/discuss-with-codex

Discuss with Codex

Have a multi-turn discussion with Codex CLI to brainstorm, debate, and converge on a solution. Use this during planning phases when you need another AI brain to hash out ideas.

Communicate with the user EXCLUSIVELY in 繁體中文 (台灣用語). Communicate with Codex in English.

The discussion topic comes from $ARGUMENTS. If empty, ask the user what they want to discuss.

STEP 1: Frame the Discussion

Read relevant files if the topic references specific code or architecture. Build enough context to have an informed discussion.

Formulate an opening prompt for Codex. The prompt should:

• Clearly state the problem or question
• Provide relevant context (code snippets, architecture constraints, trade-offs)
• Ask Codex for its analysis, opinion, or proposed approach
• Be specific enough to get a substantive response

STEP 2: Open the Discussion

Send the opening prompt to Codex:

codex exec "TOPIC_CONTEXT_AND_QUESTION" --json 2>/dev/null | jq -rs '
  "=== CODEX REASONING ===\n"
  + ([.[] | select(.item.type == "reasoning") | .item.text] | unique | join("\n"))
  + "\n\n=== CODEX RESPONSE ===\n"
  + ([.[] | select(.item.type == "agent_message")] | last | .item.text)
'

Use timeout of 300000ms.

This creates a resumable session for follow-up rounds.

STEP 3: Analyze and Respond

After receiving Codex's response:

1. Analyze what Codex said — identify key points, agreements, disagreements, gaps, and novel insights
2. Think critically — where is Codex right? Where might it be wrong? What did it miss? What's a stronger argument?
3. Formulate a follow-up that:
   • Acknowledges valid points
   • Challenges weak reasoning with specific counter-arguments
   • Raises new angles or considerations Codex missed
   • Pushes toward convergence if ideas are aligning
   • Asks targeted follow-up questions to dig deeper

STEP 4: Continue the Discussion

Resume the Codex session with your follow-up:

codex exec resume --last "YOUR_FOLLOW_UP_RESPONSE" --json 2>/dev/null | jq -rs '
  "=== CODEX REASONING ===\n"
  + ([.[] | select(.item.type == "reasoning") | .item.text] | unique | join("\n"))
  + "\n\n=== CODEX RESPONSE ===\n"
  + ([.[] | select(.item.type == "agent_message")] | last | .item.text)
'

Use timeout of 300000ms.

STEP 5: Convergence Check

After each round, assess whether the discussion has converged:

• Converged: Both AIs largely agree on the approach, key trade-offs are identified, a clear recommendation has emerged → proceed to Step 6
• Not converged: Significant disagreements remain, new important angles surfaced, or more depth is needed → go back to Step 3

Target 2-5 rounds total. After 5 rounds, force convergence — synthesize the best ideas from both sides even if disagreements remain.

STEP 6: Present Conclusions

Present the discussion results to the user in 繁體中文:

## 🧠 討論結論

### 討論主題
[原始問題/主題]

### 討論過程摘要
- 共進行 N 輪討論
- [1-2 句話描述討論如何演進]

### 共識
[Claude 和 Codex 都同意的觀點和方案]

### 分歧
[兩方有不同看法的地方，以及各自的理由]

### 建議方案
[綜合兩方意見後的最終建議，包含：]
- **方案描述**: [具體做法]
- **為什麼選這個方案**: [關鍵理由]
- **需要注意的風險/取捨**: [重要的 trade-offs]

### 替代方案
[如果有值得考慮的替代方案，列出來]

After presenting, ask the user if they want to:

• Proceed with the recommended approach
• Dig deeper on a specific point (restart discussion on that sub-topic)
• Take a different direction entirely

CRITICAL RULES

1. ALL user-facing output MUST be in 繁體中文 (台灣用語)
2. ALL prompts to Codex MUST be in English
3. Think CRITICALLY about Codex's responses — don't just agree. Challenge weak points, propose alternatives, push for better answers.
4. Codex: ALWAYS use --json 2>/dev/null | jq ... to extract output. Never use plain 2>/dev/null alone.
5. Follow-ups: ALWAYS use codex exec resume --last to maintain conversation context.
6. Codex JSONL extraction: ALWAYS extract BOTH agent_message (final response) and reasoning (thinking summaries, deduplicated via unique).
7. Use timeout of 300000ms (5 minutes) for ALL Codex calls.
8. Do NOT just relay what Codex says. YOU are an active participant — analyze, critique, build on ideas.
9. Keep the discussion focused. If it drifts, steer it back to the core question.
10. If Codex gives a shallow or evasive answer, push harder with specific follow-up questions.