Comprehensive Code Review (Codex + Gemini)

Overview

Run Codex and Gemini code reviews in parallel using the Task tool, then synthesize both results into a single consolidated report. This provides broader coverage by leveraging two independent AI models with different strengths.

Workflow

Step 1: Determine Review Scope

Ask the user what they want reviewed if not already clear:

Scope	When to use	Codex	Gemini
Branch diff	Before opening or merging a PR	Yes	Yes
Uncommitted changes	During active development	Yes	Yes
Specific commit	Auditing a single changeset	Yes	Yes
Remote PR	Reviewing a GitHub Pull Request by number or URL	No	Yes
Custom	User provides specific review instructions	Yes	Yes

Note: Remote PR scope is only supported by Gemini. When user selects PR, run Gemini review only and skip Codex.

Default to reviewing the current branch diff against main unless the user specifies otherwise.

Step 2: Run Both Reviews in Parallel

Launch two Task agents simultaneously in a single message. Both agents must run in the same message to ensure true parallelism.

• Agent 1: Activate the codex-review skill and run it with the determined scope.
• Agent 2: Activate the gemini-review skill and run it with the same scope.

Pass the user's review scope (branch, uncommitted, commit SHA, PR number, custom focus) to both agents so they review the exact same changes.

Tell each agent to capture and return the full review output.

Step 3: Synthesize Results

Once both agents return, produce a single consolidated report using this structure:

---

Comprehensive Review

Quality: X/5 · Confidence: X/5 · Verdict:

Where:

• Quality (1-5): Combined quality assessment across both reviewers
• Confidence (1-5): How much the reviewers agree (5 = full consensus, 1 = major contradictions)
• Verdict: Approved / Approved with suggestions / Request Changes

Summary

3-5 sentence overview combining both reviewers' assessments. Note where they agree and disagree.

Changes Walkthrough

File	Changes
path/to/file.ts	Brief description of what changed

Findings

All issues merged, deduplicated, and sorted by severity. Each finding must follow this format:

[Category] file/path.ts:LINE — Short title [Codex] [Gemini] [Both]

Explanation of the issue and why it matters.

Suggested fix:

code suggestion here

Where:

• Category: Bug, Security, Performance, Maintainability, Edge Case, Testing, Style
• Source tag: [Codex], [Gemini], or [Both] — indicates which reviewer(s) flagged it
• Findings tagged [Both] should appear first within their severity group (highest confidence)
• If both flagged the same issue with different suggestions, present both and note the difference

If none: "No issues found by either reviewer."

Highlights

1-3 positive patterns that one or both reviewers called out, attributed by source.

Reviewer Agreement

Metric	Details
Consensus	Issues both flagged independently (list count or summary)
Codex only	Findings unique to Codex
Gemini only	Findings unique to Gemini
Conflicts	Contradictory recommendations, if any (explain both perspectives)

Verdict

Restate the verdict with a 1-2 sentence justification referencing the combined findings.

Include both full outputs here for reference, clearly labeled:

• Codex output: full text
• Gemini output: full text

---

Rules

• Always run both reviews in parallel using two Task agents in a single message — never sequentially
• Use the same review scope and options for both reviewers to ensure a fair comparison
• Deduplicate findings that both reviewers flagged — mark them as [Both] with highest confidence and list first within their severity group
• When reviewers contradict each other, present both perspectives and let the user decide
• Attribute every finding to its source ([Codex], [Gemini], or [Both])
• If one CLI is not installed, warn the user and fall back to running only the available reviewer — do not fail entirely
• Sort findings by severity: Bug/Security first, then Performance/Maintainability/Edge Case/Testing, then Style
• The consolidated report should be shorter than both individual reports combined — synthesize, don't concatenate
• Always include raw reviewer outputs in a collapsible <details> section at the end for transparency