Question 1

It deletes tool calls based on model probabilities - what stops it dropping the one call that wrote my file?

Accepted Answer

A hardcoded PROTECTED_TOOLS set, not the model. On a 1,128-call session the output was structurally valid with zero orphans and every user and assistant message byte-identical, yet all nine state-changing calls were gone and the transcript no longer recorded that anything had been written. The README notes Jev scored write_file keepResult 1.00 in one session and dropped it in another, so protection lives in code: if the model wants a protected call gone, the call stays and only its result is truncated, because the call carries the arguments and you lose the acknowledgement, not the fact.

Question 2

Can I see what it would remove before it writes anything?

Accepted Answer

Yes. A bare --session run is a dry run that measures and writes nothing, and --out is what produces a compacted transcript. --dump writes every decision with its keepCall and keepResult values, so you can read the verdicts rather than trust the reduction ratio. The README is blunt that the ratio is not the test: Konsistenz and Integritat must read zero orphans and true, or the compaction is broken no matter how much context it saved.

Question 3

Why does setup.sh compile the upstream library instead of npm installing it?

Accepted Answer

Because the package the upstream README names is not on the registry; npm install fast-jev-compaction returns 404. setup.sh clones and builds tamaratran/fast-jev-compaction from source, which needs Node 18+ and network access. This project is adapted from that library under MIT, and the adapter reads TYPESAFE_API_KEY from the environment only, never writing it to a file.

Question 4

What happens to a long session that will not fit in one request?

Accepted Answer

The upstream library sends the whole conversation state on every request and throws when it will not fit, roughly 31,976 tokens against a 25,000 limit on a 2,242-message session. The adapter catches that error and splits into windows of 2, 4, 8 until each fits, since verdicts are per call and local. Two pitfalls are documented: short ids restart at t1, t2 per request, so verdicts must be keyed by tool_use_id, and only the last window carries the live end of the session, so it keeps the full --preserve count while the others run with --preserve 1.

Question 5

Which sessions does it read, and does it write them back?

Accepted Answer

It reads Hermes sessions from ~/.hermes/state.db, in the sessions and messages tables, and never writes back to the database. Output is a JSON file you name with --out, and a legacy JSON file under ~/.hermes/sessions/ still works as input.

Question 6

What are the default threshold, preserve, and truncate values?

Accepted Answer

threshold is 0.5, the keep probability required to keep a call or result; preserve is 6, the newest messages never touched; and truncate is 300, the characters retained from a dropped result. All three are overridable flags.

Question 7

What does it guarantee about user and assistant text?

Accepted Answer

No text is ever rewritten, so user and assistant messages come out byte-identical, and no entry is orphaned: a result never outlives its call and a call is never left without its result. The report's Konsistenz and Integrität fields are the real test.

Jev Compaction — picaye

About this project

Highlights

Questions about this entry

More Jev resources like this

From the Jev community