Runs adversarial code reviews on git diffs using Codex with structured findings and selective fixes.
npx skills add https://github.com/crowi/crowi --skill crowi-review
Please help me install this Agent Skill. Skill: crowi-review Source: https://github.com/crowi/crowi/tree/main/.claude/skills/crowi-review Command: npx skills add https://github.com/crowi/crowi --skill crowi-review
Reviewing uncommitted changes, branch diffs, or commit ranges for correctness, security, and contract issues is time-consuming, and raw AI findings often contain false positives. This Skill orchestrates an adversarial review pass over a defined scope, reports findings with severity levels, and only verifies and fixes the low-risk ones.
main...HEAD to get a severity-ranked findings report, then let the Skill fix the trivial issues while you triage the rest.Run /crowi-review with an optional path or git range to adversarially review that scope and receive a severity-ordered findings report.
High-intent search queries and answers about installing and using this skill.
Invoke the skill with a scope such as a path, git range like abc123..HEAD, or nothing to default to uncommitted changes. It resolves the scope once, runs Codex in read-only sandbox mode, and returns a severity-ordered findings report.
If Codex exits with an error or produces invalid output, the skill falls back to a Claude general-purpose subagent that performs the same adversarial review in the same findings format. The report notes the fallback and tags findings as unverified (claude-fallback).
No. It only fixes low-risk findings when the main branch is clean, verifying each against the actual code first. Larger or judgment-based findings are reported for user decision, and nothing is committed while the working tree is dirty.
No. The skill never pushes or creates PRs, since the repository follows a main-direct workflow. Fixes stop at local commits, and pushing always waits for explicit user instruction.
The skill uses verification-on-action: it does not pre-verify every finding to avoid redundant code reading. Findings are confirmed or refuted only when a fix is attempted, and untouched findings remain tagged unverified for user triage.
aurabxGenerate Conventional Commits messages from working tree or staged diffs.
phenobarbitalPush a git worktree branch, create a PR to main, and clean up.
DataDogRun Hatch-based lint, format, and quality checks on modified files.
yashs33244Merge branches, run tests, review diffs, and prepare release commits for PRs.