Claude Code 里用 Jev 做小决策、Fable 当顾问

Mikadzyki🌙 @Mikadzyki_NFT原推发布于 2026.10.05

在 X 上打开原推

OPUS 5.5 + FABLE 5.1 + JEV = A DEVELOPMENT TEAM WITH AN AI ADVISOR IN CLAUDE CODE Long tasks alternate between writing code and making decisions that shape everything that follows. An advisor provides a separate review at those moments while the main model keeps doing the work Enable it with /advisor fable In this setup, Opus 5.5 runs the session on high, while three subagents split the routine work: - Explorer reads the code - Worker makes changes and runs tests - Researcher finds the relevant documentation Fable 5.1 receives the full conversation, including tool calls and their results. Its advice therefore takes into account the work already done and the approaches that failed Instructions can set three points for consultation: - before committing to a major plan - after the same error happens twice - before declaring a long task complete These are instructions for the model, not hardcoded triggers. It decides when to consult the advisor The proposed setup also includes Jev from TypeSafe AI. It handles quick, small decisions such as classification or choosing the next branch of the workflow The advisor only uses tokens when consulted But it rereads the entire session each time, so frequent reviews of a long conversation can get expensive. The point of this split is to get a second opinion where a wrong decision would lead to substantial rework

这条是引用着另一条写的：x.com/mikadzyki_nft/status/2101141853810618747