jev-builder-loop
Grok skill that uses Jev as a judgment sensor in a builder-agent loop, mixing its probabilities with priors before choosing the next act.
Overview
jev-builder-loop is a skill for Grok's coding agent in which you write questions and environment state, Jev returns typed probabilities, and a policy mixes them with priors into posteriors before deciding the next act. Jev is used as a judgment sensor, not a coder; the goal is to cut wasted retries and operator babysitting. A script runs at loop boundaries reading state from stdin, and SKILL.md documents when to call, a question bank and the mix policy, including a hard-fact veto.
Its benchmark scenario is shipping a signed Android App Bundle after a UI change merged to main, using anonymised cases from a real session such as a green PR left waiting for review, treating a merge as a store release, and retrying Gradle with an empty JAVA_HOME. On the first snapshot, the agent without the skill scored 1/9 and the skill 6/9, while trusting Jev's argmax scored 9/9; the README explains that one argmax hit had confidence 0.29 and that Jev's own Noul contradicted its choice on a timeout, so the mix deliberately takes the conservative side. Committed answers let the benchmark run without a key. It is MIT licensed.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | What is the next act (e.g. retry_same, gather_facts, ship_or_merge)? | retry_same · gather_facts · ship_or_merge · ... |
| Yes / no | Is this a repeated failure? | yes · no |
How it works
- At a loop boundary, write state and questions
- Jev returns probabilities
- Mix with priors and apply hard-fact vetoes
- Choose the next act
Strengths
- Cuts wasted retries in agent loops
- Conservative mixing guards against overconfident argmax
- Benchmark runnable offline with committed answers
Caveats
- Small benchmark (9 cases)
- Mix can be over-cautious, as the README notes
Getting started
Clone into ~/.grok/skills/jev-builder-loop, set TYPESAFE_API_KEY, then `python3 scripts/loop.py < state.json`.At a glance
- Type
- Agent skill
- Status
- prototype
- Language
- Python
- Runtime
- Grok agent skill
- Has code
- yes
- License
- MIT
Use cases and integrations
How curated lists describe it
Grok skill: Jev as a judgment sensor in a builder-agent loop (priors × probabilities → next act).
hellogumbo/awesome-jev
Frequently asked questions
What does Jev decide?
Probabilities for the next act and related checks, which are mixed with priors.
Why not just take Jev's top answer?
The README shows low-confidence and self-contradicting cases where a conservative mix is safer.
Which agent is it for?
Grok's coding agent.
Related projects
- SystemOneHarness★ 193
Agent harness that compiles an environment's finite actions into typed questions for Jev and gates each step by confidence.
CLIbetaPython Durable Android sub-agent: a SQLite-backed worker observes the UI and Jev picks among valid actions, with verification and escalation.
MCP serverbetaPythonOmO/senpi agent plugin where Jev or Span-01 rates which skills and next tools fit the task and spots loops or completion.
Agent skillbetaTypeScriptClaude Code function hook that sends the agent's this-or-that questions to Jev and optionally ranks installed skills per prompt.
ProjectbetaTypeScriptPi coding agent extensions, including pi-jev-router, which asks Jev whether a prompt should continue, fork or start a new session.
PluginbetaTypeScriptClaude Code skill that splits work into PBIs for Codex, Antigravity or Claude sub-agents, using Jev for model choice and triage.
ProjectprototypePython