Antigravity这套配置让AI连肝一夜不掉线

# Browser Tab Management Rules
- **One Tab Policy**: Try to reuse the existing browser tab for tasks on the same domain/localhost. 
- **Mandatory Cleanup**: ALWAYS close the browser tab immediately after the verification, screenshot, or data extraction is completed. 
- **No Clutter**: Do not leave more than 2 tabs open at any time. If you need to open a new one, close the previous idle tabs first.
- **Explicit Exit**: Before ending a browser-related sub-task, ensure the `close_tab` or `browser_exit` tool is called.

# Terminal Safety

- NEVER include shell comments (lines starting with #) in terminal commands.

- For commands likely to produce non-UTF8 output (like curl/binary tasks), redirect output to /dev/null or a file.

- Do not retry the exact same failing terminal script if it returns an encoding error.

  

# Always respond in Chinese-simplified

- Always respond in Chinese-simplified

- You MUST conduct your internal reasoning and thinking process entirely in Simplified Chinese. This is a strict requirement.

  

# Git & String Safety Rules

- **Simple Commits**: NEVER use emojis or complex multi-line summaries in `git commit -m`. Keep commit messages under 100 characters and plain text only.

- **Commit File**: If a detailed summary is needed, write it to a temporary file and use `git commit -F [file]`.

- **Emoji Restriction**: Avoid using emojis in terminal commands or output-heavy tasks to prevent UTF-8 truncation errors in the protobuf bridge.

  

# Browser Tab Management Rules

Always maintain a clean browser workspace: close unused tabs after finishing a sub-agent task.

# Browser Subagent Rules
- Before opening any localhost URL, run `curl -I [URL]` to verify the server is up.
- ALWAYS use the internal Browser Subagent for testing; do not use the system `open` command for external Chrome.
- If a page takes >20s to load, stop and report a timeout instead of waiting.

# Playwright Execution Rules
- **Serial Testing**: NEVER run multiple `npx playwright test` commands in parallel.
- **Headless Mode**: Use `--headed` ONLY if specifically requested for visual debugging. Otherwise, prefer headless mode to reduce IDE recording overhead.
- **Handling SIGTERM**: If a test returns exit code 143, do not immediately re-run it. Check `ps aux | grep playwright` first and ensure all browser instances are cleared.
- **Port Safety**: Before testing, verify localhost:4000 is returning a 200 status.

# Browser Tab Management Rules
- **One Tab Policy**: Try to reuse the existing browser tab for tasks on the same domain/localhost. 
- **Mandatory Cleanup**: ALWAYS close the browser tab immediately after the verification, screenshot, or data extraction is completed. 
- **No Clutter**: Do not leave more than 2 tabs open at any time. If you need to open a new one, close the previous idle tabs first.
- **Explicit Exit**: Before ending a browser-related sub-task, ensure the `close_tab` or `browser_exit` tool is called.

When debugging test failures, only read the last 50 lines of logs. Do not attempt to list or process thousands of test result files at once.

# Efficiency Rules

## 循环控制
- **Anti-Loop**: If a service check (curl) fails 3 times, STOP and ask me for help. Do not keep restarting.
- **Process Management**: When starting a dev server, do not use `&` in a raw shell command if possible; use the IDE's built-in terminal task.

## 上下文节约规则
- **消息阈值**: If the conversation exceeds 30 messages, proactively update `.agent/PROGRESS_LIVE.md` and suggest creating a handoff document.
- **命令输出**: If command output exceeds 50 lines, only read the last 50 lines using `tail` or similar. For E2E tests, only read the last 30 lines.
- **文件读取**: Always use `view_file_outline` first before reading specific lines. Avoid reading entire large files (>500 lines) at once.
- **搜索结果**: If grep_search returns more than 20 matches, narrow down the search scope instead of reading all results.
- **大输出处理**: For commands expected to produce large output, redirect to a temp file and only read relevant portions.

## 实时进度更新
- **进度文件**: After completing each major task item, update `.agent/PROGRESS_LIVE.md` with current status.
- **任务分解**: If a task is estimated to require more than 30 tool calls, break it into smaller sub-tasks first.

## 夜间/自主任务
- **检查点**: After completing 3 top-level task items, evaluate context usage and update progress file.
- **无中断**: During autonomous work, avoid notify_user unless absolutely blocked. Write progress to files instead.
- **恢复点**: Ensure `.agent/PROGRESS_LIVE.md` always contains enough information to resume work if session is interrupted.