agentsclimarketplace

Review and address

Skill pricklywiggles/fractally-claude-marketplace/plugins/ship-it/skills/review-and-address

A Claude Code plugin marketplace: ship-it drives a batch of issues to merged PRs; stack-it takes a project from an empty folder to a verified, documented stack.

Install
npx -y skills add pricklywiggles/fractally-claude-marketplace --skill review-and-address

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

2 things to look at

  • no licenseNo license file was found in the repository. Code published without one is not open source by default, so using it at work is a question for whoever answers licensing questions where you are.
  • 0 stars0 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

Review a branch or the current diff with the project's configured reviewers run in parallel, merge their findings, and apply the warranted ones. Use for "review my changes", "review this PR/branch", "run the reviewers on the current diff with main", "review and fix the warranted feedback". Runs every reviewer in ship-it.config review.reviewers (e.g. pr-review-toolkit, vercel-react-best-practices). Also the review stage of the ship-it orchestrator. Not for watching or fixing CI (use ci-fix) or comment-only cleanup (use comment-cleanup).

SKILL.md

4.2 KB, 861 tokens by cl100k_base, as published. Nobody here has run it

review-and-address: fan out reviewers, merge findings, apply warranted

Review a diff with every configured reviewer in parallel, merge the findings, and (when asked or configured) apply the warranted ones. Reviewers only read; the one write step is the address.

1. Resolve the target and the diff

  • PR, branch, or worktree given (an argument, or passed by the ship-it orchestrator as a work-unit): review its diff, git diff <base>...HEAD (base defaults to repo.mainBranch, usually main).
  • Nothing given (standalone): review the current changes, the working-tree diff against main, or the current branch's diff with main. This is the "current diff with main" case.
  • Note the changed files; reviewers and the address step operate only on this diff.

2. Read config, else defaults

Load the resolved config via ${CLAUDE_PLUGIN_ROOT}/scripts/load-config.sh (defaults applied, @FILE refs inlined); read keys with jq:

  • review.reviewers (the list; if absent, default to a single pr-review-toolkit agent reviewer),
  • review.applyWarranted (default: true in the orchestrator; for a standalone run, report unless asked to apply),
  • verify, houseRules, safety (for the address step).

3. Fan out the reviewers (parallel, read-only)

Run every reviewer in review.reviewers over the diff concurrently: spawn one subagent per reviewer in a single message so they run in parallel. Each reviews ONLY the diff and returns structured findings. Dispatch by kind:

  • agent: spawn the Agent tool with subagent_type = the reviewer's ref (e.g. pr-review-toolkit:code-reviewer), prompting it to review the diff and return findings.
  • skill: spawn a general subagent that invokes the Skill tool with the reviewer's ref (e.g. vercel-react-best-practices) scoped to the diff, and return its findings.
  • command: run the reviewer's ref via Bash against the changed files; parse its output into findings.

If a reviewer is not installed, skip it and note it (do not fail the whole review). If review.reviewers is empty, there is nothing to review; say so and stop.

Each finding: { severity, file, location?, issue, recommendation, reviewer }.

4. Merge and de-duplicate

Combine all reviewers' findings. Collapse items that flag the same file and location into one, keeping the highest severity and listing every reviewer that raised it. Order by severity.

5. Address (apply the warranted ones)

Apply only when review.applyWarranted is true, or when this run was explicitly asked to fix:

  • Apply real, in-scope fixes: all blocker / high / medium, plus clearly-correct low / nit. Skip out-of-scope, incorrect, or noise items and record why.
  • Edit only within the target working copy and the diff's scope.
  • Re-verify with the verify commands; fix anything you broke.
  • Commit (concise imperative subject; honor houseRules: no em dashes, no AI attribution).

A report-only run (standalone, not asked to fix) stops after the merged findings; it does not edit.

Hard rules

Reviewers are read-only. The address step applies only genuine, correct, in-scope fixes, never a change that just silences a finding. Stay within the diff. When run by the orchestrator, run non-interactively: apply per applyWarranted and record skips rather than asking.

Output

  • Standalone: the merged findings (severity, file, issue, which reviewers), and, if you applied, what you applied and what you skipped with reasons.
  • Called by the orchestrator: a structured result, { findings: [...], applied: [...], skipped: [...] }, for finalize to push.

What ships with it: 1 file

703 B alongside SKILL.md

Keep looking

Skills are one crate of 328,083. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.