Ask Codex

Delegate a question or review request to Cursor's agent CLI (agent -p --approve-mcps) configured with a codex model. Gather relevant context from the current conversation and codebase, construct a focused prompt, execute via CLI, and present the response.

Core Concept

User Request ──> Gather Context ──> Build Prompt ──> agent -p ──> Present Response
                 (files, docs,      (context +        (stdout)
                  conversation)      question)

Workflow

1. Parse User Intent

Extract from the user's message:

• Question: What to ask codex (may be explicit or implied from conversation context)
• Target files: Any specific files or code the user wants codex to review
• Scope: Whether this is about a specific file, a feature, the whole project, or a general question

If the user's intent is unclear, ask for clarification using AskUserQuestion:

question: "codex에게 무엇을 물어볼까요?"
header: "Question"
options:
  - label: "현재 작업 리뷰"
    description: "지금 작업 중인 코드/파일에 대한 리뷰 요청"
  - label: "아키텍처 의견"
    description: "프로젝트 구조나 설계에 대한 의견 요청"
  - label: "직접 질문 작성"
    description: "codex에게 보낼 질문을 직접 입력"
multiSelect: false

2. Gather Context

Based on the parsed intent, collect relevant context. Minimize context to only what is necessary for the question.

Context sources (select relevant ones):

Source	When to Include
Specific files	User mentions a file or code block
Feature documents	Current session involves a feature workflow (check .workflow-adapter/doc/feature_*/)
Git diff	User asks about recent changes
Error output	User asks about an error or issue
Conversation summary	User asks a general question about current discussion

For feature workflow context, check for these files in .workflow-adapter/doc/feature_{name}/:

• context.md - Research context
• brainstorming.md - Brainstorming results
• spec.md - Feature specification
• plan.md - Implementation plan

Read only the documents relevant to the question. Summarize large documents rather than including them verbatim.

3. Build Prompt

Construct a focused prompt combining the context and question:

## Context

{gathered context - files, docs, conversation summary}

## Question

{the specific question or request for codex}

Prompt guidelines:

• Keep total prompt under 8,000 characters to avoid CLI issues
• Summarize large files rather than including full content
• Include file paths and line numbers when referencing code
• Frame the question clearly and specifically
• Write the prompt in the same language the user is using

4. Execute Agent CLI

Run the agent command via Bash with a 120-second timeout:

agent -p --approve-mcps "{prompt}"

Note: --approve-mcps is required to auto-approve MCP servers in headless/print mode.

Additional options:

• --workspace {path}: Set workspace if the question is about a specific project directory
• --model {model}: Override model if user requests a specific one (default uses configured codex model)

If the prompt is too long for a single command argument, write it to a temporary file and pipe it:

cat /tmp/codex-prompt.txt | xargs -0 agent -p --approve-mcps

Clean up the temporary file after the command completes.

Error handling:

• If agent command is not found, inform the user to install Cursor CLI
• If authentication fails, suggest running agent login
• If the command times out, suggest simplifying the question or reducing context

5. Present Response

Display codex's response with clear attribution:

**Codex Response:**

{response from agent CLI}

After presenting, ask if follow-up is needed:

question: "codex 응답에 대해 추가 작업이 필요한가요?"
header: "Follow-up"
options:
  - label: "충분해요"
    description: "추가 작업 없이 종료"
  - label: "추가 질문"
    description: "codex에게 후속 질문 전달"
  - label: "반영하기"
    description: "codex 의견을 현재 작업에 반영"
multiSelect: false

If the user wants a follow-up question, repeat from step 3 with the previous response included as additional context.

Constraints

• Write prompts in the same language the conversation is conducted in
• Keep prompts concise and focused - avoid dumping entire files when a summary suffices
• Set Bash timeout to 120000ms for agent CLI execution
• Do not expose API keys or sensitive information in prompts
• If codex suggests code changes, present them as suggestions - do not auto-apply without user confirmation
• Clearly attribute all responses as coming from codex to avoid confusion