Second Opinion

Run external LLM code reviews using OpenAI Codex or Google Gemini CLI

Get independent code reviews from external LLM CLIs (OpenAI Codex, Google Gemini) on uncommitted changes, branch diffs, or specific commits.

Overview

The Second Opinion plugin shells out to external LLM CLIs for independent code reviews powered by separate models. It supports both OpenAI Codex CLI and Google Gemini CLI, with options to run them individually or compare results side-by-side.

Author: Dan Guido

Features

Dual Model Support

Run reviews with OpenAI Codex (gpt-5.3-codex) or Google Gemini (gemini-3.1-pro-preview), or both for comparison

Multiple Scopes

Review uncommitted changes, branch diffs vs main, or specific commits

Focused Reviews

General review, security, performance, or error handling focus areas

Project-Aware

Optionally include CLAUDE.md or AGENTS.md for context-aware reviews

Prerequisites

OpenAI Codex CLI

1. Install Codex CLI

   npm i -g @openai/codex
   
2. Configure Authentication

   Requires OpenAI API key or ChatGPT Plus subscription configured for Codex

Google Gemini CLI

1. Install Gemini CLI

   npm i -g @google/gemini-cli
   
2. Authenticate

   gemini auth login
   
3. Install Extensions

   gemini extensions install https://github.com/gemini-cli-extensions/code-review
   gemini extensions install https://github.com/gemini-cli-extensions/security
   

Installation

/plugin marketplace add trailofbits/skills
/plugin install second-opinion

Usage

1. Invoke the Command

   /second-opinion
   

   or with inline arguments:

   /second-opinion check the uncommitted changes for security issues
   
2. Answer Questions

   Claude asks up to 4 questions (skips any already specified):

   1. Review tool - Codex, Gemini, or both (default: both)
   2. Review scope - Uncommitted, branch diff, or specific commit
   3. Project context - Include CLAUDE.md/AGENTS.md? (if exists)
   4. Review focus - General, security, performance, or error handling
3. Review Diff Stats

   Claude shows what will be reviewed:

   6 files changed, 103 insertions(+), 15 deletions(-)
   

   If the diff is empty, Claude stops. If very large (>2000 lines), Claude warns.

4. Run Review

   For “Both” (default), Claude runs Codex and Gemini in parallel.

   For single tool, Claude runs just that tool.

5. View Results

   Results are presented with clear headers:

   ## Codex Review (gpt-5.3-codex)
   [structured findings by priority]
   
   ## Gemini Review (gemini-3.1-pro-preview)
   [gemini output]
   
   ## Summary
   Both reviews agree on:
   - Missing input validation in auth.py:45
   
   Codex flagged but Gemini didn't:
   - Performance concern in db.py:123
   

Review Scopes

Uncommitted Changes

Reviews all uncommitted work (tracked and untracked files):

git diff HEAD
git ls-files --others --exclude-standard

Branch Diff vs Main

Reviews all changes in current branch since it diverged from default branch:

# Auto-detects default branch (main, master, etc.)
git diff <default-branch>...HEAD

Specific Commit

Reviews changes introduced by a single commit:

git diff <sha>~1..<sha>

Focus Areas

General Review

Broad code review covering:

• Code quality and maintainability
• Potential bugs and edge cases
• Best practices and patterns
• Documentation completeness

Security & Auth

Security-focused review:

• Input validation and sanitization
• Authentication and authorization
• Secrets and credentials management
• Injection vulnerabilities
• Dependency vulnerabilities (if manifest changed)

Performance

Performance-focused review:

• Algorithm efficiency (O(n) analysis)
• Database query optimization
• Memory usage and leaks
• Caching opportunities
• Blocking operations

Error Handling

Error handling review:

• Uncaught exceptions
• Error propagation
• User-facing error messages
• Logging and observability
• Graceful degradation

Codex vs Gemini

Feature	Codex	Gemini
Model	gpt-5.3-codex	gemini-3.1-pro-preview
Reasoning	xhigh	Default
Output	Structured JSON findings by priority	Natural language
Headless	codex exec with stdin	gemini -p with stdin
Extensions	None	code-review, security
Dependency Scan	N/A	/security:scan-deps (when security focus + manifest changed)

Running both (the default) provides comparison and validation. Where they agree, findings are higher confidence.

