Summarization is the default answer to a full context window: pay a model to write a shorter transcript. The Jev ecosystem's compaction tools reject that trade. They ask two binary questions per tool call, leave user and assistant text verbatim, and treat uncertainty as a reason not to touch anything. Jev 1.14's opt-in 262k context is in preview and not production-ready by TypeSafe's own account, so the judgment layer has a bigger alternative to argue with.
Keep or drop, never summarize
The family starts with fast-jev-compaction, a Claude Code plugin and npm library. Every non-pinned tool call gets two Noul questions: should the call stay, and should the result stay verbatim. State is fitted to 25k tokens in stages before anything is sent, requests are split under a 30k-token cap, and answers are merged back per call. A message that loses all its content is removed, and no result is left without its call.
save-token-jev extends the pattern with adapters for Codex, OpenCode V2, Claude Code, the Anthropic and OpenAI APIs, and a generic transcript format: each call is kept with its full result, kept with a bounded result, or removed together with its result, and nothing is rewritten. On Codex it scores before built-in compaction and injects retained context at SessionStart. omp-jev-compaction vendors the same core; the wider roster is those two questions against different hosts, all inside the loop the coding-agent survey maps.
Two questions, three outcomes
The decision shape is stable across ports. The DeepSeek Harness port's defaults show it: a keepResult at or above keepThreshold (0.5) keeps call and result verbatim; keepCall alone keeps the call and truncates the result to 300 characters; otherwise both drop. Six recent messages are preserved, state is capped at 25k tokens, requests at 30k, and everything kept stays verbatim.
The differences are in the failure posture. QuentinDanblon's Pi port sends unmodified messages on any uncertainty, including a Jev error or a state that cannot be fitted, and gives up about 10 points of net gain to remove a call only when it is provably re-runnable. winnow splits large Read, Bash, and Grep results into roughly 25-line blocks and asks one yes/no question per block; a hidden block becomes a three-line stub with a key that restores the text. Output judged to show an error is never hidden, and borderline blocks are kept. Several ports, save-token-jev included, fail open: a missing key leaves the request unpruned rather than wrong.
When they fire
compact-adviser asks two one-sentence questions per checkpoint — is the unit finished, and is this hands-on work — and composes them into a score whose threshold starts near 0.90 on an almost empty context and relaxes toward 0.50 as the host's own compaction point approaches. Hints are suppressed below a constant 40,000-token minimum, requests are capped at 32,000 serialized bytes, and an error is never treated as an affirmative judgment. Only Pi and Claude Code compact automatically; Codex and Grok are hint-only.
KamilPostrozny's Pi port gates the judgment itself: a prune must free at least the hysteresis gap — 5,632 estimated tokens at a 112,640-token window with 16,384 in reserve — or it commits nothing and lets native compaction proceed. Jev there never sees result bodies; it judges from call inputs plus output size and error metadata, so an all-drop pass is safe but content-agnostic.
What the tools measure
omp-jev-compaction has the most mechanical numbers: a 16,198-message replay produced 3 rewrites and 37 reuses over 40 requests, with 0 prefix breaks across 39 checks. A cache guard skips sessions already at 80% or more cache reads, and on 14 live sessions the project reports it protected $1.81 per request round while capturing $2.61; its OpenRouter route measured about 430 ms and $0.0000198 per decision request.
The QuentinDanblon port's ratios come from two screen-heavy sessions, about 90% tool results and mostly base64 screenshots: 31.8% and 37.7% net gain with background scoring, 33.6% and 40.9% on the request path, which cost 2,984 ms in the hook over a session against 322 ms in the background. Only the ratios are defensible, its caveat says. Past that the numbers thin: a single 700 ms measurement in the DeepSeek Harness port, a demo post with no independent benchmark, no measured results reviewed for the Codex packet tool.
The 262k counterargument
Jev 1.14, as summarized by MarkTechPost, raises context to 262k behind opt-in danger zone headers at an unchanged $0.042 per million input tokens, experimental, with TypeSafe saying plainly that none of it is production-ready. A larger window defers the choice; it does not judge whether an old tool result still matters, and the workload these tools were built for is a transcript whose bulk is stale tool output. Window size does not settle cache arithmetic either: one port's append-only decision ledger invalidates the provider's prompt cache from the first edited message, while the omp plugin keeps its truncations sticky so the prefix survives.
What none of it guarantees
Decisions are model judgments, not proofs, as one port's caveat puts it. Selection is lossy: the Codex packet tool says to keep the original input as the authoritative archive, and the savings-gate port warns that committed excerpts stay fixed until their calls leave the active branch. Privacy is part of the contract, too — installing the adviser is stated as consent to send checkpoint context to TypeSafe once a key is available, and the omp port sends conversation state including tool names, inputs, and message text on every reduction.
Used as designed — conservative thresholds, text untouched, failures ending in a no-op — the family is a cheap way to keep a long session honest. Its two questions per call ride the existing batching discipline, which keeps a decision at a fraction of a cent. The long-context header makes the layer optional; the measurements above should decide whether you take the option.
