Claude Code 三层模型分工类比 Jev 决策

marfin @marfinxx原推发布于 2026.10.16

在 X 上打开原推

THIS IS ABSOLUTE F*CKING GENIUS Anthropic just quietly shipped the native Advisor tool in Claude Code, and paired with the new Haiku 5.5 it completely breaks multi-model economics The idea: run Sonnet 5.5 as main builder, swarm Haiku 5.5 as high-speed read-only scouts, and keep Opus 5.5 on call as the expert advisor Opus-grade judgment on every architectural fork, zero Opus burn on routine file audits, grep searches, or lint passes Launch Claude Code with the native advisor flag: claude --model sonnet --advisor opus how the workload divides: -> Haiku 5.5 swarms parallel exploration: repo tree mapping, grep call site searches, and API spec parsing at 340 tok/s -> Sonnet 5.5 leads the main terminal session: writes implementations, refactors code, and runs test suites -> Opus 5.5 stays on call as advisor: receives the full conversation transcript and intervenes only on high-leverage checkpoints -> you stop burning Opus context on routine terminal iteration Opus only steps in at three specific moments: -> right before a plan locks -> the second an identical error hits twice -> right before declaring a complex task complete how the Advisor tool works under the hood: -> Server-Side Tool: runs server-side on Anthropic infrastructure, ingesting full multi-turn context without polluting local memory -> Capability Ranking: Claude Code validates model hierarchies upfront so Opus 5.5 advises Sonnet 5.5 and Haiku 5.5 -> Prompt Cache Lock: preserved prefix caching keeps 1M+ token context ingestion sub-second across repeated consultations -> Terminal Telemetry: inspect the full advisor reasoning transcript in real time simply by pressing Ctrl+O Haiku explores. Sonnet builds. Opus reviews JEV engineering is the exact same move one layer down: deterministic forks (file selection, tool dispatch, fast retry vs abort) clear in sub-milliseconds, so the frontier models only handle genuine forks the full architecture tree: > Sonnet 5.5 runs the primary session, writes implementations, and executes tests > search scout maps the repo structure on Haiku 5.5 at 340 tok/s > grep tracer locates all call sites on Haiku 5.5 at 340 tok/s > docs parser extracts schemas and specs on Haiku 5.5 at 340 tok/s > Opus 5.5 stands on call as advisor, evaluating plans and reviewing diffs > JEV resolves instant deterministic routing in sub-milliseconds paste the tree and this prompt into Claude Code: "Reconfigure my Claude Code setup around this multi-model architecture: 1. Run the primary session on Sonnet 5.5 with Opus 5.5 as advisor via --advisor opus 2. Spawn three read-only Haiku 5.5 subagents: search (repo mapping), grep (call sites), docs (API references) 3. Add one rule to CLAUDE.md: consult the advisor before locking a plan, after repeating the same error twice, and before final task handoff 4. Press Ctrl+O to inspect advisor guidance before applying recommendations 5. Route fast deterministic forks through JEV, show all changes as diffs first" -> http://code.claude.com/docs/en/advisor bookmark this, you will use it on every major codebase

这条是引用着另一条写的：x.com/mirku21/status/2098836008468992471