Claude Code 分层调度：Opus 顾问 + Jev 16ms 路由

Bober_smart @Bober_smart原推发布于 2026.10.10

在 X 上打开原推

Why is everyone silent about this? Official Anthropic tip for Claude Code: stop burning Opus 5.5 capacity on tasks Haiku 5.5 can swarm, while Sonnet 5.5 handles construction  shift the architect to standby mode using /advisor  claude --advisor opus --subagents haiku  Sonnet 5.5 drives the main session at maximum effort level, drafting modifications and running test suites Haiku 5.5 subagents scan the repository in parallel at 340 tokens per second, performing file discovery and reviewing spec documentation Opus 5.5 stays ready in the background as an expert advisor, stepping in only at three critical moments:  > rightarrow before locking down a strategy: does the blueprint miss authentication conditions or schema rules? > rightarrow when a test fails twice: are we fixing the root cause or going down an endless rabbit hole? > rightarrow before declaring completion: did final diffs introduce hidden bugs, and are baseline checks met?  Opus 5.5 evaluates. Sonnet 5.5 constructs. Haiku 5.5 executes in paralle Jev engineering applies the exact same pattern one layer below: automated decisions requiring no reasoning (which file to open, which tool to trigger, retry or cancel) run in 16ms, so top-tier models wake up only when execution paths truly diverge Plan at maximum. Delegate at medium. Keep Opus ready > On-call Opus 5.5 reviews the full session history and spots deep architectural flaws > Lead Sonnet 5.5 applies edits, writes core logic, and executes test environments > Haiku 5.5 subagents handle AST parsing, grep queries, and API doc analysis in parallel > JEV micro-fork layer handles 1,500+ formal routing paths in under 16ms > Expert remains fully silent during routine bash execution to conserve context Configure my Claude Code workspace for hierarchical advisor orchestration:  1. Check ~/.claude/settings.json and project settings for model role alignment (main, subagents, and advisor): Pin main session model: sonnet with effortLevel: high for primary execution Pin subagents model: haiku with effortLevel: medium for parallel file discovery and doc gathering Pin advisor model: opus on call for strategic analysis 2. Enable the advisor tool in ~/.claude/settings.json: Set advisorModel to claude-opus-5-5 Enable subagent parallel dispatcher pool (3x workers) 3. Set up automatic advisor consultation checkpoints in ~/.claude/CLAUDE.md: Request /advisor opus before finalizing multi-file architecture plans Automatically invoke /advisor opus when the same compiler or test error occurs twice Require an advisor diff contract audit before marking tasks complete or staging git commits 4. Limit subagent operational scopes: Haiku subagents return structured AST summaries and documentation snippets only without altering the main session context Show all configuration diffs before applying. Do not commit edits until confirmed Http://code.claude.com/docs/en/advisor

这条是引用着另一条写的：x.com/bober_smart/status/2078784709253841039