agent-computer-interfacelisted

**CONSUMES**: User browser/computer/scraping task description + current environment context **PRODUCES**: Tool selection decision + applied judgment rules + capability detection results + configuration guidance 平台绑定交互决策（cross-harness binding）：本文件及其 references 中所有 AskUserQuestion 调用是「交互决策契约」而非具体工具——当前 harness 有该工具 （Claude Code）→ 直接调用；无该工具（Codex 等）→ 以编号纯文本列出全部选项 （1. … / 2. … / 3. …）并**停止等待用户输入**，用户以编号或自由文本作答； 禁止代答、禁止把选项折叠成默认值继续执行。SAFETY 门控的调用点（人工审批 / 归档确认 / 权限升级确认类）无论何种 harness 都必须获得真人作答后才能继续。 非交互执行模式（如 codex exec）→ 视为无人可答，按 blocked 停止并上报， 不得自选默认值；已按 YOLO/预授权模式运行且该决策点有书面预授权记录 → 按其协议处理，不适用本条 blocked 分支。 # Agent Computer Interface Capability Pack **Version**: 0.1.0 **Compatibility**: Claude Code (Phase 1); Codex / Cursor / Gemini in Phase 3 **License**: Apache 2.0 --- ## What This Pack Does AI agents encounter a browser or desktop task and immediately start installing Playwright, even when Claude in Chrome is already connected. They pick Computer Use for a simple page scrape. They use screenshot-based vision for a table that Firecrawl would extract as clean JSON in one call. They spend 10+ minutes trying tools that aren't installed while the right tool is already available in their MCP server list. The root problem is not that agents can't use individual tools — it's that they lack systematic awareness of what's available and which layer fits the task. This pack embeds the judgment rules that a browser automation engineer applies automatically: detect first, match the task