coding-agent

What This Skill Does

The Coding Agent skill is designed to offload complex coding tasks to specialized AI agents like Codex, Claude Code, or Pi. It operates via background processes, allowing for asynchronous execution of coding-related operations. This skill is your go-to for anything involving significant code manipulation, from initial feature development and application building to in-depth code reviews and large-scale refactoring. It also excels at iterative coding tasks that require exploring multiple files or directories. The agent leverages the power of bash for execution, offering flexibility through various parameters like pty for interactive sessions, workdir for scoped context, background for asynchronous tasks, timeout for process control, and elevated for host execution when permitted.

Installation

To install the Coding Agent skill, use the following command:

clawhub install openclaw/openclaw/skills/coding-agent

Use Cases

This skill is ideal for:

1. Building New Features or Applications: Delegate the creation of new code modules, components, or entire applications.
2. Reviewing Pull Requests (PRs): Spawn the agent in a temporary directory to analyze and provide feedback on code changes without affecting your current workspace.
3. Refactoring Large Codebases: Undertake substantial code restructuring and optimization efforts efficiently.
4. Iterative Coding with File Exploration: Engage in coding tasks that require navigating and understanding the context of multiple files.

Important Considerations:

• PTY Mode: Codex, Pi, and OpenCode require pty:true for interactive terminal applications. Claude Code (claude CLI) does not require PTY and should use --print --permission-mode bypassPermissions for full tool access and to avoid interactive confirmations.
• Workspace Restriction: Never spawn agents within your ~/clawd workspace. Always use temporary directories or project-specific directories outside of this restricted area.
• Simple Fixes: Avoid using this skill for trivial one-liner code edits; direct manual editing is more efficient.
• ACP Harness: For thread-bound ACP harness requests (e.g., spawning Codex in a Discord thread), use sessions_spawn with runtime:"acp" instead.

Example Prompts

1. bash pty:true workdir:/path/to/my/project command:"codex exec 'Implement user authentication using JWT in the Node.js backend'"
2. bash background:true workdir:/tmp/review-pr-123 command:"claude --print --permission-mode bypassPermissions 'Review the changes in this PR for potential security vulnerabilities and suggest improvements.'"
3. bash pty:true workdir:~/projects/legacy-app command:"pi exec 'Refactor the database connection module to use connection pooling'"

Tips & Limitations

• Background Execution: For long-running tasks, utilize the background:true parameter. This returns a sessionId that can be used with process tool actions (list, poll, log, write, submit, send-keys, paste, kill) to monitor and manage the background session.
• Temporary Directories: When reviewing PRs or experimenting, leverage mktemp -d to create isolated temporary directories. Remember to cd into them before executing commands.
• Agent Specific Flags: Pay close attention to the specific flags required by each agent (e.g., pty:true for Codex/Pi/OpenCode, --print --permission-mode bypassPermissions for Claude Code).
• Security: Be mindful of the commands you execute, especially when using elevated:true. Running agents in sandboxed environments or temporary directories is generally safer.
• Code Execution: This skill inherently involves code execution. Ensure you understand the implications and potential risks, particularly when running commands with elevated privileges or in production environments.