Skip to content

Jev directory

Fast Compaction Dsh

Verdict-based context compaction for DeepSeek Harness: instead of a lossy summary, jev-latest scores every tool call and result — kept verbatim, truncated to 300 characters or dropped — and falls back to compaction-basic on any failure.
Stars
4
Language
TypeScript
Reading time
3 min
Published by
Community
Author
kolawong
Last verified
Oct 2026

Highlights

  • Two Noul questions per non-pinned call: should the call stay, and should the result stay verbatim, batched so state plus questions stays under maxRequestTokens.
  • Decisions: keepResult at or above keepThreshold keeps both verbatim; else keepCall keeps the call and truncates the result to 300 chars; else both drop.
  • User and assistant text is never removed, and everything kept stays verbatim — nothing is ever rewritten.
  • Defaults: preserveRecentMessages 6, maxStateTokens 25k, maxRequestTokens 30k, truncateHeadChars 300, keepThreshold 0.5.
  • Measured on a real region, the Jev scoring round-trip takes about 700 ms; on a missing key, Jev error or insufficient reduction, it falls back.

Watch out

Copy reviewed against the repo README on 2026-10-15; the captured README was truncated inside the agent preset patch YAML. The 700 ms figure is a single measurement on a real region. The GitHub license metadata reads NOASSERTION, and the README describes the project as a DSH port of tamaratran/fast-jev-compaction (MIT).

Questions about this entry

Short answers drawn from the linked project or page.

Reactions & coverage

Posts, threads, and videos about this entry from around the web.

X: The compaction plugin, in Chinese

X: The compaction plugin, in Spanish

From the Jev guides

Original write-ups on what this entry is for.

From the Jev AI model community

Posts from builders shipping with Jev right now.

Follow @typesafeai