An agent's status question stops an auto-approved run for a person #280

Closed
opened 2026-10-28 16:16:04 -04:00 by cmoriarty · 2 comments
Owner

What happens

Run 134 (vLLM, auto-approve on) stopped at "Needs you" in the judge. While its browser subagent ran in the background, the judge called the question tool as a "status checkpoint": "The file:// browser check is still running in the background… No action needed." Its one option was "Continue when ready". Braid opened it as a gate with expire_action: hold and no expiry, so the run waits for a person indefinitely. Auto-approve doesn't cover agent questions.

Fix

  • An agent question with a single option, or a recommended default, is answered with it on an auto-approved run, after a short pause, and the answer is logged.
  • The judge's brief says to wait for a background subagent by ending its turn, never with question.

Done when

An auto-approved run whose agent asks a one-option or defaulted question carries on without a person.

From #279, run 134.

**What happens** Run 134 (vLLM, auto-approve on) stopped at "Needs you" in the judge. While its `browser` subagent ran in the background, the judge called the `question` tool as a "status checkpoint": *"The file:// browser check is still running in the background… No action needed."* Its one option was "Continue when ready". Braid opened it as a gate with `expire_action: hold` and no expiry, so the run waits for a person indefinitely. Auto-approve doesn't cover agent questions. **Fix** - An agent question with a single option, or a recommended default, is answered with it on an auto-approved run, after a short pause, and the answer is logged. - The judge's brief says to wait for a background subagent by ending its turn, never with `question`. **Done when** An auto-approved run whose agent asks a one-option or defaulted question carries on without a person. _From #279, run 134._
Author
Owner

It loops. Run 134's judge asked four of these in about four minutes ("Judge progress", "Waiting on browser run", "Awaiting browser report", "Still awaiting report"), each with one option, "Continue when ready". Each answer woke it; with its browser subagent still running it had nothing to do, so it asked again.

So answering automatically isn't enough on its own: it would only make the loop faster. Better:

  • While a step's background subagent is still running, refuse question with: "end your turn; the subagent's report wakes you".
  • Then the auto-approve rule above covers what's left: real one-option or defaulted questions.
**It loops.** Run 134's judge asked four of these in about four minutes ("Judge progress", "Waiting on browser run", "Awaiting browser report", "Still awaiting report"), each with one option, "Continue when ready". Each answer woke it; with its browser subagent still running it had nothing to do, so it asked again. So answering automatically isn't enough on its own: it would only make the loop faster. Better: - While a step's background subagent is still running, refuse `question` with: "end your turn; the subagent's report wakes you". - Then the auto-approve rule above covers what's left: real one-option or defaulted questions.
Author
Owner

Shipped in #284 (the backend-comparison fixes), confirmed on production in run 139 (cmoriarty/breakout-planned, vLLM, auto-approve on).

Run 139's judge sent its browser subagent 12 tasks over 29 minutes: a desktop and a phone playtest through window.__app, with screenshots and console reads. It called no question tool, and the run opened no gate: the judge waited for each report by ending its turn, and the auto-approved run never stopped for a person. Round 1 passed.

Shipped in #284 (the backend-comparison fixes), confirmed on production in run 139 (cmoriarty/breakout-planned, vLLM, auto-approve on). Run 139's judge sent its `browser` subagent 12 tasks over 29 minutes: a desktop and a phone playtest through `window.__app`, with screenshots and console reads. It called no `question` tool, and the run opened no gate: the judge waited for each report by ending its turn, and the auto-approved run never stopped for a person. Round 1 passed.
Sign in to join this conversation.
No labels
No milestone
No project
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference
cmoriarty/braid#280
No description provided.