待验证50% 置信权衡取舍精确时间
Terminal-Bench 主要覆盖终端任务，无法反映模型在长文本理解、多模态、复杂推理等维度的表现
1
来源数
50%
置信度
长期有效
时效性
2026/8/11
首次发现
来源
相关事实
待验证Skill绑定的脚本通过命令行执行，脚本代码本身不进入上下文窗口，只有执行结果消耗Token57% 相似待验证In Claude Code Skills, bound scripts are executed via the command line and the script code itself does not enter the context window—only execution results consume tokens56% 相似待验证The `/loop` command is session-bound: once the terminal is closed, cancelled, or exited, all tasks immediately stop and are lost with no way to recover them.55% 相似待验证Claude Desktop cannot directly access and modify the project file system, cannot execute terminal commands, and cannot perform context-aware reasoning at the codebase level54% 相似待验证单次对话指令只在当前会话有效，关闭即失效，而全局自定义指令写一次即永久生效且跨项目通用51% 相似
引用此条事实
Stable URI
https://kongchang.com/claim/729508API
curl https://kongchang.com/api/v1/knowledge/claims/729508MCP
get_claim(id=729508)