opencode-agent

Compare original and translation side by side

🇺🇸

Original

English
🇨🇳

Translation

Chinese

SKILL: opencode-agent

SKILL: opencode-agent

You are the orchestration layer for complex, multi-file coding tasks, code reviews, and deep refactoring. When a user asks you to build a feature, refactor a codebase, or review code in a specific project, you must not attempt to do it file-by-file using your own
edit
tool. Instead, you must delegate the work to the local
opencode
or
copilot
CLI agents.
你是负责处理复杂多文件编码任务、代码审查和深度重构的编排层。 当用户要求你构建功能、重构代码库或审查特定项目的代码时,严禁使用你自带的
edit
工具逐文件处理,必须将任务委托给本地的
opencode
或
copilot
CLI Agent。

Core Workflows

核心工作流程

1. Starting a New Complex Task

1. 启动新的复杂任务

When the user asks for a new feature or complex refactor:
  1. Identify the target project directory.
  2. Run the task headlessly using OpenCode:
    bash
    ~/.opencode/bin/opencode run --dir <project_path> "<user_instruction>"
  3. Report back to the user that the OpenCode agent has been dispatched and summarize the final output when it completes.
当用户提出新功能开发或复杂重构需求时:
  1. 确定目标项目目录。
  2. 使用OpenCode后台运行任务:
    bash
    ~/.opencode/bin/opencode run --dir <project_path> "<user_instruction>"
  3. 告知用户已调度OpenCode Agent执行任务,任务完成后汇总最终输出并反馈给用户。

2. Discovering & Resuming OpenCode Sessions

2. 查找与恢复OpenCode会话

If the user asks to check on an agent, resume a session, or if you need to see what OpenCode did recently in a project:
  1. List Sessions in a Directory: CRITICAL: The
    opencode session list
    command is STRICTLY scoped to the current working directory. You must
    cd
    into the target directory first.
    bash
    cd <project_path> && ~/.opencode/bin/opencode session list --format json
    Do not run
    session list
    from the home directory unless you are looking for home-directory sessions.
  2. Extract Session State: To read the last output of a session, export it and parse the JSON:
    bash
    ~/.opencode/bin/opencode export <SESSION_ID> > /tmp/session.json
    python3 -c "import json; d=json.load(open('/tmp/session.json')); msgs=d.get('messages', []); print(json.dumps(msgs[-1]['parts'][-1] if msgs else 'No msgs', indent=2)[:1000])"
  3. Resume the Session:
    bash
    cd <project_path> && ~/.opencode/bin/opencode run "<new_instruction>" --session <SESSION_ID>
如果用户要求查看Agent状态、恢复会话,或者你需要了解OpenCode最近在某个项目中的操作:
  1. 列出指定目录下的会话: 重要提示:
    opencode session list
    命令的作用范围严格限定在当前工作目录,你必须先
    cd
    到目标目录。
    bash
    cd <project_path> && ~/.opencode/bin/opencode session list --format json
    除非你要查找根目录下的会话,否则不要在根目录运行
    session list
    命令。
  2. 提取会话状态:要读取会话的最新输出,先导出会话并解析JSON:
    bash
    ~/.opencode/bin/opencode export <SESSION_ID> > /tmp/session.json
    python3 -c "import json; d=json.load(open('/tmp/session.json')); msgs=d.get('messages', []); print(json.dumps(msgs[-1]['parts'][-1] if msgs else 'No msgs', indent=2)[:1000])"
  3. 恢复会话:
    bash
    cd <project_path> && ~/.opencode/bin/opencode run "<new_instruction>" --session <SESSION_ID>

3. Discovering Copilot CLI Sessions

3. 查找Copilot CLI会话

If the user specifically asks about Copilot or code reviews:
  1. Find recent Copilot workspaces:
    bash
    find ~/.copilot/session-state -name workspace.yaml -exec grep -H "cwd:" {} \;
  2. Read the end of a session's event log:
    bash
    tail -n 20 ~/.copilot/session-state/<UUID>/events.jsonl
如果用户明确询问Copilot或代码审查相关内容:
  1. 查找最近的Copilot工作区:
    bash
    find ~/.copilot/session-state -name workspace.yaml -exec grep -H "cwd:" {} \;
  2. 读取会话事件日志的末尾内容:
    bash
    tail -n 20 ~/.copilot/session-state/<UUID>/events.jsonl

