New: 98,000+ skills indexed from GitHub repos with 100+ stars · every skill safety-checkedBrowse top skillsTop skills
Esc to closeBrowse all skills
Submit a skillBrowse skills
Submit a skillBrowse skills

Agent skill

Jev Compactor

by autonomous-ai in autonomous-ai/openharness

Design and tune a Jev Compactor session. An agent skill from autonomous-ai/openharness.

MITAuto-check passed

Install Jev Compactor

skills CLI
$ npx skills add autonomous-ai/openharness --skill jev-compactor -a claude-code

Project install by default; add -g for ~/.claude/skills/.

GitHub CLI
$ gh skill install autonomous-ai/openharness jev-compactor --agent claude-code

Project scope by default; add --scope user for a personal install. Needs GitHub CLI 2.90.0 or later (public preview).

Manual copy
$ git clone --depth 1 https://github.com/autonomous-ai/openharness.git skills-src && mkdir -p .claude/skills && cp -r skills-src/store/agents/jev-compactor/skills/compactor .claude/skills/jev-compactor && rm -rf skills-src

Use ~/.claude/skills/ instead of .claude/skills for a personal install. The folder must contain SKILL.md.

Claude Code skills documentation · loads skills from .claude/skills/

Facts

Skill name
jev-compactor
GitHub stars
1.1k
Token cost
~1.5k tokens
SKILL.md length
871 words
Files
1
Skills in repo
100
Repo updated
First seen
Licence
MIT

At a glance

Design and tune a Jev Compactor session. An agent skill from autonomous-ai/openharness.

  • Works in 5 steps: Write the tasks. Run node… → Start easy (distraction 0.1). Let a few… → Raise distraction in steps of 0.2. Note… → …
  • Editing session.json
  • SKILL.md covers The decision, Ground truth, The knobs and What to expect (mock, default…, plus 4 more sections
  • Calls node

What it does

Jev Compactor is an agent skill from autonomous-ai/openharness. Design and tune a Jev Compactor session. Use when editing session.json, writing task vocabularies, or looking for settings where needle recall holds while the reduction is large.

Its SKILL.md is about 1.5k tokens, which your agent loads only when the skill is triggered. It is a single SKILL.md file with no bundled scripts.

The repository describes itself as: The ultimate harness for coding agents and beyond. All your agents. All your machines. One command center. Start with code, then follow your curiosity and build across… The licence is MIT.

When your agent uses it

  • Editing session.json
  • Writing task vocabularies
  • Looking for settings where needle recall holds while the reduction is large

Example prompts

  • “/jev-compactor”

Requirements

  • Node.js

Workflow steps

5 steps, taken from the first numbered list in SKILL.md.

  1. Write the tasks. Run node "$JEV_DSH/toolchain/check.mjs".
  2. Start easy (distraction 0.1). Let a few compactions run. Read .harness/verdict.json
  3. Raise distraction in steps of 0.2. Note where recall crosses recallTarget.
  4. At that edge, try one fix at a time: a higher target, a bigger budget, a bigger trimTo,
  5. Try one bad idea on purpose: give two tasks the same words. Watch Jev keep the wrong task's

What it can do on your machine

Read from SKILL.md and the folder at commit 50da5db. It shows what the files ask for, not the result of running them.

  • Tool permissions

    Pre-approves nothing: there is no allowed-tools line, so your agent's usual permission prompts apply.

    From allowed-tools in the SKILL.md frontmatter.

  • Runs code

    Shell commands in SKILL.md call:

    • node

    From the folder's file list and the shell code blocks in SKILL.md.

  • Network

    No URLs in SKILL.md.

    From URLs in SKILL.md, links to its own repository left out.

  • Credentials

    Names no API keys, tokens, secrets or passwords.

    From names ending in _API_KEY, _TOKEN, _SECRET, _KEY or _PASSWORD in SKILL.md.

Context cost

Jev Compactor loads about 1.5k tokens when it runs. Until then it costs about 48 tokens; SKILL.md has 871 words of instructions outside code blocks.

Always · name and description, kept in context so the agent knows when to use it
~48
When it runs · the whole SKILL.md, loaded when a task matches
~1.5k

Estimates: characters ÷ 4, the usual rule of thumb; real counts depend on the model's tokenizer. Scripts and assets cost tokens only if the agent reads them.

Safety

Auto-check passed

The automated check found no risky patterns in SKILL.md.

Automated static check — not a guarantee. Review scripts before installing. It scans the text of SKILL.md for risky patterns (piping downloads into a shell, reading credential files, hidden Unicode, destructive commands); files beside SKILL.md are not scanned.

