Keep the pair verbatim
First and latest tool pairs, plus manual pins, remain complete. A candidate result above the keep threshold also keeps its call and full result.
Jev Context Compaction
Edit a conversation record and review which earlier tool results Jev suggests keeping.
Try another example
Start from an editable synthetic historyLoading a sample never calls Jev.
01 / SOURCE CONTEXT1,970 / 8,000 characters
Run this example first, then edit a message or tool result to see how the recommendation changes. Nothing is removed from your original.
History to review
user messageThe dashboard pagination test fails after a refactor. Please find the regression.Always kept
git statusmain clean; no uncommitted changesPinned · kept
rgsrc/query/paginate.ts:17 paginate(items, page)
tests/pagination.test.ts:42 expected page 2…May be reduced
assistant messageI found the shared pagination helper and will inspect its boundary handling.Always kept
test runnerAn older unrelated color-snapshot test passed. Its output is no longer relevant to the pag…May be reduced
read fileexport function paginate(items, page, size) { const start = page * size; return items.slic…Pinned · kept
test runnerFAIL pagination.test: page 2 expected item 11, received item 21. Current call stack points…Pinned · kept
Selected task, constraints, messages, tool inputs and results are sent to this site’s server and its configured Jev provider. Remove secrets or private data before running. No files or real agent sessions are read or changed.
The recommendation can be reviewed and reversed locally.Edit complete history · 7 records
Conversation & paired tools
Each tool record keeps its call input and result together. First and latest tool records are always pinned.
user messageThe dashboard pagination test fails after a refactor. Please find the regreAlways kept
git statusmain clean; no uncommitted changesPinned
rgsrc/query/paginate.ts:17 paginate(items, page) tests/pagination.test.ts:42 May be reduced
assistant messageI found the shared pagination helper and will inspect its boundary handlingAlways kept
test runnerAn older unrelated color-snapshot test passed. Its output is no longer releMay be reduced
read fileexport function paginate(items, page, size) { const start = page * size; rePinned
test runnerFAIL pagination.test: page 2 expected item 11, received item 21. Current caPinned
02 / REVIEW & EXPORTAWAITING EVALUATION
⌁
The complete original stays available
Edit the synthetic history, then evaluate. The result and probabilities will appear here without altering your input.
Review omitted evidence before using exported text. Shorter context is not proof of better answers, lower total cost, or lossless preservation. Nothing is written into an agent session.
What the local policy actually does
Keep the call and a result prefix
When the call is useful but the complete result falls below the threshold, the tool keeps the call input, a chosen result prefix, and an explicit truncation marker. If that would be longer, it keeps the original.
Remove the pair together
When both probabilities fall below the threshold, call and result leave the exported context together. The full original remains visible and Restore needs no Jev request.
Implementation
This browser tool and a host plugin solve different parts
Inspect a bounded draft
You paste or edit a task, constraints, messages and up to eight paired tool records. The entire bounded State is sent for at most twelve Boolean/Noul questions in one logical evaluation. No local files or real session history are read.
Work inside Claude Code hooks
fast-jev-compaction pairs calls and results by tool ID, pins boundaries, and stages shorter State representations. Its hook path depends on Claude Code function-hook support. This page does not install or reproduce that integration, its long-history batching, or its native-compaction fallback.
Filter new output earlier
Winnow filters tool results before they enter the agent context and can cache hidden output for recall. That differs from pruning an existing history; caching and recall are not features of this page.
Compacted context is a reviewable proposal; check what was kept before using it elsewhere.
Limits & privacy
A shorter record is a proposal, not a guarantee
Before and After use the same JSON serialization for character counts. Expand a decision to inspect the full removed result. Restoring shows the complete original; exporting never writes to an agent conversation. Historical tool output can be irreplaceable or describe a side effect, so human review matters even when a tool could be called again.
Keeping the original intact does not make a pruned export lossless. Fewer characters do not establish better answers or lower total cost. The threshold and preview length are application policy; changing them only recomputes the result locally.
What this demonstration can establish
TypeSafe documents Noul as a probability for a yes/no question and supports multiple questions over one State. This tool uses two independently judged questions per candidate. Noul has no separate Choice confidence. We do not treat a returned probability as proof that deleting evidence is safe.
The source review used fast-jev-compaction at e3f262a, Winnow at 51d80b9, and the Codex proof of concept at 7147a4b. The latter’s maintainer explicitly does not recommend it for real work. A published npm package exists for the upstream experiment, but its host compatibility and installation path were not independently exercised here; consult the upstream repository and current host documentation before trying it.
This is not a replacement for native model compaction or internal state handling, and it does not generate summaries or promise lossless compression. Community interest does not establish production quality. Model judgments for this new scenario have not been benchmarked here.