Reviewing Skill

Multi-model second opinions via OpenAI Codex CLI and Google Gemini CLI.

When to Use

• Validating critical findings (security, correctness, soundness)
• Breaking ties when uncertain
• Cross-checking complex reasoning
• Architecture decisions with high stakes

Don't use for: Simple questions. Just think harder.

Priority Order

Codex (OpenAI)

OpenAI's CLI for code and reasoning tasks.

Invocation

codex exec -s read-only -m gpt-5.2 -c reasoning=xhigh "<prompt>"

Models

Reasoning Levels

Choose the right level for the task. Higher reasoning = more time.

Guidance:

• Start with high for most review tasks—it's the sweet spot
• Escalate to xhigh only when correctness is critical or reasoning is complex
• Use medium or default for simple questions where speed matters
• If xhigh times out or you're waiting too long, consider whether high suffices

# Most reviews
codex exec -s read-only -m gpt-5.2 -c reasoning=high "..."

# Critical/complex only
codex exec -s read-only -m gpt-5.2 -c reasoning=xhigh "..."

# Quick sanity check
codex exec -s read-only -m gpt-5.2 "..."

Gemini (Google)

Google's CLI for research and long-context analysis.

Invocation

gemini -s -m gemini-3-pro-preview "<prompt>"

Models

Strengths

• Long context windows (large codebase analysis)
• Research and fact-checking
• Tie-breaking between Claude and Codex

Prompt Pattern

Always request a ---SUMMARY--- marker for reliable extraction:

codex exec -s read-only -m gpt-5.2 -c reasoning=xhigh "
<context and question>

End with:
---SUMMARY---
AGREE/DISAGREE with <topic>
Confidence: HIGH/MEDIUM/LOW
Key insight: <one sentence>
"

Output Extraction

# Run and capture
codex exec -s read-only -m gpt-5.2 -c reasoning=xhigh "..." > /tmp/opinion.log 2>&1

# Extract summary
sed -n '/---SUMMARY---/,$ p' /tmp/opinion.log

Common Patterns

Second Opinion

codex exec -s read-only -m gpt-5.2 -c reasoning=xhigh "
Problem: <describe>
Current hypothesis: <what you think>

Do you agree? What would you change?

---SUMMARY---
AGREE/DISAGREE
Confidence: HIGH/MEDIUM/LOW
Key insight: <summary>
"

Tie-Breaker (Gemini)

gemini -s -m gemini-3-pro-preview "
Analysis A (Claude): <view>
Analysis B (Codex): <view>

They disagree on: <the issue>

Which is correct?

---SUMMARY---
FAVOR: A/B/NEITHER
Reasoning: <why>
"

Multi-Model Consensus

For high-stakes decisions, query both:

# Get Codex opinion
codex exec -s read-only -m gpt-5.2 -c reasoning=xhigh "..." > /tmp/codex.log 2>&1

# Get Gemini opinion
gemini -s -m gemini-3-pro-preview "..." > /tmp/gemini.log 2>&1

# Compare summaries
sed -n '/---SUMMARY---/,$ p' /tmp/codex.log
sed -n '/---SUMMARY---/,$ p' /tmp/gemini.log

Recording Findings

Post to jwz for discovery:

# Single model
jwz post "issue:<id>" --role alice \
  -m "[alice] SECOND_OPINION: codex on <topic>
Model: gpt-5.2 (reasoning=xhigh)
Agreement: AGREE|DISAGREE|PARTIAL
Key insight: <summary>"

# Multi-model consensus
jwz post "issue:<id>" --role alice \
  -m "[alice] CONSENSUS: <topic>
Models: codex, gemini
Agreement: FULL|PARTIAL|SPLIT
Synthesis: <reconciled view>"

Timeouts

# Codex with xhigh reasoning can take time
timeout 600 codex exec -s read-only -m gpt-5.2 -c reasoning=xhigh "..."

# Gemini
timeout 600 gemini -s -m gemini-3-pro-preview "..."

Fallback Chain

if command -v codex >/dev/null 2>&1; then
    codex exec -s read-only -m gpt-5.2 -c reasoning=xhigh "..."
elif command -v gemini >/dev/null 2>&1; then
    gemini -s -m gemini-3-pro-preview "..."
else
    claude -p "..."
fi