Your agent casually asked a question and froze the queue for 10 minutes — how do you all handle ask_user_question?
A real pain point. I'll share my solution and the traps I hit, but I really want to hear how the community actually deals with it.
Quick one for everyone: do your agents randomly stop to ask you questions? And what happens to the running task when they do?
Here's what happened to me: DeepSeek Harness's agent calls ask_user_question mid-task — "should this module be sync or async?" "OK to run this high-risk command?" Sounds reasonable. But it deadlocked my entire task queue for 10 minutes.
The problem: a question = a silent deadlock
The worst part: it doesn't error. The task just hangs there, pending keeps climbing, you think it's thinking — but it's actually waiting for a human answer that will never come. Because the question goes through the question channel (question/requested), which is completely separate from the approval channel (approval/requested).
My first relay only listened to approvals, so:
Agent stops on the computer, waiting for a user
↓
question-channel messages silently dropped by my relay
↓
phone has no idea
↓
task hangs forever until timeout
It took me two days to realize: it wasn't DSH hanging — the question just never reached a human.
My fix: turn questions into "your answer needed" cards
While building DeepSeek Phone Harness (remote-control your desktop agent from your phone), I made questions a card that drops right into the message stream:
- Agent asks → phone shows "your answer needed" card
- Pick an option / type a custom answer / skip
- Answer it and the task continues; ignore it and it waits
Protocol has to match the official Web GUI exactly ({ok:true, value:{sessionId, answer:{answers:[{id, selected}]}}}), or DSH won't accept it.
What I'd really like to hear
My fix only handles the "phone side." I want the broader community's practice:
- Do you let agents ask you questions at all, or disable
ask_user_question? - Granularity: which questions are worth interrupting a human for, and which should the agent just decide?
- Async handling: when a question blocks a task, what does your queue do — block, skip, or run around it?
- Has anyone set a "question budget" per task, or use an alternative tool?
Honestly my current stance is conservative (ask only if it must, always interrupt). But I suspect that's not optimal — "when to interrupt a human" is probably the deeper question than "how to interrupt."
Would love to hear your practices below. And if you've had an agent-question deadlock too, drop a comment so I can gauge how common this actually is.
Implementation is open source (MIT): https://github.com/2903077918-lgtm/DeepSeek-phone-harness
