返回 Skill 列表
extension
分类: 开发与工程无需 API Key

coding-agents

生成并控制外部编码代理(Claude Code、Codex、OpenCode、Pi),以分配任务。在以下情况下使用:(1) 用户要求运行另一个编码代理;(2) 将子任务分配给不同的AI工具;(3) 并行运行多个代理处理不同问题;(4) 在隔离环境或工作树中生成代理;(5) 用户提到“codex”、“opencode”、“pi agent”或运行“另一个claude”。

person作者: jakexiaohubgithub

Coding Agents

Orchestrate external coding agents programmatically. Use this when you want to delegate tasks to other AI coding tools.

Prefer interactive mode

Interactive sessions via tmux should be the default for most tasks. Background mode is only appropriate for the simplest tasks that require very few commands or edits (e.g., "count the files in src/", "what version of node is this project using?").

Why interactive is better:

  • You can observe progress and catch issues early
  • You can respond if the agent asks clarifying questions
  • You can interrupt if the agent goes off track
  • Agents often need multiple turns to complete real work
  • Debugging is much easier when you can see what happened

Use background mode only when:

  • The task is trivial (one or two simple commands)
  • You are confident the agent will complete without interaction
  • You are running many parallel tasks and cannot watch them all

Interactive mode requires the tmux Skill

For interactive sessions, you MUST use the tmux Skill. The tmux Skill contains:

  • "Orchestrating Claude Code sub-agents" recipe - the primary pattern for this use case
  • Socket conventions and session management
  • Sending input, monitoring output, cleanup

Read and follow the tmux Skill for all interactive agent orchestration.

For parallel work on multiple issues, use the git worktrees Skill to create isolated branches.

Available agents

Read the relevant reference file for the agent you need right now:

Each reference includes: invocation, key flags, and completion detection.

Rules

  1. Respect tool choice - if user asks for Codex, use Codex. Do not offer to do the task yourself.
  2. Be patient - do not kill sessions because they seem slow. Agents take time.
  3. Monitor without interfering - check progress but let the agent work.
  4. Use appropriate flags - permission-bypassing flags for unattended/background work only.
  5. Isolate work - use worktrees or temp directories to avoid conflicts with your current work.
  6. Clean up - kill sessions and remove worktrees when done, unless the user asked you not to.