Jev Compaction

Choosing turns with Jev, and what actually makes a handoff work

Jev cannot write a summary. It can mark each turn of a transcript:

• keep: carries a decision, a constraint, a preference, unfinished work, or an exact value, path, id, command or error that later work depends on.
• summarize: background whose gist matters. In the digest this is the turn's first 400 characters, nothing more. Jev writes no gist.
• drop: chatter, superseded attempts, repeated output.

It judges a long turn on its first 350 and last 350 characters, redacted, 40 turns per request, and sees no other turn while it does.

Read this before you use it for a handoff

This skill used to say a handoff written from Jev's digest "stops losing the one line that mattered". We measured that on seven real sessions and 104 recall questions (scorecard) and it was wrong:

Jev's marks did beat the same marks handed out by recency (11 questions to 4), so the judgement is real. The digest built around it clips every other turn to 400 characters, and that cost more than the judgement earned. Nous Research found the same shape with a different Jev design (hermes-agent PR 116246).

So, for a handoff:

1. Give the writer the whole dialogue. A current flash model reads 100,000 characters for about a cent. Do not pre-filter it.
2. Ask for up to 1,200 words, five headings: Working on, State, Decisions, Pointers, Next. At 400 words the capsule was full whatever the writer had read.
3. Name the session in the handoff and say it is searchable. One search was worth 16 to 33 points to every handoff we tried, and a session with no handoff and one search beat every handoff without one. On Hermes: session_search(query="..."), then session_search(session_id=..., around_message_id=...). Passing query together with session_id ignores the query.
4. Do not append a list of "identifiers seen". It looked free and obvious; it changed nothing with a handoff and cost 13 points without one.

The hermes-handoff plugin does all four. HANDOFF_JEV=1 puts the Jev pre-pass back if you want to compare on your own sessions with evals/compaction/run_eval.py.

When this skill is still the right tool

When the size is fixed and something has to go: a small local writer, a context you cannot grow, a digest for a person to skim. There, choosing turns with Jev beat choosing them by recency.

1. Get the transcript as a list of {role, content} messages. On Hermes: hermes sessions export --session-id <id> --format jsonl -.

2. Select:

   • Hermes: call jev_compact_select with messages.

   • Anywhere else:

     jev compact-select --digest < transcript.json      # {"messages":[...]} or a bare list
     

3. Write from digest. [KEEP VERBATIM] lines go in unchanged. [background] lines are clipped already; treat them as context, not as the record.

4. The digest is cut to its last 24,000 characters by default, oldest first, keep lines included. Pass a larger limit if early keep lines matter.

Guarantees

• The last six messages are always kept (keep_last); system messages are always kept.
• Nothing is dropped unless Jev was confident (0.7+). An unjudged turn is marked summarize, never drop. Summarize still means clipped to 400 characters.
• Turns that look like they hold a secret are not sent to Jev.
• Jev down: every turn comes back summarize. That is a worse input than the plain transcript, so on status: "fail_open" use the plain transcript instead.
• status: "partial" means some batches answered and some did not; the ids in unjudged sat at the summarize default with nobody judging them. Treat it like fail_open unless unjudged is short and you can see it does not cover the turns you care about. It used to report ok in this case, so one good batch hid every failed one.

When to compact at all

should_compact is arithmetic, not a model call: compact at 60% of the window, urgently at 85%. Do not ask a model whether the window is full.