Using it from an agent

The skill (SKILL.md) is what the agent reads. It says when to reach for jevx, how to word a question so it stands alone, what each exit code means, and what never to do with it.

Where it goes

jevx install --skills copies SKILL.md to:

• ~/.claude/skills/jevx/ (or $CLAUDE_CONFIG_DIR/skills/jevx/ when that variable is set)
• ~/.agents/skills/jevx/, the vendor-neutral location
• ~/.codex/skills/jevx/, if ~/.codex exists

Claude Code and Codex load a skill when its description matches the task. jevx skill prints the installed copy; the published one is at skill.md.

What the agent is told

• When to reach for it: classifying, filtering, triaging, routing, deduping, ranking or checking many items; when a branch depends on meaning; before asking the user “A or B?”; before handing back a turn.
• The forms: is / pick / filter / rank, then ask with saved or inline questions over one input or a batch.
• Exit codes: 0 act, 1 other branch, 3 check or ask, 4 say it failed. An error is never a no.
• Context: read ## Jev from the user’s agent files; pass facts with --context; tell the user which file to edit, never edit it.
• Saved questions: check question list first; save per repo with --local when asked.
• Reporting: word the question so it stands alone, and report when a verdict changed what it did: “jevx flagged 12 of 300 lines”.
• Never: use it as a safety gate, put secrets into --in or --proposal, or train another model on a hosted endpoint’s answers unless its terms allow it.

The table the agent uses

you are about to…	do this instead
read 300 lines to find the failures	ask --lines log --noul fail="Is this a failure an on-call engineer would act on?"
sort tickets into buckets	ask --states t.jsonl --choice team="Which team?|web=UI;api=backend;ops=infra"
decide if two records are one entity	printf 'A: %s\nB: %s' "$a" "$b" | jevx ask --noul same="Are A and B the same customer?"
check a claim against a source	jevx ask --noul ok="Does the text say payment is due in October?" < invoice.txt
rank search results	ask --lines results.txt --noul hit="Does this answer: how do I get a refund?", then sort by p
ask the user “A or B?”	ask --in "<the facts>" --choice pick="Which would the user choose?|a=<A>;b=<B>"; ask only on exit 3
hand back your turn	jevx judge --request "<request>" --proposal "<final message>"

A session, end to end

The user asks an agent to find what broke in last night’s log. Instead of reading it, the agent runs one batch:

Terminal

jevx ask --lines app.log --noul fail="Is this a failure an on-call engineer would act on?" \
  | jq -c 'select(.answers.fail.verdict=="yes")'

On the three-line sample from Ask, the same kind of batch picks line 2 (the gateway timeout, P 0.98) and nothing else. The agent reads that line, fixes the cause, and says in its report that jevx picked 1 of 3 lines and it acted on that one. If the call had exited 4, the skill says to report the failure, not to guess.

Other agents

Anything that reads ~/.agents/skills gets the same file. Plugins (hooks) are Claude Code only for now. jevx uninstall --skills removes exactly the copies it wrote and leaves other skills untouched.