multi-review

name	multi-review
description	Multi-model code review. Runs code-review skill with 2 models in parallel, then synthesizes findings.

Multi Review

Runs the code-review skill with 2 different models in parallel, then synthesizes with active validation.

Process

Phase 1: Gather Reviews

1. Create a unique temp dir + get the PR diff (same as code-review)

   # Unique temp dir for this run
   TMP_DIR="$(mktemp -d -t multi-review.XXXXXX)"
   PR_DIFF="$TMP_DIR/pr-diff.txt"
   
   # If PR number provided, use it. Otherwise current branch.
   gh pr diff [PR_NUMBER] > "$PR_DIFF"
   

2. Run 2 parallel reviews via bash

   # Claude review MUST use the `claude` CLI (not `pi -p --model ...`) so it uses Claude auth.
   claude -p --model opus --permission-mode bypassPermissions \
     "Read and follow /Users/pat/Work/pi-skills/skills/code-review/SKILL.md to review the PR. Diff is at $PR_DIFF" \
     > "$TMP_DIR/review-opus.md" &
   
   # Codex review continues to use pi.
   pi -p --model gpt-5.5 --provider openai-codex \
     "Read and follow /Users/pat/Work/pi-skills/skills/code-review/SKILL.md to review the PR. Diff is at $PR_DIFF" \
     > "$TMP_DIR/review-codex.md" &
   
   wait
   

   If Claude fails because of auth or local lock contention, rerun only the Claude command after fixing auth (claude auth) or retrying.

Phase 2: Active Validation (IMPORTANT)

Do not blindly trust the reviewers. Validate each finding yourself.

3. Read PR context first Before looking at sub-agent reviews, get the full picture:

   # What the PR claims to do
   gh pr view [PR_NUMBER] --json title,body
   
   # What it actually does
   cat "$PR_DIFF"
   
   # What others have already said
   gh pr view [PR_NUMBER] --json comments,reviews --jq '.comments[].body, .reviews[].body'
   

   Form your own impressions. Note any issues already flagged in PR feedback.

4. Collect all findings Build a deduplicated list of every issue from both reviews. Note which model(s) found each issue.

5. Validate EACH finding For every finding, actually look at the code and verify:

   • Is this a real bug/issue? (check the code, don't just trust the claim)
   • Is it a false positive? (model hallucinated or misunderstood)
   • What file/line is affected? (verify it exists and matches)

6. Score by IMPACT, not consensus Rate each validated issue by actual severity:

   • 🔴 Critical: Breaks functionality, security issue, data loss
   • 🟠 High: Real bugs, incorrect behavior, major guideline violations
   • 🟡 Medium: Performance, maintainability, edge cases
   • 🟢 Low: Style, minor improvements, nitpicks

   Consensus count (both models) ≠ importance.

   • Consensus often means "obvious issue any reviewer would catch"
   • Unique findings may be subtle insights worth MORE attention, not less

7. Flag unique findings for extra scrutiny When only one model found something:

   • WHY did only one catch it? (deeper insight vs hallucination?)
   • Validate more carefully - could be the most important find
   • Could also be a false positive - verify against actual code

8. Check for gaps What might BOTH models have missed?

   • Complex state/timing issues (e.g., async race conditions)
   • Claimed features that don't actually work (check PR description)
   • Subtle logic errors in control flow
   • Look at the PR description - are all claims implemented?

Phase 3: Synthesized Output

9. Output format

# 🔍 Multi-Model PR Review: [PR title]

## Validated Issues

### 🔴 Critical
[Issues that must be fixed - functionality broken, security, etc.]

### 🟠 High Priority  
[Real bugs, incorrect behavior - should fix before merge]

### 🟡 Medium Priority
[Performance, maintainability, edge cases - should discuss]

### 🟢 Low Priority
[Style, minor improvements - nice to have]

Each issue should include:
- **File**: path/to/file.ext#L10-L15
- **Status**: ✅ Confirmed | ⚠️ Needs verification | ❌ False positive
- **Found by**: Opus / Codex / PR feedback
- **Description**: What's wrong and why it matters
- **Suggestion**: How to fix (if applicable)

## ❌ False Positives Filtered
[List any findings that were wrong, with brief explanation]

## ⚠️ Potential Gaps
[Things all models may have missed - especially check PR description claims]

## 📊 Model Coverage
| Issue | Opus | Codex | PR | Status |
|-------|:----:|:-----:|:--:|--------|
| Issue 1 | ✅ | ✅ | - | ✅ Confirmed |
| Issue 2 | ❌ | ✅ | - | ✅ Confirmed |
| Issue 3 | ✅ | ❌ | - | ❌ False positive |
| Issue 4 | ❌ | ❌ | ✅ | ⚠️ Models missed! |

## Final Verdict
**[MERGE / FIX FIRST / NEEDS DISCUSSION]**

[Brief explanation of verdict]

Key Principles

1. Validate, don't just synthesize - You are the senior reviewer, not a secretary
2. Unique findings deserve MORE attention - They might be the deepest insights
3. Consensus ≠ importance - Obvious issues get caught by all; critical bugs may be subtle
4. Check what's missing - The worst bugs are the ones no one found
5. Compare against PR description - Do claimed features actually work?