KamilPostrozny/pi-fast-jev-compaction
User repository on GitHub
Fast JEV compaction extension for pi
- Forks
- 1
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.
CommunityJev Agents & MCP
Pi Fast Jev Compaction — KamilPostrozny
Pi package that uses Jev to decide, call by call, which old tool results can leave model context, with a savings gate that rejects prunes too small to be worth the prompt churn and bounded excerpts from status documents.
- Stars
- 2
- Language
- TypeScript
- License
- MIT
- Published by
- Community
- Author
- KamilPostrozny
- Last verified
- Oct 2026
This is one of several Jev listings for the same project. The canonical entry is Pi Fast Jev Compaction — joelhooks.
About this project
A Pi package port of tamaratran/fast-jev-compaction, adapted to Pi's non-destructive context hook. It uses TypeSafe Jev to decide, tool call by tool call, which old calls/results still need to remain in model context. User and assistant prose is not summarized or rewritten by this extension. Pi's persisted session transcript stays intact.
Quoted from the project's README. Read the full README.
Works forJev for Classification
Used inJev for Software & DevTools
- Languages
- TypeScript
Highlights
- An automatic prune must change real results and save at least the hysteresis gap — 5,632 estimated tokens at a 112,640-token window with 16,384 reserve.
- A rejected proposal commits nothing, defers no deletions, uses up its one automatic attempt and lets native compaction proceed instead.
- Upstream Jev never sees result bodies: it gets call inputs plus output size, success and error metadata, so an all-drop pass is not content-aware.
- Markdown status, handoff and progress documents keep deterministic heading-based excerpts, capped at 2,400 characters per result and 7,200 in total.
Watch out
The excerpt logic recognizes a narrow set of English Markdown headings and filenames, so arbitrary command output, partial reads without the relevant heading and facts outside the character budget can still be lost. Committed excerpts stay fixed until their calls leave the active branch and are not a continuously updated task database, and native summaries may still omit facts — re-read the actual status document before editing it or treating it as current. The token figures are local estimates used for the savings gate, Jev budgets and diagnostics, not a tokenizer-accurate promise or a cache-cost predictor; Pi usage alone controls pressure, cancellation and post-turn validation. The hysteresis margin is described as an initial calibration to evaluate, not an optimum. 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.
An estimate, stated as such. Savings are measured between the already-pruned logical results and the proposed results in estimated net tool-result tokens, and the README calls the gate an efficiency heuristic rather than a tokenizer-accurate promise or a cache-cost predictor. Those local estimates are used only for the savings gate, Jev budgets and diagnostics — Pi's own usage signal exclusively controls pressure, native cancellation and post-turn validation.
No. The extension decides, tool call by tool call, which old calls and results need to remain in context. User and assistant prose is not summarized or rewritten, Pi's persisted session transcript stays intact, and the package is built on Pi's non-destructive context hook. What it adds is retained excerpts from status documents, labeled as historical notes rather than current file contents.
A cost gate and bounded excerpts. An automatic proposal must change actual results and save at least the resolved hysteresis gap — 5,632 estimated tokens at a 112,640-token window with a 16,384 reserve — otherwise it commits nothing and native compaction takes over. Successful reads of Markdown files whose basename contains a status, handoff or progress component keep deterministic heading-based excerpts, complete paragraphs included, capped at 2,400 characters per result and 7,200 across committed excerpts.
Malformed or out-of-range probabilities fail open, so a bad reply cannot delete context on its own. The 0.5 scoring threshold, the scoring questions and the six-message protection are unchanged, and the 0.7.4 release adds diagnostics instead of new deletion power: probability distributions, per-tool summaries, estimated versus minimum saved tokens, and the acceptance reason, available through /jev-status and /jev-diagnostics without logging result bodies.
No. Upstream Jev receives call inputs plus output size, success and error metadata rather than result bodies, so an all-drop pass is not content-aware evaluation of every result's facts.
It must both change actual results and save at least the resolved hysteresis gap. At a 112,640-token window and 16,384 reserve that is 5,632 estimated tokens, with no fixed percentage-of-history floor.
Deterministic excerpts from Markdown headings for status, handoff or progress files, bounded to 2,400 characters per result and 7,200 total, labelled as historical notes rather than current file contents.
More Jev resources like this
20GitHub stars
TypeScript SDK and CLI coding agent where Jev only picks options while your code runs effects, building programs as syntax trees.
Jev Agents & MCP#community#awesomejev#catalog-import
Community
rhighs/jev-codeTypeScript
rhighs/jev-codeTypeScript22GitHub stars
Runnable agent harness that moves pick, score and yes-no loop choices to Jev or Laya in milliseconds while reserving the language model for writing, compiling DECISION.md into executable questions.
Jev Agents & MCP#community#jevbest#catalog-import
Community
zjunlp/JevLoopTypeScript
zjunlp/JevLoopTypeScript35GitHub stars
Diff-aware test selector that sends the git diff to Jev with one question per test, then prints the native filter flags and hands them to vitest, jest, node, bun, Playwright, cargo, or go test.
Jev Agents & MCP#community#jevbest#catalog-import
Community
mizchi/jev-test-filterTypeScript
mizchi/jev-test-filterTypeScript13GitHub stars
Pi extension that classifies each task with Jev over OpenRouter and routes it to a Pareto-optimal model balancing measured quality, cost, and latency under a role policy.
Jev Agents & MCP#community#jevbest#catalog-import
15GitHub stars
Node CLI that lints a file tree against fuzzy natural-language rules in one Markdown file, judging each matched file with a swappable TypeSafe-backed judge.
Jev Agents & MCP#community#jevbest#catalog-import
Community
tyler-dot-earth/patdownTypeScript
tyler-dot-earth/patdownTypeScript106GitHub stars
Decision layer exposing Jev's typed yes-no, choice, and score answers to DeepSeek Harness and MCP hosts, offline by default with every network transmission disclosed per feature.
Jev Agents & MCP#community#jevbest#catalog-import
Community
PerryLink/jevcoreTypeScript
PerryLink/jevcoreTypeScriptFrom the Jev community
Posts from builders shipping with Jev right now.