app

Adom Hydrogen

Public Made by Adom

by adom

The whole Adom experience as one signed native app: your editor, your AI agent, and the full electronics design toolchain, running locally on your own machine.

Overview Files Skills Discover Packages Releases Dependencies Issues20 Breadcrumbs Pull Requests Activity
Open general

run-prompt never reports needs_you while a Claude AskUserQuestion is waiting; no API to answer it

John Lauer · 10d ago

What happened

On Adom Hydrogen v1.0.601 (ConfRoomROG), a remote AI started a long filmed build with POST /ai-threads/run-prompt {new:true, type:"claude", background:true} and polled GET /ai-threads/run-prompt/status?job=... every minute.

About 2 hours in, the Claude thread asked an AskUserQuestion: "Live stages: which machine, and when?", with four radio options and a Submit button. The job stayed working / still_working the whole time and never reached needs_you. GET /claude/conversation-state doesn't show it either: answering:true, and there's no field for a pending question. The run sat waiting for an unknown time until a screenshot happened to show the question.

The skill says needs_you means "a permission or question is waiting". A question from the AskUserQuestion tool doesn't trigger it.

What a caller AI needs

  1. needs_you for AskUserQuestion, using the PreToolUse hook on AskUserQuestion or the same signal the permission hook uses. Also send question: {header, text, options[]} in the status so the caller can show it or answer it.
  2. A way to answer it through the API, e.g. POST /ai-threads/answer {sid|job, option: <index|label>}. Without it, the only way is OS clicks at screenshot coordinates. That went wrong once here: the shotId mapped to the 1400 px safe image, not the full one, and the first click picked the wrong radio button.
  3. /claude/conversation-state: add pending_question (and pending_permission) so a watcher that doesn't own the job can see a blocked thread.
  4. Add one line to the hydrogen-ai-threads-api skill under "Know where your prompt went" on how to detect and answer a question.

The workaround I'm using for now: treat 4 minutes with no change in msgs as "probably blocked", then screenshot.

2 Replies

John Lauer · 9d ago

Items 1, 3 and 4 are in Hydrogen 1.0.607 (insiders). Item 2, answering through the API, is not built yet, so this stays open for it.

How: Hydrogen now also installs PreToolUse and PostToolUse hooks matched to AskUserQuestion (the merge runs on every launch, so existing workspaces get them with no action).

  • 1. The job goes to needs_you (terminal, so a poller stops) and carries job.question: {questions: [{header, text, options: [{label, description}], multi_select}], at_ms}. Once answered it returns to working, and question is gone.
  • 3. GET /claude/conversation-state has pending_questions: session id to the same object, for watchers that do not own the job.
  • 4. The hydrogen-ai-threads-api skill explains it under "Know where your prompt went".
  • The hook payload cap went from 4 KB to 16 KB, because a question with four described options outgrew 4 KB and a cut payload does not parse.

Verified on a test PC, 1.0.607, with run-prompt {new:true, type:"claude", submit:true} and a prompt asking Claude to ask "Which color do you prefer?" (Red / Blue):

step state job.question pending_questions
Claude asks needs_you (terminal) header "Color", text, both options with descriptions the same, under the session id
answered "Red" in the tab working gone
turn ends done gone {}

The release sweep passed on that PC (217/217).

John Lauer · 3d ago

Item 2 is written, in 6b584ebd for the next build after 1.0.658:

POST /ai-threads/answer {job | sid | tid, option} answers the question the way a person does, inside that thread's own tab: it selects the options, then presses Submit answers, and answers ok only once the question is gone.

  • option: a label (exact, else a unique prefix) or a 0-based index.
  • options: [...] for a multi-select question, other: "text" for a typed answer.
  • answers: [{question: <index or header>, option | options | other}, ...] when Claude asks several questions at once.
  • With no job/sid/tid it answers the one question showing anywhere, and refuses if several threads are asking.
  • A wrong label returns ok:false with the real options, so a caller can correct itself instead of clicking blind. answered lists what was picked.

No screenshots or OS clicks: it works on a backgrounded Hydrogen. The skill's "A question is waiting" section now shows it.

Gate entry ai-threads-answer: run-prompt asks Claude to ask Red / Blue; at needs_you, answer {job, option:"Blue"} is ok, the job returns to working, and Claude's reply names Blue. When that passes on the build, everything in this issue is done and I will close it.

Log in to reply.