← BACK TO CASE INDEX

AGENT CONTEXT / COMMUNITY CASE

fast-jev-compaction

Context compaction for Claude Code that judges tool calls and results, then keeps the surviving text verbatim instead of rewriting it.

6,724 GITHUB STARSREPOSITORY OWNER tamaratranEXPERIMENT
VIEW ON GITHUB Stars are a verified snapshot · 2026-09-25
Repository preview of the fast-jev-compaction repository
Repository preview · tamaratran / fast-jev-compaction
OVERVIEWAGENT CONTEXT

Long agent sessions accumulate tool calls, command output, and file reads. Most compaction strategies replace that material with a rewritten summary, which paraphrases text the agent may still need exactly.

fast-jev-compaction adds a judging pass instead. It evaluates tool calls and their results, discards what is stale or irrelevant, and carries the kept text forward as-is.

WHAT JEV DECIDESJUDGMENT

Jev performs the keep-or-discard judgment over each tool call and result. The compaction path stays deterministic: the judge decides what survives, and surviving text is not rewritten.

Read the full projecthttps://github.com/tamaratran/fast-jev-compaction
VIEW ON GITHUB

Case notes by JevCases. Repository snapshot dated ; claims not independently reproduced.

← RETURN TO THE CASE INDEX