Skip to content

Context

The model sees only the question and the input. Background — whose decision it is, what counts as urgent here — comes from files you already keep. jevx reads them in this order and never edits them.

layer where who edits it
global the ## Jev section of ~/.claude/CLAUDE.md, ~/.codex/AGENTS.md, ~/.agents/AGENTS.md, ~/AGENTS.md; all read, merged, repeated lines dropped (so synced copies count once) you
folder the ## Jev section of the nearest AGENTS.md and/or CLAUDE.md from the working directory up; both merged, repeats dropped; --cwd DIR picks another start you
call --context "TEXT" or --context @file, repeatable the agent, per call
question question add … --context you, once

When CLAUDE_CONFIG_DIR is set, that directory’s CLAUDE.md is read too.

Terminal window
$ jevx context
global (none) add a "## Jev" section to one of: /Users/you/.claude/CLAUDE.md, /Users/you/.codex/AGENTS.md, /Users/you/.agents/AGENTS.md, /Users/you/AGENTS.md
folder (none) add a "## Jev" section to this repo's AGENTS.md or CLAUDE.md
order sent: global, folder, then --context on the call; a saved question's own context goes with that question

(Paths shortened to /Users/you.) With sections present, each line names the file it came from. --no-context skips the global and folder layers for one call; --context still applies.

AGENTS.md
## Jev
Decisions are for the payments platform team. Anything touching refunds or double charges is urgent.
On-call pages only for customer-visible failures, not for retries that succeed.
  • Any heading level works, with or without the space: #jev, ## Jev, ### Jev notes.
  • The section ends at the next heading of the same or a higher level. It may be empty.
  • The heading name and file names are configurable: local_context_section, local_context_file, global_context_file in the config file.
Terminal
jevx ask urgent --context "The board meeting starts in 20 minutes." < msg.txt
jevx is "Is this a critical bug?" --cwd ~/repos/payments < report.txt # that repo's ## Jev section
jevx is "Is this spam?" --no-context < msg.txt # skip the sections

A text input is sent with the context prepended; a JSON input keeps its shape and gets the context merged into its "context" field.

The skill’s own example, from the maintainer: the same message scored 0.84 with no context, 0.94 with “the meeting is in 20 minutes” and 0.64 (unsure) with “the meeting is in three months”. That one was not re-run for these docs.

Context is the part of a judgement that is yours: what counts as urgent, whose decision it is. Keeping it in the file your agent already reads means one place to edit, a diff you review, and no hidden state the CLI could drift. If a verdict looks wrong, the skill tells the agent to name the file to edit rather than edit it.

Something wrong or unclear on this page?Open an issue on GitHub.