A Claude Code plugin deletes context instead of summarizing it

A new Claude Code plugin deletes low-value tool calls from context instead of summarizing them. Each call gets scored by a small decision model.

Long agent sessions break when a summary drops the one string needed three hours later. Summarizing is lossy by definition, and you don't control which loss you get.

fast-jev-compaction skips that trade. Released September 17 and at 4,642 stars by September 20, it hooks Claude Code v2.1.274 and up and never rewrites history. It sends each non-pinned tool call to TypeSafe's jev-latest model, which scores two things: keep the call, keep the result verbatim. Anything below 0.5 gets deleted. What survives gets truncated to 300 characters. Tool results get swapped for stubs like ok, 4213 chars (omitted), holding the state under a 25k-token budget.

The same idea is moving past compaction. jev-use routes whole agent steps to a decision model whenever no text output is needed. Per jev-use's published benchmarks, a PreToolUse shell gate denies dangerous commands in about 230ms at zero LLM tokens. Its compaction pass judges 200 messages in 7 calls, taking a window from 94% to 44% with 3/3 recall.

Open versions of the underlying model followed fast. Jared Palmer's kev trains in 1 hour 45 minutes on an M5 chip and scores 0.790 accuracy against the Jev reference's 0.857. It speaks the same wire format, so an existing client can point at localhost, not a hosted endpoint.