User repository on GitHub
- Forks
- 0
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.
Jev directory
CommunityJev Integrations & Adapters
Pi Jev Compact
A Pi extension for selective, verbatim context compaction: Jev scores every tool call and its result, unneeded calls are dropped, and everything kept stays byte-identical. The README measured 15.5% reclaimed on a synthetic window, rising to 33.5% with prose included.
- Stars
- 0
- Language
- TypeScript
- Reading time
- 3 min
- Published by
- Community
- Author
- ilkerulusoy
- Last verified
- Oct 2026
- Languages
- TypeScript
- Platforms
- CLI
- Capabilities
- Open sourceRequires API key
Highlights
- Jev is asked two nouls per non-pinned tool call: call_tN (does knowing the call was made still matter) and result_tN (must the output stay verbatim).
- Decisions map to keep call and result, keep call with a truncated result, or drop both; what remains in the replay is byte-identical.
- By default only tool calls and results are candidates, matching prior art; /jev-compact text adds assistant prose under a stricter threshold.
- Measured on a synthetic 1,860-message window, reclaiming 15.5% from tool results became 33.5% when prose was added, per the README.
- The compaction writes a new session via ctx.newSession and replays survivors; the previous session file is never modified.
Watch out
Grounded in the truncated repo README and GitHub metadata read on 2026-10-15; the excerpt ends inside the session mechanics section, before install instructions. The README itself warns that all 85 tests use a fake Jev or stubbed transport and that live keep/drop quality has not been measured against any reference.
Questions about this entry
Short answers drawn from the linked project or page.
A selective context compaction extension for the Pi coding agent. Instead of asking an LLM to summarize old turns, it uses Jev to score every tool call and its result, drops what is no longer needed, and keeps the rest verbatim.
Pi's built-in compaction asks an LLM to summarize old turns, which can lose an exact error string, file path or command right when it becomes relevant again. This extension never rewrites or paraphrases: Jev says what is no longer needed, and everything that remains is byte-identical.
By default only tool calls and results are candidates, which sets a ceiling because assistant prose is often the largest part of a long session. The README measured 15.5% reclaimed on a synthetic 1,860-message window, rising to 33.5% when /jev-compact text added prose to the candidates.
No. User messages are never candidates in either mode. Prose is opt-in via /jev-compact text and held to a stricter threshold, because dropped reasoning cannot be recovered — unlike a dropped tool result, which can be produced again by re-running the tool.
Not yet. The README states all 85 tests use a fake Jev or a stubbed transport so no test observes real model judgment; live runs have compacted real sessions, but the quality of Jev's keep and drop decisions has not been measured against any reference.
Pi's session_before_compact hook accepts a single cut point (a summary plus firstKeptEntryId) and extensions receive a ReadonlySessionManager, so selective removal cannot be persisted in place. The extension calls ctx.newSession and replays the survivors; the previous session file is never modified.
No. The README states it is an independent project, not affiliated with TypeSafe AI or the Pi authors. The design credits prior art: fast-jev-compaction (MIT) for the two-noul question design and pi-vcc (MIT) for the verified Pi session_before_compact integration.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
X: The compaction plugin, in Chinese
X: Decision outputs in Grep
Hacker News: jevmem keeps Claude Code project memory by asking Jev which turns are worth saving.
avinashjetwani61 pointsJevmem – automatic project memory for Claude Code, built on JevRead the thread on Hacker NewsX: A usage router for Grok Bot
More Jev resources like this
A Pi extension that asks Jev, per completed tool call, whether the call and its full result still matter, then applies local keep, truncate or drop rules before compaction. Kept evidence is written verbatim into a Verbatim Tool Evidence block.
Jev Integrations & Adapters#community#catalog-import#logicrw
An OpenCode plugin that keeps the session on a cheap sticky parent model so the cache stays warm, and asks Jev whether a turn is hard before escalating through jev_escalate to temporary stronger subagents.
Jev Integrations & Adapters#community#catalog-import#cobanov
A context engine for OpenClaw that asks Jev which tool calls and results a session still needs, drops the rest and never summarizes: what stays is the original text, while one request scores dozens of calls in about a second.
Jev Integrations & Adapters#community#catalog-import#logicrw
Local tool-routing classifier for coding agents: a gateway and MCP tool that ask Jev which tool to call next, keep your agent's own model and login, and log every decision so you can start by observing.
Jev Integrations & Adapters#community#catalog-import#logicrw
Community
felpsdev/jev-classifierTypeScript
felpsdev/jev-classifierTypeScriptVerdict-based context compaction for DeepSeek Harness: instead of a lossy summary, jev-latest scores every tool call and result — kept verbatim, truncated to 300 characters or dropped — and falls back to compaction-basic on any failure.
Jev Integrations & Adapters#community#catalog-import#logicrw
A pi coding-agent extension that compacts long transcripts without rewriting them: Jev answers only keep/kind questions about chunked history, original lines go into a verbatim ledger, and compaction is a deterministic render plus durable topic files. Calls the TypeSafe System One endpoint directly.
Jev Integrations & Adapters#community#catalog-import#logicrw
From the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.