jevskill
Agent skill that has Jev pick the few relevant lines from big logs before your coding agent reads them, with a reversible reduce and ledger.
Overview
jevskill is a skill for Claude Code, Codex and many other agents that stops them burning context on large tool output such as CI logs. Instead of pasting a 900-line log into the session, where it is re-sent every turn, the skill asks Jev to find the few lines that matter; the README shows a 34,989-token log reduced to a 320-token shortlist. The reduction is reversible, and a ledger learns when calling Jev pays off.
The README reports an A/B test with 99.3% fewer input tokens and higher accuracy (18 of 18 versus 15 of 18), and quotes about 325 ms and $0.000013 per decision. It works via OpenRouter or TypeSafe directly and installs with npx skills add; an agent can also install it by following a hosted install guide. The README also promotes a live stream (in Polish) about building computer use with Jev. Figures are the author's own tests.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Yes / no | Is this log line relevant to the current task or failure?(inferred) | — |
| Yes / no | Is it worth calling Jev for this output? (ledger)(inferred) | — |
How it works
- Agent receives large tool output
- Skill sends candidate lines and the task to Jev
- Jev returns a shortlist of relevant lines
- Agent reads the shortlist; full output remains recoverable
- Ledger records outcomes to learn when reduction helps
Strengths
- Large reductions in context growth across a session
- Reversible reduction
- Very low cost per decision per the README
Caveats
- Benchmarks are the author's own A/B tests
- Requires an OpenRouter or TypeSafe key
- Relevance filtering can drop needed lines
Getting started
npx skills add lazniak/jevskill -gAt a glance
- Type
- Project
- Status
- beta
- Language
- Python
- Runtime
- agent skill
- Has code
- yes
- License
- MIT
Use cases and integrations
How curated lists describe it
Teach your coding agent to stop burning context. Jev (System One) via OpenRouter or TypeSafe: 325ms, 0.000013 USD per decision. A/B tested 99.3% fewer input tokens with accuracy up. Ships a reversible reduce and a ledger that learns when Jev pays off.
logicrw/awesome-jev-projectsTeach your coding agent to stop burning context. Jev (System One) via OpenRouter or TypeSafe: 325ms, 0.000013 USD per decision. A/B tested 99.3% fewer input tokens with accuracy up. Ships a reversible reduce and a ledger that learns when Jev pays off.
hellogumbo/awesome-jev
Frequently asked questions
What does Jev decide?
Which lines of a large log or tool output are relevant to the task.
Can I get the full output back?
Yes, the README describes the reduce as reversible.
Which agents does it support?
Claude Code, Codex, DSH and 20+ others via the skills installer.
Related projects
Scores installed Claude Code skills against the current project with Jev and hides irrelevant ones via skillOverrides.
Agent skillbetaJavaScriptDeepSeek Harness plugins that filter tool output, prune stale context and disclose skills progressively, with Jev or Laya judging what goes.
PluginbetaTypeScriptLocal compaction proxy for Claude Code and Codex that compacts more often with local rules and keeps full history searchable.
ProxybetaGoSkill that makes Claude Code and Codex split AI features into Jev, LLM or plain-code steps before writing code.
ProjectbetaClaude Code (and Codex) plugin that trims long Bash output chunk by chunk with Jev before it reaches the model, without summarising.
PluginbetaTypeScriptPre-message hook that keeps or drops SKILL.md files before the first model call, using local overlap or Jev keep/drop judgments.
HookbetaPython