cursor-cli-agent_skill

Overview

This skill delegates complex coding tasks to the Cursor Agent CLI, running interactive agent sessions to implement, refactor, or review code. It is built for multi-file, iterative work and supports background sessions, isolated git worktrees, and non-interactive print mode for scripting. PTY access is required for correct operation.

How this skill works

The skill invokes the Cursor CLI via a bash tool with pty:true so the interactive terminal behaves correctly. You provide a workdir and a command (or run background:true for long jobs); the skill can monitor sessions via the process tool, send stdin, and kill or poll sessions. It supports modes like --plan (read-only), --yolo (auto-approve), --print (non-interactive/JSON), and worktree isolation for parallel tasks.

When to use it

• Building or creating new features or applications that need file edits and iterative exploration.
• Reviewing pull requests in a temporary clone or temp directory (use plan mode for read-only reviews).
• Refactoring large codebases where the agent needs to open, edit, and commit multiple files.
• Long-running or parallel tasks that you want to run in background with session monitoring.
• Scripting or CI tasks that need parseable, non-interactive output via --print.

Best practices

• Always set pty:true on the bash tool — Cursor will hang without a PTY.
• Specify workdir so the agent is scoped to the intended project; never spawn agents in the ~/.openclaw/ or ~/clawd workspace.
• Use --plan for reviews and safety when you want read-only analysis before applying changes.
• Use background:true for lengthy tasks and monitor with the process tool (list, log, poll, submit, kill).
• Use -w (worktree) to isolate parallel jobs and avoid interfering with each other.
• Prefer --print for CI/CD or automation where machine-readable output is required.

Example use cases

• Refactor the auth module across multiple files using background mode and --yolo for fast execution.
• Batch-review several PRs by cloning into a temp directory and running agent --plan for each PR.
• Run a non-interactive codebase summary in CI: agent --print --output-format json 'List all API endpoints'.
• Create parallel issue branches with git worktree and launch an agent in each worktree to fix and push changes.
• Start a long-running feature build in background and receive a notification or poll logs until completion.