← ClaudeAtlas

rjv-jev-decisionslisted

Hand closed-set decisions in Claude Code or Codex to Jev, TypeSafe's System One model, instead of spending flagship tokens. Main use: pick the model tier before spawning a subagent or Codex worker (haiku/sonnet/opus, gpt-5.6-luna/terra, gpt-6-astra with effort, or local Ollama), with confidence and stakes gates. Ships PreToolUse hooks that fill in a missing model. Also checks a cheap worker's report against its acceptance items, and scores Build mode for rjv-work-plan. Triggers: use jev, typesafe, which model for this subagent, should I escalate, did the worker finish, build mode check, jev hook, model pick log.
rjvim/ai-skills · ★ 0 · AI & Automation · score 61
Install: claude install-skill rjvim/ai-skills
# Jev decisions: cheap, fast judgments for coding agents Jev is TypeSafe's System One model. It reads a state and answers typed questions with probabilities and a confidence. The question types are Choice (one of a set), Noul (probability of yes) and Score (a level). Jev generates no text. Why it fits a coding session: many decisions an agent makes are closed-set ("which model?", "is this done?", "is this risky?"). A flagship spends a thinking turn on each. Jev answers in about a second for about $0.00002 a call (input tokens only, $0.042 per million; output is free). ```text agent is about to decide something from a fixed set │ ├── code can decide it exactly (count, compare, lookup) ──> code │ ├── needs reasoning, several hops, or generating text ────> the agent itself │ └── one semantic judgment over a short state ─────────────> Jev │ confidence high ──> act on the answer │ confidence low ──> go one step safer <──┘ ``` ## Setup - `TYPESAFE_API_KEY` in the environment (keys: console.typesafe.ai/keys). - `curl` and `jq`. No SDK needed. - Scripts live in `scripts/`; question sets in `decisions/`. If the key is missing or the call fails, every script exits 3. **Exit 3 means "no decision", never "stop".** Fall back to the manual ladder below and carry on. ## 1. Pick the model for a delegated task (