Claude Code 用 Jev 思路给每个子 agent 配顾问
把 Fable 5.1 设为 advisor 挂在每个 Sonnet 5.5 子 agent 后面，只在重写多文件、重复失败、报告完成三个关键分叉点介入判断。
holy f*ck. it's in Anthropic's own docs: stop saving Fable 5.1 for your main session when every Sonnet 5.5 subagent can call it too put Fable behind every agent with --advisor claude --advisor fable Opus 5.5 leads on high, launched with the flag so a broken advisor fails loudly Sonnet 5.5 subagents explore, edit and test at medium, and every one inherits the advisor Fable 5.1 stays on call for all of them, only stepping in at three critical moments: → before a subagent rewrites a dozen files: is this the plan Opus approved? → when a subagent fails the same test twice: is it missing context the lead had? → before a subagent reports done: did it check what it was sent to check? Sonnet 5.5 works. Fable 5.1 watches every agent. Opus 5.5 signs off Jev engineering saves the advisor for the forks that matter: the trivial ones (which file, which tool, retry or stop) resolve in code, so every Fable call lands on a real decision Turn it on. Prove it's on. Let every agent use it. - the full advisor tree > Opus 5.5 on high leads the session, launched with --advisor fable > explorer, worker and tester on Sonnet 5.5 at medium, each inheriting Fable > DISABLE_TELEMETRY, or anything that blocks feature-flag fetching, keeps the advisor off > if the API refuses the pairing, Claude Code resends without the advisor and shows no error > Fable as advisor needs a one-time usage-credits consent through /model fable paste the tree and this check into Claude Code ↓ "Make sure my advisor is actually running, everywhere: 1. Search my shell profile, env and every settings file for anything that switches the advisor off. 2. Check that my main model accepts fable as its advisor, and if Fable still needs the one-time usage-credits consent, tell me to run /model fable 3. Check ~/.claude/agents and .claude/agents for explorer, worker and tester > Create only the roles that are missing > Give each model: sonnet, effort: medium so they pass the Fable pairing check > List any agent pinned to a model that can't take a Fable advisor 4. Add one rule to ~/.claude/CLAUDE.md: the main session and every subagent consult the advisor before multi-file edits, when an error repeats, and before reporting done Then tell me to restart with claude --advisor fable and look for the "Advisor Tool (experimental) is on" line. Preview every edit as a diff and apply nothing until I confirm." save and bookmark this no matter what
这条是引用着另一条写的：x.com/eng_khairallah1/status/2104868464770863565