Jev scores tool calls KEEP/DROP to cut coding agent context 156k→62k — but deletes can backfire
Moretti_a · reddit · 2026-09-18
Coding agents die from 150k-token histories of Read/Bash/Edit/retry noise. Instead of lossy self-summarization, a new approach uses TypeSafe's Jev to score each tool call KEEP/DROP, keeping user/assistant text verbatim — demo shrinks context from 156k to 62k in seconds (repo: tamaratran/fast-jev-compaction). Caveats: dropping failed attempts can cause retry loops; mid-history deletes blow the KV cache and may cost more than the noise; and chopping hidden reasoning traces can make Claude Code dumber, not cheaper. It's a fast classifier over history, not a compressor.
More from coding & agent
- Hypit: open-source framework lets AI agents clone viral videos into 100-variant workflows — nikola_mr64990 · 2026-09-18
- OutlierKit kills its own product, relaunches as API + MCP for agents — ayushtweetshere · 2026-09-18
- OutlierKit launches API and MCP to turn any AI agent into a YouTube strategist — ayushtweetshere · 2026-09-18
- Abandoned AI agent projects are uniquely hard to hand off, evaluator observes — Mariav_Dowdf · 2026-09-18
- PCP v0.1: A Draft Protocol for Purpose-Bound, Revocable AI Agent Authority — sierracatalina · 2026-09-18
- Legatus v0.1: open-source transport-independent coordination protocol for delegated agent work — sierracatalina · 2026-09-18