Why not just summarize or delete old context?

Summarization rewrites the middle, breaking provider prompt cache and risking invented details, while deletion is unrecoverable. This keeps originals, returns them byte for byte on expand, and confines rewrites to the request where price math says they pay off.

How do admission and recall actually work together?

Admission profiles each segment for keep-worthiness, type, role, lifespan, and injection, turning low-value runs into one-line pointers and quarantining injections. Recall then filters on the profile, shortlists lexically, and has Jev rerank from loose descriptions.

What is supersession and does it cost model calls?

From tool-call arguments alone with no Jev call: reruns of the same command, rereads of the same lines, or files edited since are recorded as obsolete. About a fifth of tool output goes stale this way through the agent's own edits.

Did it actually hold up on a real agent workload?

On SWE-bench Verified with OpenCode the profiled gate removed about a quarter of tool output with resolve rates unchanged at 20 of 23, caught 17 of 18 planted injections, and put recall targets in the top 3 for 80% of queries against 49% for BM25.

How do I try it with no API key?

Clone the repository, run uv venv .venv and uv pip install --python .venv/bin/python -e '.[dev]', then .venv/bin/python demo.py. The offline tour runs one act per mechanism with no API key.

How do I turn it on safely?

Start with shadow_only=True in GateConfig, which scores and logs everything but changes nothing. Replay the shadow log to choose thresholds, then switch it on; every decision is logged so calibration is fitted from what the agent later used.

When does a work-area rewrite actually pay?

Only when dropped bytes times turns left times cache price beats the rest of the tail times input minus cache price. The rewrite applies to the request, and the stored conversation is never changed.

CommunityJev Demos & Experiments

Jev Compaction — Waxmell114514

Context manager where Jev scores tool output into keep, pointer, or quarantine classes while kept lines stay verbatim and anything moved out can be expanded back byte for byte.

See highlights

Open github.com

Stars: 4
Language: Python
License: MIT
Published by: Community
Author: Waxmell114514
Last verified: Oct 2026

Waxmell114514/jev-compaction

User repository on GitHub

View on GitHub

A context compactor that can only score, never write — so an agent's memory can't hold a fact the transcript never contained. Working demo, runs offline.

Forks: 1
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.Homepage

About this project

Jev decides what a coding agent keeps in its context. It can score text but never write it, so everything the agent sees is original, and anything moved out can be brought back byte for byte. ▶ How it works, and what it did on SWE-bench, or run the tour locally with no API key: An agent's context fills up with tool output: test logs, file listings, whole source files. The usual fix is to have a model summarise old turns. Two things go wrong:

Quoted from the project's README. Read the full README.

Tagscommunity awesomejev catalog-import

Works forJev for Classification Jev for Ranking Jev for Tool calling

Used inJev for Software & DevTools

Languages: Python

Highlights

Five-mechanism pipeline of admission, recall, supersession, work area, and calibration in one library.
Work-area rewrite applies to the request only when dropped-bytes times turns-left beats cache math.
HTTP sidecar exposes admit, observe, expand, recall, workarea, and stats for any harness.
Shadow-only mode scores and logs everything before you enable any context changes.

Watch out

Session-length estimation, Bayesian threshold updates, and tiered memory are listed as not built yet; savings depend on turns taken and cache pricing. Copy reviewed against the repository README on 2026-10-04, not from running it.

Questions about this entry

Short answers drawn from the linked project or page.

Other listings of this project

Forks and copies of this project in the directory. This entry is the canonical one.

jev-compaction

Compacts Hermes transcripts by scoring every tool call with Jev: stale calls are dropped, stale results truncated, and protected tools such as write_file and patch are always kept. Nothing is ever summarised.

Jev Agents & MCP#community#jevbest#catalog-import

Community

picaye/jev-compactionJavaScript

More Jev resources like this

aside-jev

Browser workflow that keeps one Aside REPL session open, compacts page observations into allowed candidates, and has Jev pick the next action with confidence checks.

Jev Demos & Experiments#community#awesomejev#catalog-import

Community

himomohi/aside-jevPython

open-jev-typed-decision-engine

Open 150M encoder that reproduces the typed noul, choice, and score interface in one forward pass, trainable on a free GPU and runnable locally for nothing.

Jev Demos & Experiments#community#jevbest#catalog-import

Community

intikhab49/open-jev-typed-decision-enginePython

jev-kit

Claude Code kit that guards PreToolUse calls with a free pre-filter plus one Jev question for ambiguous cases, and adds sub-agent sizing, rechecks, search, daemon speedups, and monitoring.

Jev Demos & Experiments#community#awesomejev#catalog-import

Community

jonathanavis96/jev-kitPython

jev-ultrafast-mcp

MCP server where one browser_goal call hands a URL, a goal, and code-checked assertions to a server-side loop driven by a decision model over live page elements.

Jev Demos & Experiments#community#jevbest#catalog-import

Community

jiawei686/jev-ultrafast-mcpPython

jev-style

Local Qwen3.5 decision models in 0.8B and 2B sizes with fitted temperatures, systemone-shaped server, playground, guard, MCP tools, and eval harness.

Jev Demos & Experiments#community#jevbest#catalog-import

Community

lawrence3699/jev-stylePython

keys-mcode-continuous-context-browser-decision-enhancement-pack

Enhancement pack for the MiniMax Code CLI with TUI patches, continuous context renewal, a local-vLLM compaction fix, and a measured arm64 browser lane.

Jev Demos & Experiments#community#awesomejev#catalog-import

Community

drowzeys/keys-mcode-continuous-context-browser-decision-enhancement-packPython

From the Jev guides

Original write-ups on what this entry is for.

Per-Tick Decisions in Games

Sep 28, 2026

A game tick is a state plus a legal action list. How Clash Royale, Minecraft, and the earlier Mario, Pokémon, StarCraft, and drone loops ask Jev one step at a time.

Jev Demos & Experiments#gaming#agents#control

Read article

Back to all resources

From the Jev community

Posts from builders shipping with Jev right now.

Follow @typesafeai