carbon-background

Files (1)

• SKILL.md

SKILL.md (first lines)

# Jay background checks

Run the helper with the host's authorized terminal tool:
`node <plugin-root>/background/engine.mjs status <project-root>`.
Resolve the real project root first; quote paths. Read `../../background/README.md`
for the limits, host differences, exact commands and privacy disclosure.

## Default quick checks and opt-out

Lightweight local checks are on by default where the host trusts the command hooks.
Preserve existing saved disables and global pauses. Briefly explain on first use:
Jay checks changed source after roughly two quiet minutes, without running the app
or making model requests. Three minutes per run, at most three runs and ten reserved
minutes per UTC day, shared across projects. Offer “Pause background checks”; do not
repeatedly ask someone who already paused. AI review and executable project tests
still require separate approval. Host-hook approval is never bypassed.

`configure <root> '{"enabled":true}' --approved` records explicit approval.
`pause <root>` records a decline or pauses this project. `off-everywhere <root>`
stops all projects, including work already running. `allow-projects <root>` clears
only that global pause; projects with no saved setting then use the on-by-default
local checks, while explicitly disabled projects stay disabled. Explain this and ask before re-enabling.
`review <root>` returns saved evidence and records that the user reviewed this run.

## Daily chat reminder

With trusted host hooks, after 24 hours without a CARBON run being requested in a
chat, Jay briefly offers a run on that chat's next user turn. This is a question,
not permission to run. Each chat is independent; at most one nudge per 24 hours.
Do not repeat the question when the user is already requesting testing. When starting
an assessment, use the exact `thread-invoked` command supplied by the current host
hook to record it, including assessments selected without a slash command. Never
guess a thread key. No host thread ID means no automatic reminder.
`configure <root> '{"reminders":false}'` turns off nudges for this project;
`off-everywhere` also silences them. Never create a schedule or wake an idle model
just to ask. Host sessions that lack hooks retain ordinary manual commands.

Recommend one installed command based on the actual conversation and project, with a
brief reason: changed feature → focused test; keyboard/UI accessibility work →
accessibility; auth/trust boundaries → security; release decision → confidence.
Respect the edition: when a specialized command is absent, use `carbon-test` with
that focus instead of advertising unavailable functionality. Unclear context →
`carbon`. The local hint stores only a topic category, never the prompt text; use the
live chat to refine it. Do not fabricate activity or claim the suggestion is a result.

Checks stay on this machine by default. Claude Code users can separately opt into
`{"aiReview":true}`: it sends a bounded source excerpt to Claude through the signed-in
CLI, consumes that provider's allowance, and is NOT covered by a promise of free usage.
Do not enable AI review on Codex or on machines without the Claude CLI. Never switch
their provider or region silently. Ask before enabling this separate data flow.

For approved local regression tests, configure `tests` only after showing the user
the exact executable, arguments, project and side effects. These commands are not
sandboxed by CARBON: approve only isolated local tests without live services or
production credentials. Do not infer test safety from a package.json script name.

## Results

Show one brief Jay summary: checks passed/failed/deferred, possible issues by category,
strongest actionable finding and the next behavior check. Use the bundled
`background/jay-gray.webp` icon where local images are supported; otherwise say Jay.
Use monochrome bars with counts, not a made-up quality percentage or green styling.
Do not use another product's name or icon. Offer the `latest.html` report, review evidence,
ask the coding agent to fix a selected issue, or pause. Do not auto-fix, manufacture
confidence percentages, or describe source inspection as browser testing. No token
counter. A timeout, provider failure or cancellation is partial/blocked, not a pass.

Hosts must support and trust the bundled command hooks. In unsupported hosts, explain
that idle automation is unavailable and offer a normal bounded quick assessment with
available host tools. Never install a daemon, create a schedule or bypass hook trust.

<!-- usage-command:start -->
### Anonymous usage counter
Once when this command is actually invoked (not when merely listing or reading
documentation), run `node <absolute-plugin-root>/scripts/usage-analytics.mjs command carbon-background`
with the host terminal if available. Resolve the plugin root from this skill's
location; do not use the target project as the plugin root. Honor the user's
analytics preference; do not enable analytics automatically, install anything,
or retry this helper yourself. Failure or missing terminal access must never
block the command. This records a command invocation, not a completed test.
<!-- usage-command:end -->

## End-of-command issue handoff
At the end of this command, list actual findings from this invocation or the explicitly
selected assessment in chat, with severity and demonstrated/suspected status. Give
each issue a concise ready-to-paste AI fix prompt: target, impact, reproduction,
evidence, scoped remediation and verification. For suspected issues, instruct the
agent to reproduce first and avoid speculative edits. Ask: “Would you like me to fix
these issues—all of them, or selected ones?” Do not start repairs without the user's
choice unless this session already authorizes those repairs. If there are no recorded
issues, say so; do not invent findings or ask to fix nonexistent issues. Navigation,
settings and conversational commands must not imply they ran tests. Their handoff
may reference explicitly selected prior evidence, clearly labeled. Copying a prompt
does not authorize execution.

From jarbon/carbon-lite, MIT.

More from jarbon/carbon-lite