Enhanced PR Review System

[blacksyan]

Enhanced PR Review System

A context-aware, multi-phase AI code review system for GitHub PRs using OpenCode.

---

🎯 Problem & Solution

The Problem

Current OpenCode capabilities for PR reviews have some limitations:

• Limited project context - Reviews may lack deep alignment with specific project conventions.
• Strict enforcement everywhere - High friction for experimental or non-protected branches.
• One-way feedback - Minimal opportunity to clarify intent before the AI issues a decision.
• Single-line focus - Feedback is often isolated to single lines, making multi-line refactors harder to suggest.

The Solution

This system enhances OpenCode with a context-aware review layer that:

• ✅ Reads your actual code using OpenCode's native file-reading power.
• ✅ Asks questions first before reviewing (Phase 1 → Phase 2).
• ✅ Branch-aware strictness - Strict for main, relaxed for feature branches.
• ✅ Multi-line committable suggestions - One-click apply for complex fixes.

---

📋 Table of Contents

• Features
• How It Works
• Quick Start
• Usage Commands
• Configuration
• Customization
• Test Results

---

✨ Features

🔍 OpenCode File Reading

The system leverages OpenCode's native capability to read files from your repository. You can verify this in the GitHub Actions logs:

Sending message to opencode...

|   Read     {"filePath":"src/services/VideoPostLikeService.js"}
|   Read     {"filePath":"src/services/VideoPostCommentService.js"}
|   Read     {"filePath":"src/models/VideoPost.js"}

This ensures the AI analyzes your actual source files to understand the broader context beyond just the diff.

🎯 Multi-Phase Review Flow

Phase	Trigger	Purpose
Phase 1	/oc	AI asks clarifying questions to understand your intent
Phase 2	Answer	Context-aware review based on your detailed answers
Direct Review	/oc!	Skip questions and get an immediate review
Recheck	/oc recheck	Request a re-review after you've pushed fixes

🎛 Branch-Aware Strictness

The review logic adapts to your target branch:

Target Branch	Behavior
main, staging, production	Strict - Missing concurrent dependencies (e.g., migrations vs models) are marked as FAIL and block the PR.
feature/*, trial/*, develop	Relaxed - Missing dependencies are marked as INFO/SKIPPED. The AI can APPROVE while providing reminders for later.

Configure via: STRICT_REVIEW_BRANCHES=main,staging,production

📝 Multi-Line Committable Suggestions

Comments can now span multiple lines (e.g., lines 42-65) and include GitHub's suggestion syntax:

📍 Lines 42-65 in src/services/MyService.js

⚠️ Warning: This block needs a transaction wrapper to ensure atomicity.

```suggestion
async myMethod() {
  return transaction(async (t) => {
    // your fixed code here
  });
}
```

Users can click "Commit suggestion" directly on GitHub to apply the fix.

📚 Context Injection System

• 89+ auto-detected patterns to trigger relevant context loading.
• RAKAMIND.md / OPENCODE.md support - Define per-repo AI rules.
• Concurrent implementation checks - Automatically detects missing cross-layer logic (model-migration-sync, service-transaction-checks).

---

🔄 How It Works

1. User Triggers Review: Comment /oc on a PR.
2. OpenCode Reads Files: The system identifies relevant files and reads them via OpenCode's tools.
3. Phase 1 (Clarification): AI asks questions if information is missing.
4. Phase 2 (Review): AI issues a structured review with a checklist, inline comments, and a decision.

---

🚀 Quick Start

1. Add Workflow to Your Repo

Create .github/workflows/opencode.yml in your target repository:

name: opencode

on:
  issue_comment:
    types: [created]
  pull_request_review_comment:
    types: [created]

jobs:
  opencode:
    if: |
      (github.event_name == 'issue_comment' && github.event.issue.pull_request) ||
      github.event_name == 'pull_request_review_comment'
    runs-on: self-hosted
    steps:
      - uses: actions/checkout@v4
        with:
          ref: enhance/pr-agent-capabilities-v2
          repository: rakamindev/opencode

      - uses: oven-sh/setup-bun@v2
        with:
          bun-version: latest

      - name: Install dependencies
        run: cd packages/opencode && bun install

      - name: Run opencode
        env:
          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
          GOOGLE_GENERATIVE_AI_API_KEY: ${{ secrets.GOOGLE_GENERATIVE_AI_API_KEY }}
          STRICT_REVIEW_BRANCHES: main,staging,production
        run: |
          cd packages/opencode && bun run build
          cd packages/opencode && ./bin/opencode github run

2. Add Required Secrets

In your repo Settings → Secrets → Actions:

• GOOGLE_GENERATIVE_AI_API_KEY: Your Gemini API key.
• GITHUB_TOKEN: Automatically handled if using default permissions.

3. Add Project Rules (Optional)

Create RAKAMIND.md (or OPENCODE.md) in your repo root to guide the AI:

# RAKAMIND.md

## Project
- Node.js 20 / Express 4 / Sequelize 6
- Layers: controllers → services → models

## Code Rules
- Services: multi-step writes MUST use transactions.
- Models: use snake_case, paranoid: true.

---

📊 Test Results

Verified on rakamindev/paragon-api:

• PR #414 (Migrations only): APPROVE ✅ - Correctly skipped models as out of scope for a feature branch.
• PR #415 (Models): REQUEST_CHANGES ❌ - Found missing hooks in model files.
• PR #416 (Services): REQUEST_CHANGES ❌ - Identified missing transactions and provided multi-line committable suggestions.

---

📝 Changelog

• Multi-line suggestions - Better UX with committable code blocks.
• Branch-aware strictness - Reduced friction for development branches.
• RAKAMIND.md support - Custom per-repo instructions.
• Context injection - Automated deep learning of the codebase.
• Phase 1/2 flow - Improved accuracy through developer collaboration.