claude-code-mods

根代理派活规则：分包、模型与档位、执行包写法、验收与终态报告（配合 delegate 和 dispatch-gate）

1 skill

用户离开 5 分钟未回复待确认的问题时，提示模型自行决定；模型可调用工具暂停自动推进。

Observes prompt-cache usage, keeps the session's and its finished subagents' caches warm, and runs 5m/1h TTL A/B experiments.

让会话里的模型自己重启 Claude 桌面版，并在重启后唤醒同一个会话继续干活。

Proactive context management: a task ledger, a searchable history and zero-summary window switches.

Hand a task to a cheaper model (any OpenAI chat-protocol provider) through a local Anthropic-to-OpenAI proxy; /delegate manages providers, keys and models.

A soft gate on the main session: reminds it, then refuses once per stretch, to hand work to a subagent; and refuses one Claude subagent per hour past three.

Adaptive thinking effort for the main conversation only: rest at low, raise it by steps when judgement is needed.

Programmatic tool calling: run_code runs a JS program that calls read/glob/grep/edit/bash in a sandbox and returns only the final value; large results spill to refs.

Re-reading a file already in the context returns a one-line notice if unchanged, or only a diff if changed, saving context tokens.