Claude Code 工作流里用 Jev 做确定性分支

mirku @mirku21原推发布于 2026.10.10

在 X 上打开原推

holy sh*t. Official Anthropic tip for Claude Code: Once Opus 5.5 is your main model, stop leaving Haiku 5.5 sitting idle and stop burning Opus tokens on sweeps Sonnet 5.5 and Haiku 5.5 can tear through turn mass-agent sweeps into deterministic code with /workflows claude --effort ultracode Opus 5.5 authors the workflow script and synthesizes the final PR Sonnet 5.5 executes precision refactors in isolated worktrees at medium effort Haiku 5.5 runs at 340 tok/s as the dedicated pipeline filter, cutting cognitive load at three critical bottlenecks: → before dispatch: sweeps the full codebase in parallel to extract AST candidates without context bloat → on schema validation: drops malformed tool calls and false positives before Sonnet touches them → before synthesis: verifies that every candidate strictly conforms to the target JSON schema Haiku 5.5 filters. Sonnet 5.5 refactors. Opus 5.5 ships Jev engineering is the same move one layer down: the forks that need no thinker (which file, which tool, retry or stop) resolve in V8 in 16ms, and frontier models only wake up when execution paths actually split Author on high. Refactor on medium. Keep Haiku at 340 tok/s. - the full workflow architecture > Opus 5.5 on high drafts the workflow script in .claude/workflows/ > pipeline() fans out Haiku 5.5 at 340 tok/s across target files > parallel() schema fuzzers drop false positives before patch generation > Sonnet 5.5 refactors verified files in isolated git worktrees > intermediate state lives in V8 variables, keeping context drag at 0 tokens > 1h prompt cache prefix hold eliminates redundant system prompt processing paste the setup and this prompt into Claude Code ↓ "Configure my Claude Code workspace for dynamic workflows: 1. Audit ~/.claude/workflows and .claude/workflows for existing scripts: > Draft workflow definitions for codebase audits, parallel migrations, and schema validation > Pin model: haiku, effort: low for AST sweepers, and model: sonnet, effort: medium for refactor workers > Configure workers to operate in isolated git worktrees to eliminate edit conflicts 2. In ~/.claude/settings.json, enable dynamic workflows: > Set ultracode = true > Lock main session effortLevel to high > Set subagentPromptCacheTtl = '1h' to share prompt prefixes across parallel fan-outs 3. Set workflow runtime controls: > Set autoContinueAtUsageLimit = true so long batch runs automatically resume after limit resets > Set CLAUDE_CODE_WORKFLOW_PREFIX_STAGGER_MS = 5000 for prompt prefix caching > Set large workflow warning threshold to 25 agents 4. Add a rule to ~/.claude/CLAUDE.md: all multi-file refactors must pass an adversarial Haiku schema check before patches merge to main Show me every change as a diff first. No edits until I say go." ↳ http://code.claude.com/docs/en/workflows

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