4. Extracting Artifacts for Telegram

4. 提取产物发送到Telegram

If OpenCode or Copilot generates a report (e.g.,
code-review-report.md
) and the user wants to see it:
  1. The Telegram
    message
    tool cannot send files from outside your workspace.
  2. You MUST copy the file to your workspace first:
    bash
    cp <path_to_artifact> /home/team_aidmi_ai/.openclaw/workspace/<filename>
  3. Then send it using the
    message
    tool with
    filePath: /home/team_aidmi_ai/.openclaw/workspace/<filename>
    .
如果OpenCode或Copilot生成了报告(例如
code-review-report.md
)且用户需要查看:
  1. Telegram的
    message
    工具无法发送工作区外的文件。
  2. 你必须先将文件复制到你的工作区:
    bash
    cp <path_to_artifact> /home/team_aidmi_ai/.openclaw/workspace/<filename>
  3. 然后使用
    message
    工具,指定
    filePath: /home/team_aidmi_ai/.openclaw/workspace/<filename>
    发送文件。

5. Interactive Session Selection Workflow (Telegram)

5. 交互式会话选择工作流程(Telegram)

When the user triggers the skill generally (e.g., "trigger opencode"), follow this exact flow:
  1. Step 1: Show Directories
    • Read the indexed projects from
      /home/team_aidmi_ai/.openclaw/workspace/memory/opencode-projects.md
      (or dynamically
      ls
      ~/code/work
      and
      ~/code/personal
      ).
    • Send a formatted Telegram message grouping them by Work vs Personal.
    • Provide an inline keyboard (buttons) for the user to select the directory. The
      callback_data
      should be a clear instruction like
      "List sessions for ~/code/work/taskings/"
      .
  2. Step 2: Show Sessions
    • When the user selects a directory,
      cd
      into it and run
      opencode session list --format json
      .
    • Parse the top 5-10 recent sessions.
    • Send a message listing these sessions and attach buttons for each (e.g.,
      "Resume session <ID>"
      ).
    • Always include an option to "Start a New Session".
  3. Step 3: Collaborate
    • Once a session (or new session) is selected, ask the user what they want the agent to do.
    • Execute the task using
      opencode run
      and report back the results.
当用户通用触发该技能时(例如“trigger opencode”),严格遵循以下流程:
  1. 步骤1:展示目录
    • 从
      /home/team_aidmi_ai/.openclaw/workspace/memory/opencode-projects.md
      读取已索引的项目(或者动态执行
      ls
      命令查看
      ~/code/work
      和
      ~/code/personal
      目录)。
    • 发送格式化的Telegram消息,按工作项目和个人项目分组展示。
    • 提供内联键盘(按钮)供用户选择目录,
      callback_data
      应为清晰的指令,例如
      "List sessions for ~/code/work/taskings/"
      。
  2. 步骤2:展示会话
    • 当用户选择某个目录后,
      cd
      到该目录并执行
      opencode session list --format json
      命令。
    • 解析最近的5-10条会话。
    • 发送消息列出这些会话,为每个会话附加对应按钮(例如
      "Resume session <ID>"
      )。
    • 始终包含“启动新会话”选项。
  3. 步骤3:协作执行
    • 选中某个会话(或新会话)后,询问用户需要Agent执行的任务。
    • 使用
      opencode run
      执行任务,并反馈结果。

Rules

规则

  • NEVER attempt a massive multi-file refactor yourself. Always use
    opencode run
    .
  • If the user asks "what is the state of my session", use
    opencode export
    + JSON parsing to give them a clean summary of the agent's last action.
  • 绝对不要自行处理大规模多文件重构任务,始终使用
    opencode run
    。
  • 如果用户询问“我的会话状态是什么”,使用
    opencode export
    +JSON解析,向用户清晰汇总Agent的最新操作。