Codex Invocation Details

codex exec \
  --sandbox read-only \
  --ephemeral \
  --output-schema codex-review-schema.json \
  -o "$output_file" \
  - < "$prompt_file"

• Uses OpenAI’s published code review prompt (fine-tuned into model)
• Diff is piped via stdin with the prompt
• --output-schema produces structured JSON findings
• -o captures only final message (no thinking/exec noise)
• Falls back to gpt-5.2-codex on auth errors
• Timeout: 600 seconds

Gemini Invocation Details

Uncommitted General Review

gemini -p "/code-review" --yolo -e code-review

Branch/Commit Diffs

git diff <scope> > /tmp/review-diff.txt
{ printf '%s\n\n' 'Review this diff for issues.'; cat /tmp/review-diff.txt; } \
  | gemini -p - --yolo -m gemini-3.1-pro-preview

Security Focus

# With custom security prompt piped to stdin
gemini -p - --yolo -m gemini-3.1-pro-preview -e gemini-cli-security

Safety Note: Gemini CLI is invoked with --yolo, which auto-approves all tool calls without confirmation. Required for headless operation.

Dependency Scanning

When security focus is selected AND the diff touches dependency manifest files, Claude also runs:

/security:scan-deps

Manifest files detected:

• package.json, package-lock.json, yarn.lock, pnpm-lock.yaml
• requirements.txt, setup.py, pyproject.toml, uv.lock, poetry.lock
• Cargo.toml, Cargo.lock
• go.mod, go.sum
• Gemfile, *.gemspec
• composer.json, composer.lock

The scan is skipped if no manifest files are in the diff, since it analyzes the entire dependency tree (time-consuming with zero value when deps weren’t touched).

Usage Examples

Both Tools (Default)

User: /second-opinion

Claude: [asks 4 questions]

User: defaults (both tools, uncommitted, no context, general)

Claude: [shows diff stats: 3 files, +45 -10]

Claude: [runs codex exec and gemini in parallel]

Claude:
## Codex Review (gpt-5.3-codex)

### High Priority
- auth.py:45 - Missing input validation on username parameter

### Medium Priority
- db.py:123 - N+1 query in user lookup loop

## Gemini Review (gemini-3.1-pro-preview)

Found potential issues:
1. The username field in auth.py is not validated...
2. Consider adding rate limiting to the login endpoint...

## Summary
Both reviews flagged the missing input validation in auth.py:45.
Codex also identified a performance issue in db.py:123.

Inline Arguments

User: /second-opinion check uncommitted changes for security issues

Claude: [scope: uncommitted, focus: security already specified]
Claude: [asks 2 questions: tool, project context]

User: codex only, yes include CLAUDE.md

Claude: [shows diff stats: 6 files, +103 -15]

Claude: [assembles prompt with CLAUDE.md + security focus + diff]

Claude: [runs codex exec]

Claude:
## Codex Review (gpt-5.3-codex)

### Critical
- api/auth.py:67 - SQL injection vulnerability in raw query

### High Priority
- api/auth.py:45 - Password stored in plaintext logs
- config.py:12 - Hardcoded API key

Large Diff Warning

User: /second-opinion

Claude: [asks questions]

User: both, uncommitted, general

Claude: [shows diff stats: 45 files, +3200 -890]

Claude: "Large diff (3200+ lines). This may take several minutes.
         Proceed, or narrow the scope (specific files, branch, commit)?"

User: "proceed"

Claude: [runs both reviews with 600s timeout]

Error Handling

Error	Action
codex: command not found	Show install: npm i -g @openai/codex
gemini: command not found	Show install: npm i -g @google/gemini-cli
Gemini extension missing	Show install: gemini extensions install <url>
Model auth error (Codex)	Retry with gpt-5.2-codex
Empty diff	Stop and inform user
Timeout	Inform user, suggest narrowing scope
Tool partially unavailable	Run only available tool, note the skip

Codex MCP Tools

This plugin bundles Codex CLI’s built-in MCP server (codex mcp-server), which auto-starts when the plugin is installed:

• codex - Start new Codex session with prompt, model, sandbox, approval policy
• codex-reply - Continue existing session by thread ID for multi-turn conversations

These tools work independently of the /second-opinion command. Use them for direct programmatic access to Codex.

Related Skills

• Modern Python - Reviews can check for modern Python best practices
• GH CLI - Can review branch diffs before opening PRs
• Git Cleanup - Can review branches before cleanup