Use it to run agents that explore files, make changes, and report progress back to you.

How this skill works

All agent runs are executed through a bash tool that must allocate a pseudo-terminal (pty:true). For one-shot work you can run exec-style commands inside a git repo; for long or autonomous tasks you start the agent in background mode and monitor it with process actions (log, poll, list, write, submit, kill). The agent sees only the provided workdir so operations stay scoped and auditable.

When to use it

• Building or creating new features or apps that require file exploration and iterative edits
• Reviewing pull requests — spawn the agent in a temporary clone or git worktree, never the live project folder
• Refactoring large codebases or running multi-step transformations that need background monitoring
• Running parallel fixes using separate worktrees for each issue
• Not for tiny one-line fixes or simple edits — edit those directly instead of spawning an agent

Best practices

• Always set pty:true — coding agents are interactive terminals and will break without a PTY
• Provide a focused workdir; use mktemp/git init for scratch repos when needed
• Run longer jobs with background:true and monitor via process actions instead of polling stdout continuously
• Never start agents inside protected/live folders (e.g., ~/clawd or any live OpenClaw project)
• Use --full-auto for building tasks and vanilla flags for careful reviews; prefer safety flags when unsure
• Append an auto-notify wake trigger so you get immediate alerts when background jobs finish

Example use cases

• Spawn Codex in a temp git repo to prototype a new CLI feature and return when tests pass
• Clone a PR into /tmp and run Claude Code to perform an automated review, then post results as a GitHub comment
• Create multiple git worktrees and run parallel Codex sessions to fix a batch of issues concurrently
• Run a Pi agent in background to refactor a module, send input when it asks for decisions, and collect the session log when done