Claude Context Cut from 1M to 86K

The post claims the reduction, not its safety. A critic in the thread, Theo from t3.gg, argues the opposite case: compaction is not a filter, and its job is to clean history rather than delete entries wholesale. The post does not address that objection or show a task surviving the prune.

A single self-reported session with no repeat runs and no baseline. Nearly 1M down to 86K in about a second describes one unusually large session being reviewed, and the post gives no token counts before and after, no success rate on the work that followed, and no failure cases.

A Claude Code plugin from tamaratran/fast-jev-compaction, the same project whose author announced it as scoring every tool call and dropping the irrelevant ones. The post hands you a one-line install prompt but no configuration, no confidence threshold, and no way to inspect or revert a prune.

The post mentions neither credentials nor price, only that you ask Claude to install and configure it. A Jev call per tool call would bill against the published $0.042 per million input tokens, but the post never says how many calls a prune makes, so the per-run cost cannot be worked out.

Tool calls. The post describes reviewing all the unnecessary tool calls, and the original announcement frames it as scoring every tool call and dropping what is judged irrelevant. Neither post says what happens to ordinary conversation text sitting between those calls, so tool calls are the only review scope either of them describes.

Independent. Alex Volkov is a user and podcaster sharing an install prompt, not the author of fast-jev-compaction, which lives in tamaratran's repository. He also notes that Jev helped edit the podcast episode he recorded about it, so read the post as an enthusiast's account of somebody else's tool rather than a maker's launch note.

Jev does, by scoring each call, and the drop rests on the model's judgement rather than a rule you set. Neither the post nor the original announcement states the confidence threshold, whether you can tune it, or how a borderline call is treated, so the definition of unnecessary stays inside the plugin.