second-opinion

Second Opinion

A Claude Code plugin that gets second opinions from Codex and Gemini with intelligent context sharing. Run multi-model code reviews, Q&A, and adversarial debates — all from within Claude Code.

Different models have different "attention" patterns. Running all three catches more than any individual review.

Why Multi-Model?

When Claude reviews your code alone, it has one perspective. With Second Opinion, you get three independent viewpoints that complement each other:

Model	Strength	Example Finding
Codex (GPT)	Domain logic, code tracing	"Cancel overwrites succeeded status to canceled"
Gemini	Systems thinking, concurrency	"Race condition in firstOrCreate for concurrent confirms"
Claude	Security hygiene, API design	"Leaks internal Stripe status string to client"

Each model finds things the others miss. The plugin runs them in parallel (no serial waiting) and presents a consensus table showing where they agree and disagree.

Commands

/second-opinion "question"

Get three independent answers to any question, presented as a comparison with synthesis.

/second-opinion "What's the best way to structure a Laravel module?"
/second-opinion "Should this function use early returns or a single return?"

What happens:

1. Claude assembles a context package (summarizes conversation, picks relevant files)
2. Spawns Codex and Gemini in parallel as background processes
3. Does its own analysis independently
4. Reads all outputs and presents a comparison + synthesized answer

/review-models [--uncommitted | --base branch]

Get code review from all three models simultaneously.

/review-models                          # Review uncommitted changes
/review-models --uncommitted            # Same as above
/review-models --base main              # Review branch changes vs main
/review-models --base <commit-sha>^     # Review from a specific commit

Output includes:

• Individual review summaries from each model
• A consensus table showing which issues each model found
• Severity counts with agreement levels
• Prioritized fix recommendations

/debate "topic"

Get adversarial perspectives — one model argues FOR, the other AGAINST. Claude moderates.

/debate "Should we use Inertia.js or a separate SPA for the dashboard?"
/debate "Monorepo vs polyrepo for our microservices"
/debate "Server-side rendering vs client-side rendering for this app"

What happens:

1. Claude frames the topic as a binary decision
2. Sends Codex the "argue FOR" framing, Gemini gets "argue AGAINST"
3. Both run in parallel
4. Claude moderates: argument analysis, weaknesses exposed, verdict with caveats

Real-World Example

Here's an actual review of a Stripe Terminal Tap to Pay feature (5 commits, 8 files, ~968 lines):

/review-models --base 20b9e5174^

What Each Model Found

Codex traced the code execution path and found:

• Cancel endpoint overwrites succeeded status to canceled unconditionally (critical bug — corrupts payment state)

Gemini analyzed concurrency patterns and found:

• Race condition in firstOrCreate for concurrent confirm calls (critical)
• Race condition in Terminal location creation (warning)
• Missing test coverage for null branches in extractGeneratedCardId
• Redundant re-query in catch block

Claude reviewed security and API design:

• Same cancel status bug as Codex (independently confirmed)
• Leaking internal Stripe status string in error messages
• Empty no-op if block in SetupIntentService
• Missing tests for TerminalController

Consensus Table

| Issue                               | Severity | Codex | Gemini | Claude |
|-------------------------------------|----------|-------|--------|--------|
| Cancel overwrites succeeded status  | critical | Found | Missed | Found  |
| Race condition in firstOrCreate     | critical | Missed| Found  | Missed |
| Terminal location race condition    | warning  | Missed| Found  | Found  |
| Leaked Stripe status in error msg   | warning  | Missed| Missed | Found  |
| Empty no-op if block                | warning  | Missed| Missed | Found  |
| Redundant Person query in txn       | warning  | Missed| Found  | Missed |
| Redundant re-query in catch block   | nit      | Missed| Found  | Found  |

Result: 2 critical bugs, 5 warnings, 2 nits — no single model found all of them.

How It Works

Claude (full context) → assembles context package → pipes to CLI via stdin
                                                  → both run in parallel
                                                  → reads outputs
                                                  → synthesizes comparison