claude-code - 💡(How to fix) Fix AskUserQuestion auto-resolves with empty answers in headless/no-TTY mode (Python Agent SDK) [2 comments, 3 participants]

Description

When using the Python Agent SDK (claude-agent-sdk) in a headless/no-TTY environment (e.g., Docker container, CI, server-side automation), AskUserQuestion auto-resolves immediately with empty answers:

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

The tool completes in ~37ms without any opportunity for the integration to collect answers from the user, even when a can_use_tool callback or PreToolUse hook is registered.

Expected behavior

In headless/no-TTY environments, AskUserQuestion should either:

1. Await the can_use_tool callback before resolving, allowing the integration to collect answers via its own UI and return them in updated_input
2. Respect PreToolUse hooks that return continue_: False by not executing the tool at all (currently the tool executes before the stop signal takes effect)

The Agent SDK docs describe option 1 as the intended flow, but it doesn't work in practice.

Actual behavior

1. Claude calls AskUserQuestion(questions=[...])
2. The CLI detects no TTY is available
3. The tool immediately resolves with empty answers — before any callback or hook can intervene
4. Claude continues with the empty answers as if the user responded

Workaround

The only reliable workaround is returning permissionDecision: "deny" from a PreToolUse hook, which prevents the tool from executing entirely. The question event can be emitted separately (e.g., via PubSub), and the user's answer is delivered as a new prompt in a subsequent turn.

async def pre_tool_use_hook(self, input, tool_use_id, context):
    if input and input.get("tool_name") == "AskUserQuestion":
        return {
            "continue_": False,
            "stopReason": "Waiting for user to answer the question.",
            "hookSpecificOutput": {
                "hookEventName": "PreToolUse",
                "permissionDecision": "deny",
                "permissionDecisionReason": "Questions sent to user. Waiting for response.",
            },
        }
    return {"hookEventName": "PreToolUse"}

This works but loses the tool's conversational context — the answer arrives as a fresh prompt rather than a tool result, so Claude doesn't always associate it with the original question.

Environment

• claude-agent-sdk==0.1.63 (Python)
• Bundled CLI: @anthropic-ai/[email protected]
• Running in Docker (no TTY attached)
• Python 3.12

Related issues

• #30983 — canUseTool callback never awaited (closed as not planned)
• #29530 — Empty response without rendering UI
• #47114 — v2.1.104 regression, auto-resolves even in interactive CLI
• #29618 — Auto-resolves under acceptEdits permission mode
• #16712 — No way to provide answers via stdin when resuming a session
• #34592 — Unavailable in sub-agent contexts
• anthropics/claude-agent-sdk-python#327 — Can't trigger AskUserQuestion in Python SDK