deep-pr-reviewlisted
Install: claude install-skill zaffnet/whetstone
# Deep PR Review
1. Parse the GitHub PR URL from the user message, if any. Use the models they named; otherwise pick the strongest model from each provider you can reach (for example one OpenAI, one Anthropic, one Google model) so the reviews are independent.
2. Launch one independent review per reviewer in parallel (host Task/subagent). Each reviewer returns findings tagged by the following severity tags: `act on`, `consider`, `noted`, or `dismissed`.
3. If two or more reviewers agree on a finding, merge them into a single finding and record the agreement count.
4. Sort the findings by severity (`act-on` > `consider` > `noted` > `dismissed`) and by the number of reviewers that agree on it.
5. Each finding will be posted as a single comment. Don't post a comment yet.
## No PR
When the user gives no PR URL, review the current branch against `origin/main` (`git diff origin/main...HEAD`), or the whole working tree if they ask for that. Run the same reviewers and merge the same way, then write the findings to `deep-review.local.md` in the repo root (it matches `*.local.*` in `.gitignore`) instead of posting anything. Use the same severity tags and the same writing rules; cite `path:line`.
## Post
- Inline review comments only after showing them to the user and asking for confirmation. Show comments one by one, i.e., show a comment, ask for confirmation, show the next comment, and so on.
- If you are afraid the comment shown will be blocked by the question asked using AskUser