skills.rest
sndrgrdn
sndrgrdnCOMMUNITY
Explore all skills by this author →
WebsiteGitHub

deep-review

Reviews a pinned git diff with four parallel judge agents and a mechanical pass/fail gate.

Updated Apr 23, 2026
Search Tags:#static analysis#code review#pull request#software architecture#git diff#merge gate#review panel
One-click install
npx skills add https://github.com/sndrgrdn/skills --skill deep-review-sndrgrdn
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: deep-review
Source: https://github.com/sndrgrdn/skills/tree/main/deep-review
Command: npx skills add https://github.com/sndrgrdn/skills --skill deep-review-sndrgrdn

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes references (resource) components.

What problem does it solve? Code review of a branch or PR often stays shallow, inconsistent, or swayed by the author's framing. This Skill runs a structured panel review of a pinned diff, dispatching four independent judge agents (Design, Craft, Fitness, Stability) that must back every finding with evidence, then applies a mechanical gate to produce a single PASS, FAIL, or INCOMPLETE verdict. ## Core Features & Use Cases - Four specialized judges: Design checks whether the right modules exist, Craft checks code quality rules and smells, Fitness challenges whether the change should exist at all, and Stability traces the blast radius of altered behavior. - Evidence-gated findings: Every blocking finding must cite a rubric rule tagged as a presumptive blocker and meet its evidence bar (file:line citations, module tables, premise tables, traced dependents), so verdicts rest on proof rather than opinion. - Pinned, reproducible scope: The review range is pinned to explicit base/head SHAs or a PR's merge base, with a fact-only census (stated purpose, premises, siblings, trajectory, house record, diff stats) built before any judging. - Use Case: Before merging a large feature PR, ask for a deep review of the PR; the Skill gathers the diff and context, runs the four judges in parallel, and returns a decision brief listing blocking findings, unchecked obligations, and per-seat summaries. ## Quick Start Run a deep review of PR number 123 with all four judges and give me the decision brief.

Frequently Asked Questions about deep-review

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I run a multi-perspective code review on a pull request?▼

Provide the PR number or URL and request a deep review. The Skill pins the PR's base and head commits, builds a fact-only census of the diff, then dispatches Design, Craft, Fitness, and Stability judges in parallel and returns a gated decision brief.

How do I review only part of a branch, like design and stability?▼

Name the judge subset when invoking the review, such as Design plus Stability. Unselected judges are omitted from scope, and the verdict gate resolves only over the selected seats.

What makes a finding blocking in this code review process?▼

A finding is blocking only when it cites a rubric rule explicitly tagged as a presumptive blocker and meets that rule's evidence bar, such as file:line citations or a traced dependent. Findings under untagged rules are non-blocking regardless of the judge's label.

Can I review a diff without a pull request?▼

Yes, supply an explicit base and head revision. Both are resolved to full commit SHAs with git rev-parse, the base must be an ancestor of the head, and the pinned diff must be non-empty before the review proceeds.

What tools does the deep review workflow require?▼

It requires git for revision resolution, diff capture, and history inspection, plus the GitHub CLI (gh) for PR metadata and discussion cross-checks. Missing commit objects or unrelated histories stop the review at the pinning step.

What are the limitations of an automated panel code review?▼

The review is bounded by the pinned diff and in-repo evidence; it cannot execute the code or verify runtime behavior beyond what is traceable in source. Rules that cannot be verified are reported as unchecked, which yields an INCOMPLETE verdict rather than a pass.

Related Skills

Explore More Software Engineering Skills→
giruclawbotgiruclawbot

github-pr

Generate conventional commit PR titles and Summary/Changes/Testing descriptions.

Community
Intermediate
matthewaltonmatthewalton
1

address

Fixes findings from CI-posted pull request reviews and pushes the commit to the PR branch.

Community
Advanced
NotysotyNotysoty
9

Agent Evaluation Framework Builder

Design evaluation suites for LLM agents with datasets, metrics, and CI integration.

Community
Advanced
swarm-ai-researchswarm-ai-research

PR Review

Generate severity-tagged findings, inline comments, and verdicts for pull requests.

Official
Advanced