SKILL.md

The full file from autonomous-ai/openharness at commit 50da5db, republished under its MIT licence (© autonomous-ai). 871 words, ~1,517 tokens.

Download SKILL.mdSave it as .claude/skills/jev-compactor/SKILL.md (or your agent's skills folder).
name
jev-compactor
description
Design and tune a Jev Compactor session. Use when editing session.json, writing task vocabularies, or looking for settings where needle recall holds while the reduction is large.

Craft: Jev Compactor

Jev Compactor runs a made-up coding-agent session and lets Jev compact its context window. Your craft is the session design: tasks Jev can tell apart, and settings with a clear, honest finding. Everything is synthetic. It is a demo of the idea, not a real compaction plugin.

The decision

When the window passes budget, the viewer sends Jev:

  • state: the current task (title and keywords) and the last few messages.
  • one choice question per tool result, 100 per call:
[Read] src/refund/ledger_cents.ts · 4,210 tokens. Preview: «export function apply_refund(...) { ... }»
Is this tool result still needed for the current task?
  keep  still needed, keep it word for word
  trim  only the gist matters now, keep the head
  drop  irrelevant now

Questions cannot see each other. Jev judges each block from its own text and the shared state. Messages are never judged.

Ground truth

Every tool result secretly belongs to one task, or is junk. For the current task:

  • Read and Edit results are detail needles. Every token counts. Ideal verdict: keep.
  • Grep, Bash and WebFetch results are gist needles. About the first 300 tokens count. Ideal verdict: trim.
  • Everything else (other tasks, junk) is not needed. Ideal verdict: drop.

Needle recall = needed tokens that survived / needed tokens before. Junk removed = the same for tokens that were not needed. About 4% of real results have a preview that only shows boilerplate, so even an easy session is not perfect. That is a real limit, not noise.

The knobs

knobwhat it doestry
tasks[].vocabularythe words Jev can tell tasks apart by8 to 12 single words, no word in two tasks
distractionhow much junk borrows the current task's words0.1 easy, 0.5 hard, 0.9 very hard
targetthe window must end under target x budget0.4 default, 0.3 squeezes hard, 0.6 is gentle
trimTotokens a trimmed block keeps400 default, under 300 starts to lose gist needles
budgetwhen compaction fires200000 default, 1000000 for a long session
noise, focus, mixhow much junk, how much on-task work, which eventsmore junk means bigger cuts
taskEveryevents before the made-up user moves on0 keeps one task forever, which fills the window with needles

What to expect (mock, default tasks, 1500 events)

distractionneedle recalljunk removedreduction
0.12about 98%about 97%about 81%
0.3about 96%about 86%about 72%
0.5about 90%about 82%about 70%
0.9about 78%about 78%about 70%

The "summarize instead" baseline keeps about 49% of needle tokens on the same session.

Why recall falls: junk that talks like the task gets kept. The window stays above the target, so the pressure pass trims the keeps Jev was least sure about. Some of those are real needles.

A good tuning loop

  1. Write the tasks. Run node "$JEV_DSH/toolchain/check.mjs".
  2. Start easy (distraction 0.1). Let a few compactions run. Read .harness/verdict.json: the summary has recall and reduction, the findings have cost, calls and the baseline.
  3. Raise distraction in steps of 0.2. Note where recall crosses recallTarget.
  4. At that edge, try one fix at a time: a higher target, a bigger budget, a bigger trimTo, sharper vocabularies. Say which one helped and what it cost in reduction.
  5. Try one bad idea on purpose: give two tasks the same words. Watch Jev keep the wrong task's blocks. Report it.
Show full SKILL.md (350 more words)Show less

The person's own transcript

Set "source": "my-session.jsonl" in session.json and the pane analyses a real transcript from the workspace instead of the made-up session.

  • Find it: Claude Code writes one .jsonl per session under ~/.claude/projects/<project folder>/ (the project path with / turned into -). ls -t ~/.claude/projects/*/*.jsonl | head shows the newest. Ask the person which one.
  • Copy it into the workspace: cp "<file>" ./my-session.jsonl. Inside the workspace, not under .harness, at most 64 MB. Never edit the original.
  • Never paste its contents into chat. Do not read it. Read compaction-plan.json and .harness/verdict.json instead: they hold tool names, input summaries and numbers only.
  • With a Jev key, the task message, the last few messages and the head (about 300 characters) of each tool result go to the Jev API. With no key on the machine (environment or credentials file) the offline mock judges and nothing leaves it. Tell the person which one ran.
  • There is no ground truth: no recall, no truth strip, no baseline. Report tokens before and after, the reduction, keep / trim / drop counts, questions, calls, time, cost and the biggest drops.
  • The task is the last user message. If the cut looks wrong, suggest another task button, or a pin, then "Compact now". trimTo is the only tuning knob that matters here.
  • It is an analysis of what Jev would cut, not a plugin. No live session changes.

