User repository on GitHub
"Work appears completed or recorded. Run /compact to save tokens."
- Forks
- 27
- Last pushed
- Updated Oct 2026
Repo stats from the GitHub API, cached Oct 2026.
CommunityJev Agents & MCP
Compact Adviser
An agent plugin for Pi, Claude Code, Codex CLI and Grok Build that asks Jev two classification questions to decide whether a session sits at a safe compaction boundary, then hints or compacts for you.
- Stars
- 210
- Language
- TypeScript
- License
- MIT
- Published by
- Community
- Author
- kunchenguid
- Last verified
- Oct 2026
About this project
compact-adviser is an agent plugin that answers a single question: should I /compact now? It uses Jev to instantly judge whether the current session is likely at a boundary that's safe to compact. It can give you a hint to run /compact - or, on Pi and Claude Code, if you opt in, it can run it for you at the right time automatically. Codex CLI and Grok are hint-only: nothing outside their sessions can trigger /compact.
Quoted from the project's README. Read the full README.
- Languages
- TypeScript
Highlights
- Two one-sentence Jev questions per checkpoint, is the unit finished and is this hands-on work, composed in code into a single score.
- The threshold starts near 0.90 on an almost empty context and relaxes toward 0.50 as the host's own compaction point approaches.
- Automatic compaction exists on Pi and Claude Code only; Codex and Grok are hint-only because nothing outside their sessions can trigger the command.
- Requests are capped at 32,000 serialized bytes and an error is never treated as an affirmative judgment, so a failure stays silent rather than lying.
Watch out
Installing is stated as consent to send eligible checkpoint context to TypeSafe once a key is available, and redaction of known key patterns is described as best-effort with no guarantee, which is not enough for material that must not leave the machine. Codex support is macOS and Linux only with hooks running through POSIX sh, hints are suppressed below a constant 40,000-token minimum, and the adviser stays inert in non-interactive sessions. Grok needs manual hook and status-line setup because it never loads a plugin's own hooks file. 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.
The documented payload is bounded user constraints, up to the last 64 visible replies and tool results clipped, short tool-result excerpts, any existing summary, saved-artifact names and omission markers. System prompts, hidden reasoning, images, environment variables and complete transcripts are not sent, and the key never enters the model context or the request body. Because redaction is best-effort with no guarantee, the README's own advice is to uninstall or set mode Off for material that must not leave the machine.
The gate is deliberately strict early. The floor sits at 0.90 while the context is mostly empty and relaxes toward 0.50 only as the host's own compaction point nears, and no hint appears at all below a 40,000-token minimum, while the session is not idle, or when the last turn was not a settled final answer. Automatic mode applies that same gate plus a first-use confirmation, and you can raise or lower the trigger with an explicit context budget.
On Codex the hook must be reviewed once in the hooks view, and because Codex rebuilds PATH, an older or unfindable Node silently disables it, which is what the absolute node-path variable exists for; Codex also needs Node 22.18 or newer. On Grok the install command must have run to write the Stop hook, and the status row only appears if you paste the printed status-line block into the named config file and are not in minimal render mode.
No, and that is the design. The plugin asks two one-sentence classification questions in a single request and combines them into one score in code; the README is explicit that it never asks the model to write a summary, so no generated text enters the judgment. The API base can be redirected for a self-hosted endpoint, and it must be https, with plain http accepted only for loopback hosts.
Pi 0.82.0+, Claude Code 2.1.274+, Codex CLI 0.153.0+ (macOS and Linux only) and Grok Build 1.0.34+, each requiring Node 22+ and a TypeSafe API key.
It starts at 0.90 while the context is mostly empty, through about 10%, and relaxes toward 0.50 by about 90% full, because a wrong hint costs most when there is still room.
Context below a constant 40,000-token minimum, a session that is not idle, or a last turn that was not a settled final answer all produce no hint.
More Jev resources like this
1.5kGitHub stars
Open-source platform that puts ACP-compatible agents such as Claude Code, Codex, and Pi inside the team's Slack, Telegram, Discord, and GitHub workflows, with reusable Jev decisions for routing, response gating, and reviewer selection.
Jev Agents & MCP#community#jevbest
488GitHub stars
A coding agent built on pi that hands its non-code decisions — what enters the context, whether a command is safe, whether the work is done — to a small judge model at 35 decision points per turn.
Jev Agents & MCP#community#jevbest
Community
qybaihe/muTypeScript
qybaihe/muTypeScript572GitHub stars
A CLI plus agent hooks that ask Jev one question per rule in your AGENTS.md on every edit or turn and return a probability, so a violating edit gets repaired inside the same turn. Claude Code, Codex, OpenCode, Pi.
Jev Agents & MCP#community#jevbest
Community
coldteadotai/abideTypeScript
coldteadotai/abideTypeScript147GitHub stars
A local-first desktop agent for macOS that carries work from a conversation through to a reviewable result, split into Home, Cowork and Crew. The part worth a developer's attention is Crew, where Jev suggests an assignee for a single unassigned task and you accept or ignore it.
Jev Agents & MCP#community#jevbest
Community
Foxtailsss-Andy/Anna-AgentTypeScript
Foxtailsss-Andy/Anna-AgentTypeScript116GitHub stars
Stanley is a CLI where you describe a code-review request in plain language and Jev routes it to one of ten bounded workflows. Each workflow is deterministic code that gathers bounded evidence and asks Jev fixed-choice questions per hunk; unmatched requests go to the Pi agent.
Jev Agents & MCP#community#jevbest
Community
devagrawal09/jev-codeTypeScript
devagrawal09/jev-codeTypeScript48GitHub stars
Agent harness for work that outlives one chat, with resumable turns, routed state machines, and observational memory across sessions and processes.
Jev Agents & MCP#community#jevbest#catalog-import
Community
dzhng/duet-agentTypeScript
dzhng/duet-agentTypeScriptFrom the Jev community
Posts from builders shipping with Jev right now.