Rebuilding the agent loop so Opus writes code and Jev makes decisions
A typesafe Claude Code plugin replaces every yes/no, pick-one and ranking question in an agent with a single Jev probability call, with thresholds in plain code and low-confidence cases escalated to Opus.
you’re paying Opus $20 per million tokens to answer yes or no questions most of an agent loop isn’t coding. it’s deciding. → Opus 5.5 plans and writes the code → Jev answers every yes/no, pick one, rank it → plain code acts on the probability → Opus only wakes up when Jev isn’t sure Opus writes. Jev decides. paste this into claude code 👇 claude plugin marketplace add typesafe-ai/skills claude plugin install typesafe@typesafe-ai Rebuild my agent loop as: Opus writes, Jev decides, code acts. 1. Find every place my agent asks a model a yes/no, pick one, or rank question (routing, "is this done?", "is this safe?", retries). 2. Replace each with one Jev call (model: jev-latest) that returns probabilities. Output is free, so ask all the questions in a single call. 3. Put the thresholds in plain code. Low confidence escalates to Opus 5.5, never down. 4. Log every Jev call and every escalation so I can see the savings. Show me the diff before changing anything.
This post quotes another post: x.com/1373665408193036293/status/210869352490011…