Comprehensive Review

Comprehensive Code Review

Run parallel specialized code reviews via subagents, covering architecture, security, performance, code quality, requirements compliance, and bugs. Merge findings and let the user act on them. Works with both GitHub PRs and local branch diffs.

Variables

• {TEMP_DIR} — the OS temporary directory (e.g. /tmp on Unix, %TEMP% on Windows). Use it for all intermediate files.

Workflow

Step 1: Determine review mode

Check if the user provided a GitHub PR link.

• PR mode: A PR URL is provided matching https://github.com/<OWNER>/<REPO>/pull/<PR_NUMBER>. Extract owner, repo, and PR number.
• Local mode: No PR URL provided. The review will be based on the diff between the current branch and its base branch, plus any uncommitted changes.

Step 2: Fetch diff and task description via subagent

Call a subagent to gather all change details, save the diff, and checkout the correct branch.

CRITICAL: You MUST spawn a subagent for this step. Do NOT perform the diff-gathering, branch detection, or complexity assessment yourself.

Construct the subagent prompt as follows:

Gather the code change details for a comprehensive code review.

Mode: <PR mode or Local mode>
<If PR mode: Owner: <OWNER>, Repo: <REPO>, PR Number: <PR_NUMBER>>

Instructions:
- PR mode: Fetch PR details via GitHub API (title, description, diff). Use `gh pr view <PR_NUMBER> --repo <OWNER>/<REPO> --json title,body` for metadata and `gh pr diff <PR_NUMBER> --repo <OWNER>/<REPO>` for the diff. Save diff to {TEMP_DIR}/review-diff-<branch>.patch.
- Local mode: Detect the current branch and its base branch (try main, master, develop). Use `git diff <base>...HEAD` for committed changes and `git diff` for uncommitted changes. Combine into {TEMP_DIR}/review-diff-<branch>.patch.
- Determine complexity: simple (<200 diff lines, single file/module), medium (200-800 lines, 2-5 files), hard (>800 lines, 6+ files, or touches architecture).
- Return: diff file path, diff line count, title, comprehensive task description, and complexity assessment.

Use a subagent tool to spawn the subagent. Use a powerful model since this involves complex reasoning to assess complexity.

The subagent will return:

• Diff file path: absolute path to the saved diff file (must start with /, e.g. {TEMP_DIR}/review-diff-feature.patch)
• Diff line count: total number of lines in the diff file
• Title: the PR title or summary from commits
• Description: comprehensive task description with all requirements
• Complexity: one of simple, medium, or hard

Remember these values for use in subsequent steps.

Step 3: Run specialized reviews

The review strategy depends on the complexity returned by the fetch-diff subagent.

Review Criteria

Each subagent reviews the diff through one specialized lens:

Note: This skill is self-contained. Referenced files (criteria/*.md, fetch-diff.md, scripts/post_review.js) are placeholders for future expansion. All review instructions and criteria are defined inline above.

Strategy A: Simple complexity

For simple PRs, perform the review yourself (the root agent) without calling subagents.

1. Read the Review Criteria table above to understand each review lens.
2. Read the diff file.
3. Apply all criteria to review the change yourself, producing findings as a flat list without priorities (same format as subagents would). You will assign priorities in Step 4.

Strategy B: Medium complexity

For medium PRs, launch 6 parallel subagent calls — one per review criterion.

CRITICAL: You MUST spawn subagents for this step. Do NOT perform the reviews yourself.

Use a subagent tool to spawn each subagent. Select the single most powerful model from each available provider. Alternate these models across the 6 criteria (e.g., provider A's best model for criteria 1, 3, 5 and provider B's best model for criteria 2, 4, 6). If only 1 provider is available, use its most powerful model for all 6.

Construct prompts for subagents as follows:

Review the following change through the {criterion} lens. Apply the criteria defined in the Review Criteria table for this category.

## <title>

### Task Description
<task description>

### Diff
Read the diff from file: <absolute path to diff file> (total lines: <diff line count>)

IMPORTANT: Do NOT invoke the Skill tool. Do NOT run tests, builds, linters, or type-checks — your review is based on static analysis only.

