search

/dual-search — Cross-Model Fact Verification

Runs the same factual question through Claude (WebSearch) and Codex CLI in parallel, then compares results. If they disagree, each model researches the disagreement point and the loop repeats until convergence or max rounds.

When to Use

• Factual claims that could be hallucinated ("X is open source", "Y supports Z")
• Licensing, version, compatibility questions
• Any statement where being wrong has real consequences

Input

The user provides a factual question or claim to verify.

/dual-search "Claude Code는 오픈소스인가?"
/dual-search "ripgrep은 Go로 만들어졌나?"

If no argument is provided, use AskUserQuestion to ask what to verify.

Execution Flow

Round 1: Parallel Search

Launch both searches simultaneously in a single message with two tool calls:

Claude Search — spawn an Agent:

Agent(subagent_type="general-purpose", prompt="""
Search the web for: {QUESTION}
Find at least 3 sources. For each source, note:
- URL
- What it says about the question
- Whether it confirms or denies

Return a structured answer:
ANSWER: [your conclusion]
CONFIDENCE: [HIGH/MEDIUM/LOW]
SOURCES:
- [url1]: [what it says]
- [url2]: [what it says]
- [url3]: [what it says]
""")

Codex Search — run via Bash:

codex exec "Search the web and answer this question with sources: {QUESTION}. Return your answer as: ANSWER: [conclusion], CONFIDENCE: [HIGH/MEDIUM/LOW], SOURCES: [list urls and what each says]" 2>&1

Round 2+: Convergence Loop (if disagreement)

Compare the two results. If they disagree:

1. Identify the specific point of disagreement
2. Formulate a narrower, more specific follow-up question targeting the disagreement
3. Run both searches again with the narrowed question
4. Compare again

Example:

Round 1:
  Claude: "Claude Code is open source (MIT)"
  Codex:  "Claude Code is proprietary (All Rights Reserved)"
  → DISAGREE on: license type

Round 2 (narrowed question):
  "What is the exact text of LICENSE.md in the anthropics/claude-code GitHub repository?"
  Claude: "© Anthropic PBC. All rights reserved."
  Codex:  "All Rights Reserved, not OSI-approved"
  → AGREE: proprietary

Circuit Breaker

• Max 3 rounds. If still disagreeing after 3 rounds, report both positions and let user decide.
• Each round narrows the question further — never repeat the same question.

Output Format

## Dual Search Result

**Question**: {original question}
**Verdict**: ✅ AGREE / ⚠️ DISAGREE (after {N} rounds)

### Claude (WebSearch)
**Answer**: ...
**Confidence**: HIGH/MEDIUM/LOW
**Sources**:
- [url]: summary
- [url]: summary

### Codex (GPT)
**Answer**: ...
**Confidence**: HIGH/MEDIUM/LOW
**Sources**:
- [url]: summary
- [url]: summary

### Convergence Analysis
- **Agreed on**: ...
- **Disagreed on**: ... (if any)
- **Resolution**: ... (how disagreement was resolved, or "unresolved — user judgment needed")

### Rounds
- Round 1: {question} → {agree/disagree on what}
- Round 2: {narrowed question} → {agree/disagree on what}
...

Rules

• NEVER skip the Codex search. Both must run.
• NEVER fabricate Codex results. If Codex call fails, report the failure.
• Each round's question must be MORE SPECIFIC than the previous.
• Always show sources from both sides.
• Present disagreements honestly — don't force agreement.
• Korean output for explanations, English for technical terms.