jev compact

Shrink an agent transcript without summarizing anything. Jev decides, per old tool call, whether the call and whether its full output still matter. Stale results are cut to a short head plus a note, stale calls are removed, everything else is returned exactly as it was. User and assistant text is never touched.

jev compact [transcript] [options]

When to use

Before handing a long session to another agent, to audit what a compaction would drop, or as the engine behind the Claude Code plugin's compaction hook, which replaces the built-in summary in-session. Summaries are lossy: a path, an exact error, or a constraint can vanish. This never rewrites, only deletes what Jev says is no longer needed, and the assistant can always re-run a tool.

Input

Form	Example
Claude Code session log	@~/.claude/projects/<project>/<session>.jsonl
Messages JSON	[{"role","text","toolUses":[{"tool_use_id","tool","input"}],"toolResults":[{"tool_use_id","text"}]}]; text and toolUses may be omitted

Options

Option	Meaning	Default
-g, --goal <text>	The ongoing task, so Jev knows what still matters	last three user prompts
--keep-threshold <p>	Keep a call or result when Jev's probability is at least this	0.5
--preserve-recent <n>	Newest messages never touched; the first is always kept	6
--max-state-tokens <n>	Budget for the history sent to Jev	25000
--max-request-tokens <n>	Budget for history plus one batch of questions	30000
--truncate-head <n>	Characters kept from a dropped result before its note	300
--min-reduction <p>	Below this ratio the result is worth_it: false	0.25
--concurrency <n>	Question batches in flight at once, 1 to 64	4
-o, --out <path>	Write the compacted transcript as messages JSON
--fail-on <list>	Exit 2 on low-reduction	none

How it fits Jev's 32k window

Jev never sees the raw transcript. The state is the whole conversation with every tool output replaced by a one-line note (ok, 4213 chars (omitted)). If that exceeds the budget it is shrunk in stages: tool inputs truncated, long texts abridged, old messages collapsed to a note, old calls reduced to one line, old call-less messages left out. Questions (two per candidate call) are split across as many requests as needed, each resending the same state. They run concurrently, at most --concurrency at a time (default 4), because every request in flight carries another copy of the state. Compressing the state never changes the output; it only shapes what Jev reads to decide. Token counts are estimates calibrated slightly above Jev's real counts.

Output

reduction, worth_it, stats (messages and chars before and after, kept/truncated/dropped/pinned counts, state tokens and stage, requests, ms), decisions[] (id, tool, action, keepCall, keepResult), messages (the compacted transcript).

Example

jev compact @session.jsonl --out compacted.json

61% smaller  528 → 341 messages · 1,204,311 → 470,902 chars
209 tool calls: 58 kept · 71 results truncated · 77 calls dropped · 3 pinned · state ~24944 tokens (old messages collapsed) in 5 request(s), 2140 ms

Call  Tool  Action    P(call)  P(result)
----  ----  --------  -------  ---------
t1    Bash  truncate  0.91     0.12
t2    Read  drop      0.08     0.03

In-session use: see plugin/hooks/README.md. Adapted from fast-jev-compaction (MIT).