SiNZeRo-简记·思行
Sunshine and Imagination

FIELD NOTES

Latest Notes

Recent notes from AI systems, models, infra, research, and whatever is worth keeping.

HARNESS / CLAUDE CODE

Claude Code keybindings:把 Ctrl+X 变成 Agent Leader Key

把 Claude Code 的 keybindings 组织成一套 Agent/TUI 控制层,重点覆盖 queueSubmit、diff、subagents、history 与 tmux。

Claude Code 的 ~/.claude/keybindings.json 不只是“改快捷键”。对 Agent-heavy 的工作流,更有价值的做法是:把 Ctrl+X 当成一层 Agent control plane,把 diff、排队消息、后台任务、subagent 控制等高频动作收进同一套 chord。官方文档:https://code.claude.com/docs/en/keybindings最值得记的默认快捷键 快捷键 Action 用途 Ctrl+X Enter chat:queueSubmit 排队发消息;Claude 正在工作时不会 interrupt 当前 turn Ctrl+X Ctrl+K chat:killAgents 停掉当前 session 的所有后台 subagents Ctrl+X Ctrl+B task:background 把当前 task 放后台,避开 tmux 的 Ctrl+B prefix Ctrl+G chat:externalEditor 用外部编辑器写长 prompt / spec Ctrl+S chat:stash 暂存当前...

Read →
HARNESS / MULTI-AGENT

Multi-agent Model Routing by Role

Model allocation and cost/task defaults for Controller, Worker, Reviewer, Explorer, and Advisor roles.

These three dimensions make the comparison clearer: Role Default recommendation Cost-efficient alternative (cost/task) Quality alternative Notes Controller Sol-medium Luna-max Astra-low Stable judgment matters by default; escalate hard tasks to Astra. Worker Luna-max Luna-max Sol-medium For bounded tasks, optimize for lower cost per completed task. Reviewer (weak) Luna-max Luna-max Terra-xhigh A weak reviewer...

Read →