Claude Code spider routes tool calls through Kai and Jev, waking Opus only for edge cases
An interceptor layer under Claude Code runs a local Kai decision step first, then hands cloud-bound branches to Jev, and only wakes Opus 5.5 when neither can answer.
OPUS 5.5 + JEV + DECISION 2.0 + CLAUDE CODE MODS on one spider, and Opus stops waking up for yes/no questions... most of what Opus does all day isn't hard work which file. which tool. retry or stop. is this risky you pay frontier prices for that so I put a spider under Claude Code. it catches every tool call, prompt and output before Opus sees it, and only wakes Opus for what it can't route the full tree: > guard mod → .env, secrets, rm -rf. plain code, no model, 0 tokens > Decision 2.0 Kai, 0.6B → runs on your laptop, answers 64 questions about a request in one pass > Jev → forks that need the cloud, $0.042 per 1M input, output free > Sonnet 5.5 swarm → routine edits and tests > Opus 5.5 → asleep. wakes only for the edge cases > second agent → reads everything Opus writes, sends you the short list bookmark this, you'll want it the next time Opus eats your limit paste the tree and this prompt into Claude Code ↓ 1. Create a Claude Code mod in ./spider with .claude-plugin/plugin.json, hooks/hooks.json and hooks/register.tsx 2. On http://tool.call, deny any read or edit of .env and secrets, and any bash with rm -rf. No model call 3. Route every routine question (which file, which tool, retry or stop, how risky) to Decision 2.0 Kai running locally with Transformers 4. If Kai isn't sure, send it to Jev 5. Hand routine edits and tests to Sonnet 5.5 subagents 6. Only wake Opus 5.5 when Kai and Jev both can't answer, and log why 7. Run claude plugin validate. Show me every change as a diff first ↳ https://claude.com/blog/claude-code-mods
This post quotes another post: x.com/0xsolty/status/2104665538655854981