you're paying Claude Code to play hide-and-seek with your own codebase. Find the payment code. Filter a wall of tool output. Decide which old result still matters. All of that consumes context and model tokens before the bug is fixed. Four community projects give Jev parts of that first pass, so your coding agent can work from selected evidence: → give Claude decision tools / jev-mcp Inside Claude Code: /plugin marketplace add rashedInt32/jev-mcp /plugin install jev@jev-mcp Configure a TypeSafe key via TYPESAFE_API_KEY. Claude gets tools for classification, scoring, yes/no checks and batches of questions. Check the connection with claude mcp list. github.com/rashedInt32/jev-m… → find code by asking what it does / jevgrep In your terminal, with Node.js 22+: npm install -g @dzhng/jevgrep jg auth jg "how are payments retried?" ./my-project It judges relevance across the repository and returns file locations plus source excerpts. The agent still reads the evidence and makes the change. If you already know the exact symbol, ordinary search may be enough. github.com/dzhng/jevgrep → teach the agent when to delegate / jevify npx skills add altryne/jevify --skill jevify Choose your agent in the installer, then start a new session. The skill teaches bulk relevance checks, ranking and routing. It needs a working Jev runtime; installing it doesn't automatically intercept every tool result. github.com/altryne/jevify → select what stays in context / fast-jev-compaction Jev helps choose which tool calls and results to retain, truncate or remove. Retained text stays verbatim. The Claude Code plugin has function-hook requirements, and can fall back to the built-in summary. Check its setup before installing. github.com/tamaratran/fast-j… The receipt worth reading: jevgrep's author reported the same 8/10 solves in a ten-task Python SWE-bench comparison, with total cost falling from $7.62 to $5.66 — 25.8% lower, including estimated Jev cost. That was the 0.4.3 run with Sol, not a Claude Code test or a guarantee for your repo. github.com/dzhng/jevgrep/blo… The split is useful: Jev selects and checks; the coding model inspects the evidence, implements and tests. Retrieval and compaction can miss something important, so compare results on the same task before trusting the savings. Pick the tool for the part of your session that keeps repeating. My Jev article below goes deeper into this decision layer. Save the commands for your next unfamiliar repo.