Multi-Model Code Review

Multi-Model Code Review

Run code review using 4 free OpenCode models in parallel, then have them peer-review each other to reach consensus findings.

Usage

/​multi-model-review
/​multi-model-review backend/​src/

Behavior

1. Gather the Diff

Run git diff to get the changes:

git diff HEAD

If arguments were provided, scope the diff:

git diff HEAD -- <path>

If there are no uncommitted changes, say so and stop.

2. Run 4 Models in Parallel (CRITICAL: Use bash, not Task agents)

Launch 4 concurrent bash commands IN PARALLEL using separate tool calls:

# Model 1: big-pickle (fast, concise)
~/​.opencode/​bin/​opencode run -m opencode/​big-pickle "You are a code reviewer. Run git diff HEAD first to see uncommitted changes. Focus on bugs, security issues. Output format: ## Code Review\n### Findings\n1. [P1] Title — file:line\n### Verdict: PASS | NEEDS FIXES" 2>/​dev/​null

# Model 2: gpt-5-nano (thorough)
~/​.opencode/​bin/​opencode run -m opencode/​gpt-5-nano "You are a code reviewer. Run git diff HEAD first to see uncommitted changes. Focus on bugs, security issues. Output format: ## Code Review\n### Findings\n1. [P1] Title — file:line\n### Verdict: PASS | NEEDS FIXES" 2>/​dev/​null

# Model 3: minimax-m2.5-free (balanced)
~/​.opencode/​bin/​opencode run -m opencode/​minimax-m2.5-free "You are a code reviewer. Run git diff HEAD first to see uncommitted changes. Focus on bugs, security issues. Output format: ## Code Review\n### Findings\n1. [P1] Title — file:line\n### Verdict: PASS | NEEDS FIXES" 2>/​dev/​null

# Model 4: trinity-large-preview-free (critical)
~/​.opencode/​bin/​opencode run -m opencode/​trinity-large-preview-free "You are a code reviewer. Run git diff HEAD first to see uncommitted changes. Focus on bugs, security issues. Output format: ## Code Review\n### Findings\n1. [P1] Title — file:line\n### Verdict: PASS | NEEDS FIXES" 2>/​dev/​null

IMPORTANT: Use bash tool with the exact prompt above - do NOT use Task agents, they don't invoke skills properly.

3. Save Individual Reviews

Save each output to:

reviews/
├── review_big-pickle.md
├── review_gpt-5-nano.md
├── review_minimax-m2.5-free.md
└── review_trinity-large-preview-free.md

4. Peer Review Round

Once all 4 complete, create a new prompt with all 4 reviews and ask for peer review:

You have 4 code reviews of the same changes. Review each finding and identify:
1. Which findings are valid (confirmed by code evidence)
2. Which are false positives
3. Any issues missed

Reviews:
[paste all 4 reviews]

Output format:
## Peer Review

### Validated Issues (2+ models agree)
- [issue]

### Disputed/​False Positive Issues
- [issue]

### Missing Issues
- [issue]

## Consensus Verdict: PASS | NEEDS FIXES

5. Synthesize Consensus

Create reviews/​SUMMARY.md with:

• Table comparing models
• Issues grouped by consensus level
• Final verdict

Important Rules

• Use bash commands directly - Not Task agents, not skill invocation
• Run all 4 in parallel - Use separate bash tool calls simultaneously
• Always pipe to /​dev/​null - Avoids noise
• Quote prompts properly - Use single quotes around prompts to prevent shell interpretation of special chars

---

Bonus: Auto-Fix with Opus

After consensus, you can have Opus 4.6 implement fixes:

~/​.opencode/​bin/​opencode run -m anthropic/​opus-4-6 "You are a code reviewer. Read the SUMMARY.md in reviews/ directory for consensus issues. Fix each P1 issue found there. Run tests after fixes."

Or use bash with proper quoting:

~/​.opencode/​bin/​opencode run -m anthropic/​opus-4-6 'You are a code reviewer. Read reviews/​SUMMARY.md for consensus issues. Fix each P1 issue. Run tests after.' 2>/​dev/​null