Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-09-20 · high
It scores tool calls with a small model and deletes the low performers, already passed 4,642 GitHub stars in three days.
Why now: The plugin went up on GitHub September 17 and had 4,642 stars by September 20.
fast-jev-compaction scores every tool call in a Claude Code session instead of rewriting the whole conversation into a summary. Standard compaction turns long sessions into prose, dropping exact file contents, command output and error text an agent might need again. With this plugin, a lost detail traces to one deletion, not a paragraph that dropped half its source with no flag anywhere in the transcript.
Built for Claude Code v2.1.274 and up, the plugin sends the conversation to TypeSafe's jev-latest model. It returns two probabilities for each non-pinned tool call, one for keeping the call and one for keeping its result verbatim. Anything scoring under 0.5 gets deleted outright, and calls that survive get truncated to 300 characters.
The scorer itself doesn't see raw tool output. Results are replaced with stubs like "ok, 4213 chars (omitted)." The state then runs through three compression stages, truncate, abridge, collapse, to stay under 25,000 tokens. The full request has to stay under a 30,000 token budget.
What's not in the repo: how jev-latest was trained, its parameter count, or whether it runs anywhere besides TypeSafe's hosted endpoint. If it's proprietary and hosted-only, every compaction decision routes through a third party's inference call, a dependency for something that runs on every long session.
Each link below shares sources, entities, or timing with this story.
The open-source tool now sits at 98.9k GitHub stars and pairs a skill with a proxy that compresses code, logs and JSON before they hit the model.
The same release adds a plugin eval command and fixes a permission rule that leaked across Claude Code's settings sources.
Two of the six hid commands from the approval dialog itself, using tab padding, invisible Unicode, and split Bash strings.
Version 13.24.0's manifest said one thing, its bundled scripts said another, and the mismatch sent the background worker into a kill and respawn loop on every hook event.
Compaction is where long sessions go to die. The model summarizes what happened, the summary drops the exact string you needed three hours later, and you don't find out until the agent confidently references a file path that never existed. It's the biggest source of silent con...
The release also adds a command that migrates Python projects off the old anthropic 0.x SDK and stops WebFetch from serving stale content all session.
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.