muse-jev-playbook
Playbook for using Jev as a cheap gate before expensive agent steps: confidence policy, question recipes, a skill and a router.
This repository is mostly documentation or write-up rather than runnable code.
Overview
muse-jev-playbook shows how to use TypeSafe's Jev as a fast, cheap decision layer inside an agent workflow, triaging, classifying, scoring and gating work before expensive steps such as browsing, deep research, retries or subagents. It is adapted from grok-bot-jev and generalised for Muse/Hatch agents, custom harnesses or plain scripts.
It includes concept docs, an architecture note on where the layer sits, a confidence policy, prompting guidance, a shadow-to-active measurement method, a drop-in skill, six recipe packs (triage, rank options, act-or-wait, retry gate, research cap, approval gate), a reference router and a flight-date case study. The core loop builds a compact state, makes one Jev call with parallel choice, score and noul questions, then acts at confidence of at least 0.80, surfaces between 0.50 and 0.79, and escalates below 0.50. The README says a call costs a fraction of a cent and typically returns in about a second.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which path should the agent take? | reuse_cache · stop_retry · run_deterministic · chat_only · research_capped · allow_subagent · ask_human · proceed_full |
| Yes / no | Should we retry this failed step?(inferred) | yes · no |
| Score | How much research effort does this request need?(inferred) | — |
How it works
- Check the kill switch (bypass jev, no jev, enabled false).
- Build a compact state with goal, kind, cache hints, errors and constraints.
- Make one Jev call with parallel questions.
- Act at >= 0.80, surface at 0.50-0.79, escalate below 0.50.
Strengths
- Explicit, reusable confidence policy.
- Ready-made recipe packs and a drop-in skill.
- Guidance on measuring effect honestly with shadow rollout.
Caveats
- Mostly documentation and recipes; depends on the hosted TypeSafe API.
- Thresholds are suggestions that need tuning per workload.
Getting started
Install the TypeSafe SDK, set TYPESAFE_API_KEY, then follow QUICKSTART.md to run the router in src/ or the recipes.At a glance
- Type
- Project
- Status
- prototype
- Runtime
- agent skill, scripts
- Has code
- yes
- License
- MIT
Use cases and integrations
How curated lists describe it
Jev decision layer for Muse: a fast, cheap TypeSafe AI gate before expensive agent work — confidence policy, recipes, reference router, honest measurement.
logicrw/awesome-jev-projectsJev decision layer for Muse: a fast, cheap TypeSafe AI gate before expensive agent work — confidence policy, recipes, reference router, honest measurement.
hellogumbo/awesome-jev
Frequently asked questions
What does the playbook gate?
Expensive agent steps such as browsing, deep research, retries and subagents.
What thresholds does it recommend?
Act at 0.80 or above, surface between 0.50 and 0.79, and escalate below 0.50.
Do agents need to handle keys?
On Muse/Hatch the credential is pre-connected; developers set TYPESAFE_API_KEY.
Related projects
Decision engine that combines Jev probability distributions with loss matrices to decide when a clarifying question is worth asking.
ProjectprototypeTypeScriptRust MCP server for Codex that hands classification, relevance, candidate selection and evaluation to Jev with accept/verify policies.
MCP serverbetaRustAgent skill and CLI that chains Jev question design: find decision calls, check fit, draft, lint, calibrate thresholds and plan a cascade.
Agent skillprototypePythonPython sandbox turning Jev issue-category and reproduction judgments into testable routing proposals, with offline replay.
ProjectprototypePythonInteractive web workbench for exploring and benchmarking Jev's Choice, Noul and Score via OpenRouter, with threshold transfer curves.
ProjectprototypeTypeScriptIndependent 52-page Chinese report on Jev that checks vendor claims and reproduces 50 Chinese tests, with traceable data.
ProjectresearchPython