A Claude Code plugin replaces compaction summaries with per-tool-call delete decisions scored by a cheap non-autoregressive model
fast-jev-compaction (created 2026-09-17, 4,642 stars by 09-20) hooks Claude Code v2.1.274+ and never rewrites history: it sends the conversation to TypeSafe's jev-latest decision model, which returns two probabilities per non-pinned tool call (keep the call, keep the result verbatim), then deletes only what falls below a 0.5 threshold and truncates surviving calls to 300 characters. The state shown to the scorer replaces tool results with stubs like 'ok, 4213 chars (omitted)' and is progressively compressed through truncate, abridge and collapse stages to stay under 25k state tokens and a 30k request budget. For builders this is a concrete alternative to summarization-based compaction, which loses verbatim detail and is the single biggest source of silent context corruption in long agent sessions.
↳ Follow the thread