Small decisions. Interesting possibilities.Submit contentSubmit

claude-jev

A Claude Code plugin that hands small judgments to Jev: prompt routing, subagent model tier, rule checks on edits, and a compaction that keeps the rows Jev selected.

Added to Jevfast

How it uses Jev

One request per prompt carries the prompt and the previous turn as state and asks intent as a choice among chat, lookup, fix, feature and ops, scope as a 0-2 score, and needs_tools as a noul. A subagent spawn adds four brief checks and a model-tier choice, edit and stop hooks ask rule-specific questions, and a routing hint is injected only above 0.75 confidence.

What Jev decides

Prompt routing bundle (illustrative)Example answers · not a recorded Jev response · Source ↗
Question 1 · intent
YOUR APP
INSTRUCTION

What kind of request is this for an AI coding assistant?

STATE

The state is the prompt, and for follow-ups the previous turn, so the routing questions can use what the conversation is already about. The three probes below are asked together in one request.

JEV · CHOICE
  1. chat
  2. lookup
  3. fix
  4. feature
  5. ops
Question 2 · scope
YOUR APP
INSTRUCTION

How much work does fulfilling this request take?

STATE

The state is the prompt, and for follow-ups the previous turn, so the routing questions can use what the conversation is already about. The three probes below are asked together in one request.

JEV · SCORE
02
Question 3 · needs tools
YOUR APP
INSTRUCTION

Must the assistant use tools such as reading files, searching, running commands or editing code, rather than just replying?

STATE

The state is the prompt, and for follow-ups the previous turn, so the routing questions can use what the conversation is already about. The three probes below are asked together in one request.

JEV · NOUL
YesNo

App workflow

  1. Hook the prompt

    The plugin's UserPromptSubmit hook takes the prompt, with the previous turn as context, and prepares one request.

  2. Route with confidence

    Intent, scope and needs_tools come back together, and a routing hint is injected only when the intent answer clears 0.75.

  3. Guard the work

    PreToolUse, edit and Stop hooks ask the rule questions that need judgement, and subagent spawns add brief checks and a model-tier choice.

  4. Fail open

    Any error, missing key or timeout produces no output, so the host keeps its own default behaviour and no hook can block a prompt.

Why it is interesting

It moves several small host decisions into typed questions with explicit confidence floors, so the expensive model is asked only what needs judgement and an uncertain or unavailable answer leaves Claude Code's normal behaviour untouched. The compaction hook shows the same idea applied to context: the rows Jev kept replace the summariser's output.