JEV GUIDE

Jev for Claude Code context compaction: choosing a project

Two different approaches appear under “Jev Claude Code context compaction”: filtering fresh tool output and pruning accumulated history. Locate your context problem before choosing an integration.

Jev Gallery · Reviewed 2026-09-19

Long incoming output: inspect Winnow

Winnow judges relevance before tool output enters context and caches hidden text for recall. Start here when one file read introduces large amounts of unrelated content.

Check whether relevant code remains visible, hidden-content notices are useful and recall restores the full original. Character reduction alone misses the cost of losing required information.

Source: Winnow · reviewed source ↗

Accumulated history: inspect fast-jev-compaction

fast-jev-compaction judges old tool calls and results for retention, truncation or removal while preserving retained text. The repository also includes Claude Code hook integration.

Check call/result pairing, recent messages and user constraints. Distinguish trimming the state shown to the judge from rewriting the final conversation returned to the assistant.

Source: fast-jev-compaction · reviewed source ↗

Compare on the same tasks

This table compares reviewed implementation paths, not measured speed, quality or cost.

Compare on the same tasks
Check	Winnow	fast-jev-compaction
Integration point	Incoming tool output	Accumulated tool-call history
Main acceptance check	Relevant blocks and exact recall	Constraints and valid call/result pairs
First fixture	Large file with relevant sections	Synthetic conversation with stale results
Shared concern	False removal and API failures	False removal and API failures

A practical reproduction sequence

Use synthetic data and a fake judge to verify setup and restoration first. A simulated compression ratio proves no live relevance accuracy.

Mark required retry/backoff lines in a synthetic file.
Mark the “do not touch legacy” constraint in a synthetic conversation.
Record an uncompressed baseline, then compare task completion and retained information.
Test failure handling and preserve a way to restore the original context.

Current evidence boundary

Local checks passed 80 Winnow tests and 29 fast-jev-compaction tests using simulated judgments. Winnow’s demo cache restored the original text exactly. No live Jev call or complete Claude Code workflow was run, so no actual cost-saving ratio is established.