A good report: "my-session.jsonl, 1,144,012 -> 67,087 estimated tokens (94.1% cut). 167 tool results: keep 13, trim 19, drop 135. 2 calls, mock. Biggest cuts: package-lock.json reads and npm install logs."

Reporting

Give the settings and the measured numbers together, for example: "distraction 0.3, target 0.4, trimTo 400: recall 96%, reduction 76% over 12 compactions, mock." Always say it is a synthetic session, and whether the mock or live Jev ran. Never describe the "summarize instead" lane as a real product. It is a simple baseline for contrast.

Definition of done

  • session.json passes check.mjs.
  • Your own tasks and repo name, not the template's.
  • One settings line that holds the recall target with a large reduction, and one that breaks it, both with numbers.

© autonomous-ai, MIT. Rendered from Markdown: HTML in the file is shown as text, images as links, and headings moved down two levels. Raw file

Files

Just SKILL.md in store/agents/jev-compactor/skills/compactor of autonomous-ai/openharness.

Open the folder on GitHubat commit 50da5db

Compare with similar skills

Jev Compactor next to the 5 skills that share the most tags, products or categories with it. Stars are the repository's; “used in” counts other GitHub owners with a copy.

Jev Compactor compared with similar skills
SkillStarsUsed inTokensAuto-checkLicenceRepo updated
Jev Compactor this skillautonomous-ai/openharness1.1k—~1.5kAutomated safety check: PassMIT
LLM Fine Tuningsickn33/agentic-awesome-skills47k1 repos~2.3kAutomated safety check: PassMIT
Fine Tuning With TrlOrchestra-Research/AI-Research-SKILLs13k7 repos~2.9kAutomated safety check: PassMIT
Agent Platform Tuninggoogle/skills21k—~9.6kAutomated safety check: PassApache-2.0
Jev Socialsickn33/agentic-awesome-skills47k1 repos~3.4kAutomated safety check: PassMIT
Agent Platform Tuning Managementgoogle/skills21k—~1.9kAutomated safety check: PassApache-2.0

Similar skills

  • LLM Fine Tuning

    sickn33/agentic-awesome-skills

    Set up infrastructure for fine-tuning LLMs with QLoRA, LoRA, and full fine-tuning using Hugging Face TRL, Axolotl, and distributed training with DeepSpeed or FSDP.

    47k GitHub starsUsed in 1 repo~2.3k tokens
    AI & LLM EngineeringAuto-check passed

  • Fine Tuning With Trl

    Orchestra-Research/AI-Research-SKILLs

    Fine-tune LLMs using reinforcement learning with TRL - SFT for instruction tuning, DPO for preference alignment, PPO/GRPO for reward optimization, and reward model training.

    13k GitHub starsUsed in 7 repos~2.9k tokens
    AI & LLM EngineeringAuto-check passed

  • Agent Platform Tuning

    google/skills

    Official

    Agent Platform Model Tuning. An agent skill from google/skills.

    21k GitHub stars~9.6k tokensUpdated today
    AI & LLM EngineeringAuto-check passed

  • Jev Social

    sickn33/agentic-awesome-skills

    Run read-only, browser-grounded Instagram, TikTok, or LinkedIn research through Jev routing and socai CLI, returning source-linked evidence and reports.

    47k GitHub starsUsed in 1 repo~3.4k tokens
    Auto-check passed

  • Agent Platform Tuning Management

    google/skills

    Official

    Manages GenAI tuning jobs in Agent Platform. An agent skill from google/skills.

    21k GitHub stars~1.9k tokensUpdated today
    AI & LLM EngineeringAuto-check passed

  • Peft Fine Tuning

    Orchestra-Research/AI-Research-SKILLs

    Parameter-efficient fine-tuning for LLMs using LoRA, QLoRA, and 25+ methods.

    13k GitHub starsUsed in 9 repos~3.1k tokens
    AI & LLM EngineeringAuto-check passed

More from autonomous-ai/openharness

All 100 skills in this repo

  • G-code Slicer Tool

    autonomous-ai/openharness

    Slices 3D mesh files into printer-profiled plain G-code through real slicer CLIs, with backend discovery, input inspection, dry runs and static validation.

    1.1k GitHub stars~1.5k tokensUpdated today
    Auto-check passed

  • Home Assistant Automation Builder

    autonomous-ai/openharness

    Turns a home-automation request into standard, testable automations.yaml, run against Home Assistant Core's real triggers and verified with its own trace tool.

