Overview
jev-pruner is Tamara Tran's hook that sits between shell completion and the LLM. It chunks stdout, asks Jev a Noul question per chunk about whether any line must remain visible, keeps chunks above threshold, and archives the full output locally for forensics.
Problem: Agent shells return megabytes of build logs and test output; stuffing it all into context drowns the model before anyone reads the failure line.
Built for: Claude Code and Codex users who want Bash hook compaction on command stdout without summarizing away stack traces.
First indexed on Jev Directory: 2026-09-22
Creator and team
- Name
- Tamara Tran
- Handle
- @tamaratran
How Jev is used
- Role in the product flow
- Bash hook compaction on command stdout before tool results return
- Primitives
- Noul
- State in
- Chunked stdout from the finished command plus hook context about what the agent was trying to learn.
- Decision out
- Trimmed stdout for the model with full transcript archived; chunks dropped only when Jev says nothing important remains.
- Command runs to completion in Claude Code or Codex Bash tool
- Hook intercepts stdout before the LLM receives it
- Split into chunks; parallel Noul: does any line in this chunk need to stay?
- Emit surviving chunks; store complete stdout in archive per README
jev-pruner complements fast-jev-compaction, it does not replace it. Compaction attacks long-lived tool rows in the transcript; pruner attacks one-shot terminal floods right after they happen. Jev never rewrites log lines into summaries. It only votes keep or drop at chunk granularity, which preserves exact error strings when they matter. If you already run tamaratran/fast-jev-compaction for history hygiene, add pruner when npm test or cargo build spews more text than your context budget allows. Same TypeSafe key story as Tamara's other plugins.
Features and stack
Features
- Claude Code and Codex Bash hook integration
- Chunk-level Noul gates, not LLM summaries
- Full stdout archive alongside trimmed view
- Sibling project to fast-jev-compaction (different repo)
Stack
- TypeScript
- Claude Code hooks
- Codex
- TypeSafe System One
Pricing: Bring your own TYPESAFE_API_KEY; cost scales with chunk count per command.
Links
FAQ
- How is this different from fast-jev-compaction?
- fast-jev-compaction trims paired tool_use and tool_result rows in the transcript. jev-pruner trims fresh shell stdout in the Bash hook path. Different repos, same Noul philosophy.
- Do I lose the full log?
- README describes archiving complete stdout while only the Jev-approved chunks return to the model.
Related learn guides
Original Jev guidance that pairs with this product pattern.
- Jev use cases
The patterns builders actually search for: moderation, routing, triage, RAG verify, and agent gates.
- System One model
The model family behind Jev: parallel typed questions, one forward pass, probabilities you can threshold.
Related products
Hand-picked neighbors with rich profiles or overlapping tags.