Fast Jev Compaction — tamaratran

It never rewrites anything. Only tool calls and tool results that Jev marks as no longer needed are deleted, or truncated to a head plus a one-line note. User and assistant text stays verbatim and in order, and text messages are abridged only in the state Jev sees, never in what comes back.

It throws rather than guessing. Jev failures, malformed answers, a missing key, or a history that cannot be fitted into the state ceiling all raise, and the caller decides the fallback. The Claude Code hook catches that and falls back to the built-in summary, showing a fallback toast with the reason.

Only with function hooks enabled, which the README treats as an early-access feature. It needs Claude Code 2.1.274 or newer, CLAUDE_CODE_ENABLE_FUNCTION_HOOKS set to 1 wherever Claude Code runs, and TYPESAFE_API_KEY in the same settings file. Older versions have no hook point to replace the summary.

The README's own caveat is that calibration is a group-level property, not proof that a result is safe to delete. Two guards limit the damage: the newest six messages and the first message are pinned and never touched, and a call survives whenever its keep probability clears the threshold even if its result gets truncated.

npm install fast-jev-compaction, export TYPESAFE_API_KEY, then call compactMessages(transcript, { preserveRecentMessages: 4 }) and read result.messages, result.decisions and result.stats. The Message type is a subset of Claude Code's SessionMessage, so a real session transcript passes in as is. Options cover apiKey, model, baseUrl, fetch, goal defaulting to the last three user prompts, keepThreshold at 0.5, preserveRecentMessages at 6, maxStateTokens at 25000, maxRequestTokens at 30000 and truncateHeadChars at 300. reductionRatio lets you decide whether the run was worth it.

In stages, each applied only if the previous was not enough: tool inputs truncated to 1000, then 200, then 60 characters; long texts abridged to head plus tail, oldest non-pinned messages first; old non-pinned messages collapsed to a [...] N chars omitted note; old tool calls reduced to one line each such as t12 Read file_path=src/a.ts, ok 480ch; old call-less messages left out; and runs of old call-only messages folded together. Tokens are estimated without a tokenizer, a word per six letters, calibrated to land a little above Jev's own counts.

Two noul questions per non-pinned call: should the call stay, knowing it was made with its input still matters, and should the result stay verbatim, meaning its contents are still needed and re-running the tool would not do. Against keepThreshold, a result over the threshold keeps both, otherwise a call over the threshold keeps the call and truncates the result to truncateHeadChars plus a one-line note, and otherwise both are removed. A message that loses all its content is dropped, untouched messages come back as the same objects, and no result is ever left without its call.