second-opinion

name	second-opinion
description	Get multiple AI perspectives (Gemini + Codex) on a question. Use for architectural decisions, security reviews, or when you want validation from other AIs.

Second Opinion — Multi-AI Consultation

Query both Gemini and Codex with the same question to compare perspectives. Pass your question as $ARGUMENTS, optionally followed by file paths for context.

Usage

/second-opinion What's the best error handling strategy for this API?
/second-opinion Is this auth flow secure? server/auth.ts
/second-opinion Should we use WebSockets or SSE here? server/realtime.ts client/src/hooks/useStream.ts

Workflow

Step 1: Model Selection

Check configured models for both CLIs:

• Read ~/.gemini/settings.json for the Gemini model
• Read ~/.codex/config.toml for the Codex model

If either is missing, ask the user which frontier model to use for that CLI via AskUserQuestion. Reuse prior session choices if available.

Step 2: Context Gathering

Parse $ARGUMENTS — the question is the text, any paths at the end are context files. If file paths are provided, read them and prepare a context block (limit to ~200 lines per file, truncate if needed).

Format the prompt:

Question: [the question]

Context:
--- [filename] ---
[file contents]
---

Step 3: Query Both AIs

Query Gemini:

echo "FORMATTED_PROMPT" | gemini -m GEMINI_MODEL "Answer the question above based on the provided context" 2>/dev/null

Query Codex:

codex exec -m CODEX_MODEL "FORMATTED_PROMPT" -o /tmp/codex-second-opinion.txt 2>/dev/null; cat /tmp/codex-second-opinion.txt 2>/dev/null; rm -f /tmp/codex-second-opinion.txt

Step 4: Present Comparative Results

## Gemini's Take (model: X)
[response summary]

## Codex's Take (model: Y)
[response summary]

## My Synthesis
[where they agree, where they differ, and your own recommendation]

Rules

• Query both AIs even if one fails — present whatever you get
• If a CLI errors out, note it and continue with the other
• Keep context focused — don't include entire large files, extract the relevant sections
• Always add your own synthesis — don't just relay responses
• Do NOT use auto-approval flags on either CLI
• Clean up any temp files