---

name: ask-codex description: Consult Codex for questions, feedback, reviews, or discussions. Use when Claude needs a second opinion, wants to validate an approach, or the user explicitly calls /ask-codex.

Ask Codex

Overview

Consult Codex for questions, feedback, reviews, or discussions. The conversation iterates until Codex signals completion (e.g., "LGTM" for reviews, or a complete answer for questions) or until 10 rounds are reached.

Consultation Modes

Workflow

1. Determine the consultation mode

   • If the user specifies a mode, use it.
   • If the user asks a question or requests discussion, use question or discussion mode.
   • If auto-invoked on plan mode exit and a plan file matching ~/.claude/plans/plan-*.md exists, default to plan review (even if there are uncommitted changes).
   • If there are uncommitted changes (git status --porcelain is non-empty), default to change review.
   • Otherwise, ask the user what they want to consult about.

2. Collect the content

   • Question: Gather the user's question and any relevant context (code snippets, error messages, etc.).
   • Discussion: Gather the topic and relevant context for exploration.
   • Plan review: Read from ~/.claude/plans/plan-*.md (use the most recently modified if multiple exist) or ask the user for the plan.
   • Change review: Collect both tracked and untracked changes:

     Copy# Tracked changes (staged + unstaged)
     # Use fallback for repos without commits
     if git rev-parse --verify HEAD >/dev/null 2>&1; then
       git diff HEAD
     else
       git diff --cached
       git diff
     fi
     
     # Untracked files - safe handling for special characters
     # Skip binary files and files larger than 100KB
     git ls-files -z --others --exclude-standard | while IFS= read -r -d '' f; do
       # Skip binary files (use -- to handle filenames starting with -)
       if file -- "$f" | grep -q 'text'; then
         # Skip files larger than 100KB
         size=$(stat -f%z -- "$f" 2>/dev/null || stat -c%s -- "$f" 2>/dev/null)
         if [ "$size" -lt 102400 ]; then
           echo "=== New file: $f ==="
           cat -- "$f"
         else
           echo "=== New file: $f (skipped: >100KB) ==="
         fi
       else
         echo "=== New file: $f (skipped: binary) ==="
       fi
     done
     

   • Custom: Ask the user to provide the content.
   • Also collect the original user prompt (task background) from the conversation.

3. Prepare the plans directory

   • Run mkdir -p ~/.claude/plans to ensure the directory exists.

4. Write the consultation input file

   • Create a file at ~/.claude/plans/consultation-input.md with this format:

     Copy## Original Request
     <original user prompt>
     
     ## Consultation Mode
     <question | discussion | plan review | change review | custom>
     
     ## Content
     <question, topic, plan, diff + untracked files, or custom content>
     
     ## Task
     <varies by mode>
     - Question: Answer the question. When complete, output "ANSWERED" on its own line.
     - Discussion: Explore the topic. When a conclusion or actionable insight is reached, output "CONCLUDED" on its own line.
     - Plan review / Change review: Review the content. If there are problems, describe them. If OK, output "LGTM" on its own line.
     - Custom: Address the request. When complete, output "DONE" on its own line.
     

5. Set up the temp environment

   • Prepare a temp directory under ~/.claude/tmp (same as codex-review):
     • Claude Code runs each shell command in a fresh process, so export does not persist across commands.
     • Always run the temp setup and codex exec in the same shell invocation.

     Copymkdir -p ~/.claude/tmp
     TMPDIR="$(mktemp -d ~/.claude/tmp/ask-codex.XXXXXXXX)"
     ZDOTDIR="$TMPDIR/zsh"
     mkdir -p "$ZDOTDIR"
     XCRUN_CACHE_PATH="$TMPDIR/xcrun_db"
     DARWIN_USER_TEMP_DIR="$TMPDIR"
     export TMPDIR ZDOTDIR XCRUN_CACHE_PATH DARWIN_USER_TEMP_DIR
     

6. Run Codex

   • Execute the temp setup and codex exec in the same shell invocation:

     Copymkdir -p ~/.claude/tmp && \
     TMPDIR="$(mktemp -d ~/.claude/tmp/ask-codex.XXXXXXXX)" && \
     ZDOTDIR="$TMPDIR/zsh" && \
     mkdir -p "$ZDOTDIR" && \
     XCRUN_CACHE_PATH="$TMPDIR/xcrun_db" && \
     DARWIN_USER_TEMP_DIR="$TMPDIR" && \
     export TMPDIR ZDOTDIR XCRUN_CACHE_PATH DARWIN_USER_TEMP_DIR && \
     cat ~/.claude/plans/consultation-input.md | codex exec --sandbox read-only -
     

   • Timeout: 5 minutes (300 seconds).

7. Parse the result

   • Trim whitespace from the output.
   • Check if any line matches a completion signal (case-insensitive):
     • "LGTM" → approved (for reviews)
     • "ANSWERED" → answered (for questions)
     • "CONCLUDED" → concluded (for discussions)
     • "DONE" → done (for custom)
   • If matched, mark as responder complete.
   • Otherwise, extract the response as follow-up needed from responder.

8. Evaluate the response (questioner's perspective)

   • After receiving Codex's response, Claude evaluates:
     • Does the response fully address the original question/concern?
     • Are there unclear points or ambiguities that need clarification?
     • Did new questions arise from the response?
   • Determine questioner status:
     • Questioner satisfied: No remaining questions or concerns.
     • Questioner has follow-ups: Additional questions or clarifications needed.

9. Iteration loop

   • Continue iteration if ANY of the following:
     • Responder signaled follow-up needed (no completion signal).
     • Questioner (Claude) has remaining questions or unclear points.
   • Stop iteration if BOTH:
     • Responder signaled complete (LGTM/ANSWERED/CONCLUDED/DONE), AND
     • Questioner (Claude) is satisfied (no remaining questions).
   • If continuing:
     • Report the response to the user.
     • If no completion signal was returned, explicitly ask for the completion signal (LGTM/ANSWERED/CONCLUDED/DONE) in the next reply.
     • For question/discussion: Formulate specific follow-up questions based on unclear points or new questions that arose.
     • For change review: Apply fixes to the code, then re-collect the diff and untracked files.
     • For plan review: Adjust the plan based on feedback, or ask for clarification on unclear feedback.
     • Update the consultation input with follow-up questions/context.
     • Return to step 4 and resubmit.
   • Maximum 10 iterations.

10. Finish

    • Complete: Report the full response content to the user (answer, conclusion, or review result), then note the completion status and proceed.
    • Not complete after 10 rounds: Report the current state and ask the user how to proceed.

Auto-invocation

This skill is automatically invoked when:

• Claude Code exits plan mode (cross-review before execution).
• Claude Code completes any task with code changes (cross-review before proceeding).
• The user has not explicitly skipped the consultation.

Note: For ad-hoc consultations at any point, use /ask-codex manually.

Notes

• Use --sandbox read-only since consultation does not require code modifications by Codex.
• Honor explicit user requests to skip the consultation.
• Plan files use the naming convention plan-*.md to distinguish from consultation-input.md.
• The consultation input file at ~/.claude/plans/consultation-input.md persists for debugging and history.
• For change review, both git diff and untracked file contents are collected to ensure complete coverage.
• Completion signal detection uses exact line match (after trimming) to avoid false positives like "Not LGTM".
• Binary files and files larger than 100KB are skipped to avoid input bloat.