Dev Local Review

Run a local code review before pushing to catch issues before they hit the PR and burn a review round. Uses the built-in CodeRabbit code-reviewer agent (subagent_type: "coderabbit:code-reviewer") which runs entirely locally at zero cost.

This is not a substitute for the GitHub PR review. The local review cannot access CodeRabbit's learnings from past PRs or perform repo-wide architectural analysis. It is a fast, free pre-push gate that catches the obvious issues.

How it works

The skill uses Claude Code's built-in coderabbit:code-reviewer agent type. This agent reads the diff locally and performs a thorough code review covering correctness, style, documentation accuracy, and common issues. No external service, CLI tool, or API key required.

Usage

Invoke the skill before pushing:

/dev-local-review

Or with a specific base branch:

/dev-local-review --base develop

When to use

• Before pushing fixes during /dev-review-pr — catch new issues locally

instead of waiting for the GitHub bot

• Before creating a PR with /dev-create-pr — clean up obvious issues

before the first review

• Any time you want a quick code quality check without pushing
• After /dev-final-pass for a second opinion from a different reviewer

Workflow

1. Determine what to review

Check git status and git diff --stat to understand the scope of changes. Use the base branch from the argument (default: main).

2. Run the CodeRabbit code-reviewer agent

Launch a single agent with subagent_type: "coderabbit:code-reviewer". The prompt must include:

• The list of modified and new files (from git diff --stat and git status)
• A summary of what the changes are about (infer from file paths and recent

context)

• Instructions to be thorough and report specific file paths and line numbers
• The project's code quality requirements (C99, SDL3 GPU, CLAUDE.md conventions)

Agent(
  subagent_type="coderabbit:code-reviewer",
  prompt="Review all changes compared to . [file list, context, instructions]"
)

3. Triage findings

When the agent returns, present findings sorted by severity (Major, Medium, Low, Informational). For each finding, determine:

• Actionable: A real issue that should be fixed before pushing
• Intentional: Code that is correct as-is (e.g., strict 0.0f epsilon

for determinism tests)

• Out of scope: Suggestions about code not changed in this PR

4. Fix actionable findings

Implement all actionable fixes. Do not ask about each one individually — fix them all, then report what was changed.

5. Re-run review (MANDATORY — cannot skip)

After fixing, run the agent again to verify no new issues were introduced. The re-run prompt should note which issues were already addressed so the agent focuses on finding new problems.

6. Repeat until zero feedback (MANDATORY — strict rule)

This is an absolute rule with no exceptions. The review loop MUST continue until the CodeRabbit agent returns with zero findings. The agent executing this skill is not allowed to decide on its own that remaining feedback is "just a nitpick" or "informational" and stop the loop early.

The only two valid exit conditions are:

1. The review agent reports zero findings. The loop ends. Proceed to

"After a clean local review."

1. The agent explicitly asks the user whether to stop (presenting the

remaining findings), and the user agrees to stop. Without user confirmation, the loop continues.

Prohibited behaviors:

• Deciding that remaining findings are "low severity" and stopping
• Classifying all remaining findings as "intentional" without user input
• Stopping after a fixed number of rounds (e.g., "typically 2-3 rounds")
• Summarizing remaining findings and moving on without fixing or asking
• Treating nitpicks, style suggestions, or informational comments as

automatic reasons to exit the loop

Each round of the loop:

1. Run the CodeRabbit review agent
2. If zero findings → exit loop (clean)
3. If findings exist → fix actionable ones, ask user about ambiguous ones
4. Go to step 1

If the same finding keeps reappearing after being fixed, investigate the root cause rather than stopping the loop. If genuinely stuck, ask the user for guidance — do not silently exit.

What it catches

Based on project experience, this review reliably catches:

• Missing bounds checks and safety guards
• Stale comments (e.g., ` comment listing fabsf, sqrtf` but code

also uses fminf)

• Documentation/code mismatches (README describes N phases but code has M)
• Missing inline comments on struct fields
• Resource cleanup gaps on error paths
• Magic numbers that should be named constants
• Naming convention violations
• Bare C stdlib calls without SDL_ prefix (fabsf, sinf, memset, strcmp,

malloc, etc.) — use SDL equivalents or approved project wrappers (forge_isfinite, forge_fmaxf, forge_fminf) for cross-platform portability

What it cannot do

• Access CodeRabbit's learnings from past PR interactions
• Perform repo-wide architectural analysis beyond the changed files
• Interact with PR feedback threads or resolve conversations
• Update GitHub approval state
• Guarantee a clean PR review (the GitHub bot may find additional issues)

After a clean local review

A clean local review means the obvious issues are handled. Push and request a full PR review:

1. Push the commits (single push, never piecemeal)
2. Wait for CodeRabbit's PR review — it may find additional issues the

local review missed

1. If the PR review finds issues, use /dev-review-pr to handle them

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: Nebulavenus
• Source: Nebulavenus/forge-gpu
• License: Zlib

Install and usage instructions live in the source repository linked above.