Claude Code's Silent 60-Second Timer: A Misfeature Postmortem

On July 1, 2026, Anthropic shipped Claude Code v2.1.198 with an undocumented feature: a 60-second auto-continue timer on the AskUserQuestion tool. If you did not respond within that window, Claude would proceed using "best judgment" and continue working without your input.

The feature was rolled back three days later after a GitHub issue collected 384 upvotes and 143 comments. But the incident reveals something important about how agent tooling ships, how safety assumptions can silently change, and why the changelog matters more than you might think.

What Actually Happened#

The AskUserQuestion tool is a blocking safety gate. When Claude Code encounters ambiguity - which database schema to use, which API key to prefer, whether to delete old migrations - it pauses and asks the human. The tool exists precisely because some decisions should not be delegated.

In v2.1.198, that blocking gate became a 60-second countdown. The behavior worked like this:

1. Claude asks a question
2. A timer starts (invisible for the first 20 seconds)
3. At 40 seconds remaining, a countdown appears
4. At 0 seconds, Claude picks "best judgment" and continues

For users running multiple parallel agents, watching all countdown timers simultaneously was impossible. For anyone who stepped away from their terminal - to read documentation, check another file, or simply think - Claude would proceed without them.

Half-answered dialogs were the worst case. If you started typing but did not submit, Claude would auto-submit the partial response combined with its own generated choices.

The Documentation Gap#

Anthropic's changelog for v2.1.198, v2.1.197, and v2.1.199 made no mention of the feature. The public documentation on archive.org shows zero references on July 1st. Documentation only appeared after the fix on July 3rd, describing the reversed behavior.

Users discovered the escape hatch (CLAUDE_AFK_TIMEOUT_MS) through peer discussion, not official channels.

The author of the detailed postmortem had Claude Code investigate itself. The findings:

• The feature shipped as compiled binary code, not in any public repository
• No source commits exist for either the addition or removal
• Reverse-engineering the v2.1.198 binary revealed complete instrumentation: countdown UI, analytics tracking, and schema fields
• The analytics event tengu_ask_user_question_afk_auto_advance specifically tracked partial-answer scenarios

The binary diff between v2.1.197 and v2.1.198 contained 156 lines of new human-readable strings - small enough to catch with diligence, but only if you knew to look.

What HN Is Saying#

The Hacker News discussion ranges from frustrated users to people who actually wanted the feature.

On the safety implications:

"If you miss the window, Claude Code helpfully does what it thinks is best and continues on its way. Turning a blocking gate into a 60s countdown silently voids that assumption."

"I built Human blockers into my agentic workflows with great intention, so naturally this annoyed me deeply."

On the legitimate use case:

"It's very frustrating to leave a Claude session running and come back to find it did nothing because it got stuck on a question."

On the broader product direction:

"I really hate this direction both Anthropic and OpenAI are following. They are in this silly competition whose model/harness can go unattended the longest, no matter what."

"As soon as tokens stop being subsidized I would not trust any harness made by a company that also charges for the compute."

The comparison to OpenAI came up repeatedly - Codex has the same 60-second timeout, but has refused to make it configurable. At least Anthropic reverted and added an opt-in toggle.

On the broader state of Claude Code UX:

"I would love Claude Code to be a little less vibe-coded. The underlying model is excellent, but we're being pretty much forced into using CC to use the subscription model."

"The most recent one that's had me annoyed is the Fullscreen TUI feature, which is super unintuitive, implementing its own text highlighting and copy-on-select mechanics, overriding your terminal's native right click."

Why This Matters Beyond Tokens#

The incident highlights three systemic issues:

1. Silent behavioral changes

Combined with Claude Code's default auto-update behavior, users had no control over when this appeared. You could go to sleep with a trusted tool and wake up with different safety properties.

The fix for this exists:

// ~/.claude/settings.json
{
  "env": {
    "DISABLE_AUTOUPDATER": "1",
    "FORCE_AUTOUPDATE_PLUGINS": "1"
  }
}

The second variable preserves plugin updates while freezing the CLI - otherwise, disabling updates also freezes all plugins.

2. Closed-source safety infrastructure

The source repository for Claude Code contains no actual code - only release notes, examples, and automation scripts. As the postmortem author notes:

"Did a human review the feature? Did a human merge the feature? Did a human release manager diff the release?"

These questions remain unanswerable. The shipped binaries are analyzable, but the governance process is invisible.

3. Misaligned incentives

Multiple commenters noted the economic tension:

"Right now the interests align, but as soon as more tokens -> more profit (instead of more revenue and more losses) the perverse incentives will be too big to avoid."

A feature that burns tokens while you are away serves the platform's usage metrics. The user experience of "Claude worked all night" feels good even when the actual work quality suffers from missing human checkpoints.

How to Protect Yourself#

If you want Claude Code to actually wait for your input:

Option 1: Environment variable

export CLAUDE_AFK_TIMEOUT_MS=2147483647

This sets the timeout to roughly 24 days.

Option 2: Per-session config

Use /config to toggle auto-continue off for the current session.

Option 3: Pin your version

Use the DISABLE_AUTOUPDATER setting above and only update when you have read the changelog - assuming the changelog is complete.

Option 4: Use the SDK instead

The Agent SDK gives you programmatic control over the conversation flow. Several commenters noted they switched to SDK usage specifically to avoid CLI behavior changes:

"I've been using the thing through their agent SDK for several months now so I wouldn't have to deal with any of the wonky shit they change every second week in the CLI."

The Practical Takeaway#

The feature itself was not necessarily bad - some users genuinely wanted unattended operation. The problems were:

1. No opt-in - it shipped as default behavior
2. No documentation - users discovered it through surprise
3. No changelog entry - breaking safety assumptions without notice
4. 60 seconds is too short - barely enough time to read the question, let alone research an answer

Anthropic fixed points 1-3 after the backlash. But the incident reveals that agent harness development is moving fast, and safety properties you rely on can change without warning.

For critical workflows, the lesson is simple: do not trust tool defaults. Audit your agent's permission surface, pin versions when stability matters, and assume any blocking gate might become a countdown in the next release.

Official Sources#

Continue Reading#

• Building SaaS with AI Agents in 2026: The Complete Workflow
• Interview Mode: Let Claude Code Ask the Questions First
• Using Claude Code for a Second Opinion on MRI Scans - What Actually Happened