Claude Code: Anatomy of a Misfeature

TL;DR Highlight

Claude Code auto-skipped questions after 60 seconds idle, and the changelog never mentioned it

Who Should Read

Developers running several Claude Code agents, or letting them touch risky work like deployments

Core Mechanics

2.1.198 auto-continued after 60 seconds
- If AskUserQuestion, the tool that asks a human mid-task, got no reply in 60 seconds, it told the model to proceed with its best judgment.
- 2.1.198 shipped on July 1, 2026.
Partial answers got submitted too
- Answer one of three questions and step away, and the other two are filled in by the model.
- The "re-ask later" hint doesn't help because the re-asked question hits the same timeout.
The countdown only showed in the last 20 seconds
- The default threshold is 20s (CLAUDE_AFK_COUNTDOWN_MS), so the first 40 seconds looked like a normal blocking question.
- With several agents running, you can easily miss it.
Permission prompts were exempt from the timer
- The timer applies only to AskUserQuestion, and permission and plan approvals never auto-resolve.
- But with allowlists or bypassPermissions, questions like "staging or production?" were the last gate left.

Community Reactions

A Claude Code team member apologized
- Thariq, who built AskUserQuestion, said the change was meant to stop long-running jobs from blocking on questions.
- He said it should have been opt-in and in the changelog.
Opinion that it must never be an approval gate
- One commenter said irreversible actions should always hard-block for user input.
- Another said permission enforcement should live outside the process it polices.
Some users liked the feature
- One commenter described coming back to find Claude stalled after 5 minutes on a question when they had stepped away.
- Another said models are good enough to proceed and be corrected afterward.
Data loss risk raised from experience
- A commenter noted Claude Code doesn't auto-commit when stopping for a question, so uncommitted file edits could be lost.
- Another said a terminal click meant to focus the window got read as an answer.

How to Apply

Turn off auto-continue with an env var
- On versions before 2.1.200, set CLAUDE_AFK_TIMEOUT_MS to a huge value in settings.json env (a commenter used 2147483647).
- On 2.1.200 or later, check that Question auto-continue timeout in /config is never.
Add external guardrails to agent deployments
- Find where you allowlisted deploy commands or enabled bypassPermissions.
- Don't rely on the question tool; block risky commands with a sandbox, separate VM or pre-call approval hook.
Decide your update policy
- Set DISABLE_AUTOUPDATER to stop auto-updates.
- This also stops plugin updates, so pin a version and check releases manually.

Try It

copy & run

# Run in a terminal: check AskUserQuestion mentions in the public changelog repo

git clone https://github.com/anthropics/claude-code.git
cd claude-code

# Find the changelog commit that mentions auto-continue
git log -1 --format='%h %ai' -S'no longer auto-continue by default' -- CHANGELOG.md

# Count AskUserQuestion mentions
grep -c 'AskUserQuestion' CHANGELOG.md

# Next: add "CLAUDE_AFK_TIMEOUT_MS": "2147483647" to env in ~/.claude/settings.json

Terminology

AskUserQuestionThe tool Claude Code uses to ask a human to choose among options mid-task. It's a separate path from permission approval.

AFKShort for Away From Keyboard. It's the feature's internal name.

allowlistA list of commands allowed to run without asking. If a deploy command is on it, no permission prompt ever appears.

opt-inA feature you must turn on yourself, so it's off by default.

minifierA tool that shrinks code by shortening names, which is why functions in the binary look like Thl.