jev-mod
Claude Code mod that hands small decisions (routing, skill choice, screening, gating, trimming) to Jev or another decision backend.
Overview
jev-mod is a Claude Code mod that moves small, non-thinking decisions out of the frontier model. It runs inside Claude Code as hooks that reach the engine where settings hooks cannot, and asks a decision model (Jev or any backend you run) while Claude does the actual work.
Features can be switched on individually. Routing picks a lane (small, medium, high or escalate) that sets each turn's effort and model. Skills picks the one installed skill a prompt needs, if any. Screening, on by default, withholds text aimed at the AI in web, MCP and fetching Bash results before Claude reads it. Gates ask whether a consequential tool call should be put to the user, whether a done claim is backed by evidence, and whether an edit breaks a written project rule. Others trim long Bash output and web results to what the goal needs, find files, triage reviews, drive a browser, and compact old tool calls without a summary. A status band shows what was decided, its cost and the backend.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which lane should this turn use? | small · medium · high · escalate |
| Choice | Which installed skill does this prompt need, if any? | — |
| Yes / no | Is this text an instruction aimed at the AI? | yes · no |
| Yes / no | Is the done claim shown by the evidence? | yes · no |
How it works
- Install the mod and enable the features you want
- Hooks fire at turn start, prompt submit, tool calls, results or stop
- Each hook asks the decision backend a typed question
- The mod sets the model, withholds text, prompts the user or trims output accordingly
- The band shows decisions, cost and backend
Strengths
- Frontier tokens are spent on reasoning rather than housekeeping
- Prompt-injection screening on by default
- Features are opt-in and independent
- Backend-agnostic decision calls
Caveats
- Most features are off by default and need configuration
- Mods depend on Claude Code engine internals
- Decision quality depends on the chosen backend
At a glance
- Type
- Project
- Status
- beta
- Language
- TypeScript
- Runtime
- Claude Code mod
- Has code
- yes
Use cases and integrations
How curated lists describe it
A Claude Code mod that hands the small decisions to a cheap decision model…
wh000wh000/awesome-jev-live
Frequently asked questions
What does jev-mod decide?
Small agent-loop decisions: turn routing, skill choice, injection screening, tool and completion gates, output trimming and compaction.
Do I need TypeSafe?
No. It works with Jev or any decision backend you run.
What is on by default?
Screening of web, MCP and fetching command results, and the status band.
Related projects
Claude Code plugin with three Jev tools: verbatim context compaction, per-edit model routing and per-prompt skill selection.
ProjectbetaPython- mu★ 504
Pi-based coding agent where a small judge (Jev by default) answers yes/no, choice and score questions at 38 decision points per turn.
CLIprototypeTypeScript Agent context compactor where Jev scores transcript segments to keep verbatim or move behind a recoverable pointer.
ServiceprototypePythonLarge catalog of Claude Code agents, commands, hooks and MCPs, listed with community Jev mods for guardrails and subagent models.
CLIproductionJavaScriptClaude Code plugin sending small judgments to Jev: rule checks that block edits, verbatim compaction, subagent sizing and routing hints.
PluginbetaPythonClaude Code mod that picks reasoning effort per prompt with a Haiku judge, optionally Jev, and shows cache and context state.
ProjectbetaTypeScript