lorenzini

Claude Code skills that wait for a third-party pull-request reviewer and decide whether its verdict actually means pass.

A shark's ampullae of Lorenzini are electroreceptors: they find prey buried in sand, where there is nothing to see. That is what these skills are for. Every reviewer so far has reported findings in places its own count ignores, and every bug in this repository has been the same mistake — treating "I cannot see a problem" as "there is no problem".

What is here

The CodeRabbit/Copilot split is a vendor constraint, not a preference: CodeRabbit's OSS tier requires a public repository with fewer than ten stars to have its reviews triggered by hand, so those repositories moved to Copilot. Check which applies with gh api repos/OWNER/NAME -q .stargazers_count rather than assuming.

Each skill polls until a verdict lands on the current head commit and prints one machine-readable RESULT= line. They do not review code; they adjudicate what a reviewer said about it.

Install

Symlink into ~/.claude/skills/:

git clone [email protected]:Nanako0129/lorenzini.git ~/side-project/lorenzini
for s in codex copilot coderabbit; do
  ln -s ~/side-project/lorenzini/$s-review-wait ~/.claude/skills/$s-review-wait
done

Requires gh (authenticated) and jq.

The one rule

Never infer a pass from absence.

A verdict requires an explicit positive completion marker. No marker means keep polling — not "nothing found, therefore clean". Five separate bugs have come from breaking this rule in five different ways, and all five failed open: they reported success. docs/fail-open-ledger.md records each one, what it cost, and how it was caught.

That ledger is the most valuable file here. The scripts encode its conclusions, but the conclusions look arbitrary without the failures that produced them, and a guard whose reason has been forgotten gets "simplified" away by the next person.

Testing a change

The regression cases named in each SKILL.md are live public pull requests. Their state changes for reasons unrelated to this code — pysnmp/pysmi#328 moved from SUGGESTIONS to PREMERGE between two runs an hour apart because that repository's maintainer resolved a thread in between. Confirm any behaviour change with a controlled comparison (same captured payload, one input varied) before believing either the pass or the failure.