Agent Skills: CodeReview Skill

CodeReview Skill

Use this skill to run AI-powered code review via roborev — a local, MIT-licensed review tool with explicit OpenCode support.

USE WHEN

• Completing a BUILD phase and want to catch issues before committing
• Running VERIFY phase and need reproducible quality evidence
• Reviewing a PR diff before merging
• Auditing plugin handler code for pattern violations (no console.log, handler structure)
• You want AI-powered architectural review of your changes
• Trigger phrases: "review code", "check my changes", "roborev", "code quality check"

MANDATORY

1. roborev must be installed — brew install roborev-dev/tap/roborev
2. One-time init — roborev init (installs post-commit git hook)
3. .roborev.toml must exist at repo root with agent = "opencode" and review_guidelines
4. Invoke via code_review tool (preferred) or roborev review --dirty in EXECUTE phase
5. Cite exit code 0 as VERIFY evidence: code_review tool returned exit 0 — no findings

OPTIONAL

• roborev skills install — installs the roborev OpenCode skill (adds roborev commands to agent)
• mode: "last-commit" — review the last git commit instead of dirty working tree
• mode: "fix" — feed findings to the agent for automatic fixes
• mode: "refine" — run auto-fix loop until review passes
• path argument — focus review on a specific file or glob

What roborev Does

roborev analyzes your staged/uncommitted changes (or last commit) using an LLM and surfaces:

• Code quality issues
• Security concerns
• Architectural violations
• Style inconsistencies
• Bugs and edge cases

All review runs locally. No account or cloud service required.

Quick Reference

# Review uncommitted changes (most common)
roborev review --dirty

# Review last commit
roborev review

# Feed findings to agent for fixes
roborev fix

# Auto-fix loop until review passes
roborev refine

# Install git post-commit hook (one-time)
roborev init

# Install OpenCode skill for roborev
roborev skills install

Algorithm Integration

When to invoke

The Algorithm invokes code review in two ways:

1. Via code_review tool (plugin-provided)

Call the code_review tool directly from any Algorithm phase:

Use code_review tool with mode="dirty" to review uncommitted changes before commit.

2. Via roborev CLI in EXECUTE/VERIFY phase

# In EXECUTE: review before committing
roborev review --dirty

# In VERIFY: evidence that review passed
roborev review --dirty && echo "PASS" || echo "FINDINGS"

Recommended Algorithm workflow

BUILD → commit changes
EXECUTE: roborev review --dirty
  → If PASS: continue to VERIFY
  → If FINDINGS: address in next BUILD iteration
VERIFY: cite roborev exit code 0 as evidence

Installation

macOS / Linux (Homebrew)

brew install roborev-dev/tap/roborev

Go

go install github.com/roborev-dev/roborev@latest

Verify

roborev --version

One-Time Setup

# 1. Install git post-commit hook (auto-reviews on every commit)
roborev init

# 2. Install OpenCode skill (adds roborev commands to agent)
roborev skills install

# 3. Verify config exists at repo root
cat .roborev.toml

Configuration (.roborev.toml)

This repo's config is at .roborev.toml in the root. Key settings:

agent = "opencode"

review_guidelines = """
# PAI-OpenCode Review Guidelines
...
"""

The review_guidelines field gives roborev domain-specific rules for this project — including the no-console.log constraint, file-logger pattern, and model routing rules.

Troubleshooting

roborev not found

# Install via Homebrew
brew install roborev-dev/tap/roborev

# Or check if it's in PATH
which roborev
echo $PATH

Review times out

Default timeout is 2 minutes. For large changesets, focus the review:

roborev review --dirty -- src/specific/file.ts

No changes found

Make sure you have uncommitted changes:

git diff
git diff --cached  # staged changes

Post-commit hook not running

Re-run roborev init to reinstall the hook:

cat .git/hooks/post-commit  # verify hook exists
roborev init                # reinstall if missing

PAI-OpenCode Specific Guidelines

When running code review on this project, roborev checks for:

1. No console.log — all plugin logging via fileLog() / fileLogError()
2. Handler pattern — new capabilities = new handler file + import + registration
3. No hardcoded models — model routing via opencode.json only
4. TypeScript strict — no implicit any, explicit return types on exports
5. Biome formatting — tabs, 100 char line width, double quotes

Related

• .roborev.toml — project review configuration
• ADR-018 — architectural decision for roborev integration
• .opencode/plugins/handlers/roborev-trigger.ts — plugin handler providing code_review tool
• .github/workflows/code-quality.yml — CI pipeline (Biome check on PRs)