    1.1k GitHub stars~1.2k tokensUpdated today
    Auto-check passed

  • Score Music Composer

    autonomous-ai/openharness

    Turns a musical brief into LilyPond concert-pitch music, checked parts for each instrument and a playable practice pack.

    1.1k GitHub stars~1.1k tokensUpdated today
    Auto-check passed

  • OrcaSlicer 3MF and G-code Workflow

    autonomous-ai/openharness

    Turns an STL and explicit printer and material requirements into compared OrcaSlicer plans, an editable 3MF project, checked G-code and a portable handoff.

    1.1k GitHub stars~1.5k tokensUpdated today
    Auto-check passed

  • Sheets and Docs Report Builder

    autonomous-ai/openharness

    Builds an editable DOCX report, a formula-driven XLSX workbook and a fresh LibreOffice PDF preview from one structured source file, then checks them together.

    1.1k GitHub stars~708 tokensUpdated today
    Auto-check passed

  • Bambu Labs

    autonomous-ai/openharness

    Dry-run, upload, and cautiously initiate local Bambu Lab print jobs from validated plain .gcode, using Bambu LAN FTPS/MQTT handoffs.

    1.1k GitHub starsUsed in 1 repo~2.1k tokens
    Auto-check: warnings

Questions about Jev Compactor

What does Jev Compactor do?

Design and tune a Jev Compactor session. An agent skill from autonomous-ai/openharness. Jev Compactor is an agent skill from autonomous-ai/openharness. Design and tune a Jev Compactor session.

When should I use Jev Compactor?

Jev Compactor fits situations like: editing session.json; writing task vocabularies; looking for settings where needle recall holds while the reduction is large.

How do I install Jev Compactor in Claude Code?

Run `npx skills add autonomous-ai/openharness --skill jev-compactor -a claude-code`. Or copy the skill folder (store/agents/jev-compactor/skills/compactor in autonomous-ai/openharness) into .claude/skills/jev-compactor in your project. Claude Code loads it when a task matches its description.

How do I install Jev Compactor in Codex?

Run `npx skills add autonomous-ai/openharness --skill jev-compactor -a codex`. Or copy the skill folder (store/agents/jev-compactor/skills/compactor in autonomous-ai/openharness) into .agents/skills/jev-compactor in your project. Codex loads it when a task matches its description.

Can I use Jev Compactor in Cursor, Gemini CLI or GitHub Copilot?

Cursor, Gemini CLI, GitHub Copilot and OpenCode also load SKILL.md folders. With the skills CLI, run `npx skills add autonomous-ai/openharness --skill jev-compactor -a cursor` (or -a gemini-cli, github-copilot or opencode for the others). To copy it by hand, put the folder in .cursor/skills/jev-compactor, .gemini/skills/jev-compactor, .github/skills/jev-compactor and .opencode/skills/jev-compactor in your project.

What does Jev Compactor need to run?

Going by SKILL.md and its folder, Jev Compactor needs the command-line tools its instructions call (node). Our summary lists: Node.js.

Does Jev Compactor access the network?

SKILL.md contains no URLs. Any network use would come from the scripts or tools the agent runs. This is read from the text; nothing was executed.

Is Jev Compactor safe to install?

Our automated static check of SKILL.md found no risky patterns, such as piping downloads into a shell, reading credential files or hidden Unicode. It is not a guarantee. Review the folder before installing.

What licence does Jev Compactor use?

Jev Compactor is published under the MIT licence (the repository's licence). It allows redistribution, so the full SKILL.md is shown on this page.

How many tokens does Jev Compactor use?

About 1.5k tokens (SKILL.md is roughly 6.1k characters). Agents keep only the skill's name and description in context until a task matches; then they load SKILL.md in full.

What are the alternatives to Jev Compactor?

Skills that share tags, products or a category with Jev Compactor: LLM Fine Tuning (sickn33/agentic-awesome-skills, 47k stars), Fine Tuning With Trl (Orchestra-Research/AI-Research-SKILLs, 13k stars), Agent Platform Tuning (google/skills, 21k stars) and Jev Social (sickn33/agentic-awesome-skills, 47k stars). The comparison table on this page puts their stars, adoption, token cost, safety result and licence side by side.

Who maintains Jev Compactor?

autonomous-ai (a GitHub organization) maintains it in autonomous-ai/openharness, which has 1,149 GitHub stars. The repository holds 100 skills in this directory. The repository was last updated on October 8, 2026.

Source: autonomous-ai/openharness on GitHub. Facts on this page come from the repository at the commit we read; the author's words are quoted as theirs.