laight plugins for Claude Code
Two independent Claude Code plugins. Install either one, or both.
| Plugin | In one line |
|---|
| cross | Two AI reviewers with different models and lenses review your PR blind; a third pass, with your session's context, writes one GitHub review that says where each comment came from. |
| baton | Every Claude Code session leaves a short handoff at commit time, and the next one checks it against git before trusting it. No more cold starts. |
/plugin marketplace add berni23/claude-plugins
/plugin install cross@laight
/plugin install baton@laight
Both need gh signed in for their GitHub parts, and both read your
project's conventions (check command, tracker, hard rules) from its CLAUDE.md / AGENTS.md
instead of assuming any.
cross — two blind AI reviewers, one final PR review
A single AI reviewer has two problems. It misses things, and it reports things that aren't true:
a "race condition" that can't happen, a "missing null check" two lines above. You end up reviewing
the review.
[!IMPORTANT]
/cross:cr and /cross:fix-cr are not a replacement for human review. They catch and filter
more than a single AI reviewer does, but every finding is checked by reading, not proven, and
every fix is the AI's judgement. Use them to prepare a PR for a human reviewer, and have a person
look at the review and the fixes before merging anything that matters.
cross cuts down both:
- Two reviewers, reviewing blind. One Opus reviewer leads with runtime, state and failure modes.
One Sonnet reviewer leads with spec, contract and test coverage. Neither sees the other's output.
- A third pass writes the final review. The session you run it from reads both reviews, checks
each finding against the code, drops false positives and noise (recording why), and adds what it
sees itself, including from the ticket or requirements it already has in context. It's a
reviewer, not a referee. Nobody runs the PR's code; everything is checked by reading.
- One review, written as one reviewer's. It posts as a normal GitHub review. Each comment
carries a severity,
[high], [medium] or [low], that the final pass sets itself.
Whether both reviewers found a finding, only one did, or the final pass added it is evidence for
that severity and for whether to post the finding at all, not a rule: a pedantic point both raised
can be dropped, and a crash only one caught can be [high]. Who raised what (A+B, A, B, or
S for the final pass's own, often from the ticket in your session) stays in the local
scratch.review; the PR never mentions agents, models or tiers. The final pass never clears a
real defect with context, and never goes looking for context it doesn't have.
Everything either skill posts goes out under your GitHub account, so it is marked as an agent's:
every review comment starts with [cr-agent] and every reply or filed issue with [fix-cr-agent].
[cr-agent] [high] retry() re-enqueues the job before the lock is released, so a fast worker picks it up
while the first attempt is still writing. Two writes to the same row. Confirmed by reading
worker.ts:88–104: the finally that releases the lock runs after enqueue().
Example of a posted comment.
Then /cross:fix-cr answers the review, and it works on any review, not just one from
/cross:cr, including your colleagues'. Two independent planners decide for each finding whether
to fix it, reject it (with the reason) or turn it into an issue. The session that wrote the code
compares their plans, shows you a decision table, then fixes the real findings one at a time,
starting each with a test that fails, and replies on every thread.
open PR → /cross:cr → /cross:fix-cr → push
/plugin install cross@laight
GitHub-only. The more clearly your CLAUDE.md states its hard rules, the sharper the reviews get.
- Run
/cross:cr checked out on the PR's branch, pushed: the reviewers read your local tree.
/cross:fix-cr doesn't commit or push. It leaves the fixes in your working tree and asks
you to review them; it commits or pushes only if you tell it to.- Opus and Sonnet are the defaults. If one isn't available to you, the skills fall back to the
other and say so in the review.
| What it does |
|---|
/cross:cr [PR] | Runs the two reviewers in parallel, sorts their findings into both / one only / contradicted, then writes the final review itself (re-checking single-reviewer findings, dropping noise, adding its own with your session's context) and posts one COMMENT review with inline comments tagged [high], [medium] or [low]. Who raised each finding, open questions and dropped findings go to a local, git-ignored scratch.review. |
/cross:fix-cr [PR] | Two independent Opus planners (one aiming for the smallest change, one for the root cause) give every finding a verdict: fix, reject, already fixed, out of scope → issue, spec disagreement, or needs you. The session compares their plans using the same A+B/A/B/S tiers, shows you the decisions, applies the fixes test-first and replies on every thread. |
code-reviewer agent | The reviewer /cross:cr runs twice. Reviews against your project's own spec and checks each finding against the code before reporting it. |
fix-planner agent | The planner /cross:fix-cr runs twice. Reads the code, never edits it. |
Why it's built this way
여기까지가 README 앞부분입니다. 나머지는 원문에서 읽을 수 있습니다. README 전문 보기 ↗