Strategy C: Hard complexity

For hard PRs, launch 2 parallel subagent calls per criterion (12 total) — one per criterion per model, using 2 different models from different providers for diverse perspectives.

CRITICAL: You MUST spawn subagents for this step. Do NOT perform the reviews yourself.

Model selection: Choose exactly 2 models — the single most powerful model from each of 2 different providers. If only 1 provider is available, use its most powerful model for all 12 calls (fall back to Strategy B behavior with 2 calls per criterion).

Use a subagent tool to spawn each subagent.

For each of the 6 criteria, launch 2 subagents — one with each model. Use following prompt:

Review the following change through the {criterion} lens. Apply the criteria defined in the Review Criteria table for this category.

## <title>

### Task Description
<task description>

### Diff
Read the diff from file: <absolute path to diff file> (total lines: <diff line count>)

IMPORTANT: Do NOT invoke the Skill tool. Do NOT run tests, builds, linters, or type-checks — your review is based on static analysis only.

All 12 calls should be launched in parallel.

Step 4: Merge, filter, and prioritize results

Subagents return findings as flat lists without priority or severity labels. The root agent is responsible for deduplication, false-positive filtering, and priority assignment.

4a. Collect and deduplicate

1. Collect all findings from each review (self-review or subagent responses).
2. Group findings by file and line number.
3. Merge issues that describe the same problem (same file, similar line range, same category). When merging, combine their review types. For hard PRs, findings from different models on the same criterion that agree strengthen confidence; findings from only one model should be noted as lower confidence.
4. Keep the best description, suggested fix, and Diff line from among duplicates. Track Diff line values internally for use in Step 7 but do NOT include them in the output shown to the user.

4b. Filter false positives

Review each finding and discard it if:

• The subagent itself expressed doubt about whether it's a real issue.
• You can verify from the code context that the issue does not apply (e.g., the code is already protected by a guard the subagent missed, or the flagged pattern is intentional and correct).
• The finding is about a pre-existing issue not introduced by the change.

Be conservative — when in doubt, keep the finding.

4c. Assign priorities

Assign a priority to each remaining finding using these levels:

When assigning priorities, consider:

• Cross-criteria signal: A finding flagged by multiple criteria (e.g., both architecture and security) is likely more important.
• Blast radius: Issues affecting hot paths, public APIs, or security boundaries deserve higher priority.
• Confidence: Findings confirmed by multiple models (hard PRs) or backed by concrete evidence deserve higher priority than speculative ones.
• Context: The same issue type may deserve different priorities depending on the codebase and change context.

4d. Format output

Sort by priority (P0 first), then by file path. Each finding must be a standalone entry — do NOT bundle multiple distinct issues into a single row even if they are related.

## Comprehensive Review Findings

| # | Priority | Issue | File:Line | Review type |
|---|----------|-------|-----------|------------|
| 1 | P0 | Description | link to specific line in file | architecture(opus-4-6-think), security(gpt-5-3-codex) |
| 2 | P1 | Description | link to specific line in file | bugs(opus-4-6-think) |
| ... | | | | |

### Details

#### 1. [P0] Issue title
**File:** link to specific line in file
**Review type:** architecture(opus-4-6-think), security(gpt-5-3-codex)

Description and why it matters.

**Suggested fix:**
\`\`\`
code
\`\`\`

Step 5: Ask user how to handle each finding

Skip this step if the user already specified what to do with findings in their initial prompt (e.g., "fix all issues", "post comments for all P0s", etc.). In that case, proceed directly to Steps 6/7 based on their instructions.

Otherwise, send a single message asking the user which findings to fix or post as comments. List each finding with its number, priority, and short description, and ask the user to reply with their choices.

• PR mode: Ask which issues to fix and which to post as PR comments.
• Local mode: Only ask which issues to fix. Do NOT mention posting comments.

Step 6: Apply fixes for issues marked "Fix"

For each finding the user chose "Fix":

1. Read the relevant file and understand the surrounding context.
2. Apply the suggested fix (or an appropriate fix if the suggestion is incomplete).
3. After all fixes are applied, present a summary of changes made.

Step 7: Post selected issues as PR comments (PR mode only)

Skip this step entirely if no findings were marked "Post comment".

Only applies in PR mode. Post line-specific comments via the GitHub API.