AskUserQuestion tool returns 'completed' without displaying prompt to user

Bug Description

The AskUserQuestion tool silently completes without displaying the question prompt to the user. The tool returns "User has answered your questions: . You can now continue with the user's answers in mind." with no actual user selection — because the user never saw the question.

Steps to Reproduce

1. In a Claude Code CLI session (Opus 4.6), invoke a skill that uses AskUserQuestion
2. The tool call is made with valid parameters (question, header, options array, multiSelect)
3. Expected: User sees a prompt with selectable options in the terminal
4. Actual: No prompt displays. The tool immediately returns as "completed" with an empty answer.

Impact

• Severity: High — Any skill relying on AskUserQuestion for user input silently fails
• In our case, 15 consecutive AskUserQuestion calls returned empty across a single session
• The assistant had no way to detect the failure (return format looks identical to a successful call)
• User confirmed: "nothing showed" on their end for any of the 15 prompts

Example Tool Call (one of 15 that failed)

{
  "questions": [{
    "question": "How would you like to handle these 14 captures?",
    "header": "Triage",
    "options": [
      {"label": "Approve all", "description": "Route everything as shown"},
      {"label": "Go one-by-one", "description": "Review each item individually"},
      {"label": "Modify", "description": "Change specific classifications"},
      {"label": "Skip all", "description": "Don't process any captures right now"}
    ],
    "multiSelect": false
  }]
}

Response Received (for all 15 calls)

User has answered your questions: . You can now continue with the user's answers in mind.

Note the empty space between "questions:" and "." — no answer content.

Environment

• Claude Code CLI (terminal, not web)
• Model: claude-opus-4-6
• OS: macOS Darwin 24.6.0
• Session type: Interactive (not headless)

Suggested Improvement

The return format for AskUserQuestion should make it distinguishable when the user actually selected an option vs when the tool completed without user interaction. Currently there is no way for the assistant to detect this failure mode.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>