用 Jev Gate 给 Claude Code 做决策分流
把 Jev 接进 Claude Code hooks，用 70-500ms 的类型化决策过滤常规判断，只把不确定的交给 Opus，省下大量 token。
I don't understand why everyone isn't doing this yet Claude Code now has an advisor mode. Put a $0.04 decision model in front of it and you burn a fraction of the tokens Main idea: Jev Gate for Claude Code Most routine decisions in a session never need an LLM. So stop paying Opus to make them: Main session (Sonnet 5.5, medium effort): writes the code and delegates to subagents Advisor (Opus 5.5): sees the full transcript but only steps in before a plan, on a recurring error, and before "done". Never writes code Gate (Jev by TypeSafe): runs inside Claude Code hooks. Typed answers with a probability in 70-500ms - which agent, safe to run, diff ready. Confident answers are handled in code. Uncertain ones go to Opus Reviewer subagent (Opus 5.5, high effort): reads the diff, makes no edits Turn on the advisor in your terminal: /advisor opus Then paste this into Claude Code: Configure my Claude Code workspace for a Jev-gated setup: In .claude/agents, create a worker subagent on Sonnet (medium effort) and a reviewer subagent on Opus (high effort, read-only tools) In .claude/settings.json, set advisorModel to opus and add hooks: UserPromptSubmit - ask Jev "which agent?" (choice) PreToolUse on Bash - ask Jev "safe to run?" (noul), block below 0.5 Stop - ask Jev "diff ready?" (score), ask the advisor if confidence is low Write the hook script with @typesafe-ai/sdk and read the key from TYPESAFE_API_KEY In CLAUDE.md, add a rule: no PR without reviewer approval Show config diffs before applying edits > https://code.claude.com/docs/en/advisor
这条是引用着另一条写的：x.com/solst1ne/status/2101724066369654935