Claude Code: "auto mode classifier gave no verdict", how to fix it

You ask for an ls and Claude Code says it can't run it. Bash stops, Edit stops, and only file reads keep working. The message mentions an "auto mode classifier" and "no verdict". Between September 28 and 29, 2026, people hit this on macOS, Windows, WSL, Linux and the VS Code extension. This guide explains what the error means and shows the four ways back to work, with the commands.

What does "auto mode classifier gave no verdict" mean?

It means auto mode's safety check failed, not that your action was judged dangerous. With no verdict, Claude Code denies the action instead of running it unreviewed.

The full message, as it appears in issue #97854:

The server-side auto mode classifier gave no verdict (error), so auto mode cannot determine the safety of Bash. This is a transient failure of the check, not a judgment about the action: a later response may get a verdict.

In auto mode, a second model, the classifier, reviews each action before it runs, instead of a permission prompt. When the server returns no answer, the command stays blocked. The tool name changes (Bash, Edit, Agent, an MCP tool) and so does the reason in parentheses: the official errors page gives (timed out) as an example. File reads and code search don't go through the classifier, which is why they keep working.

Why are so many people seeing this error now?

Because auto mode became the default. According to the permission modes documentation, from version 2.1.283 Claude Code starts every interactive terminal and VS Code session in auto mode, on every plan. Before that, it only did so on Pro, Max and Team.

The result: many people who never picked auto mode now depend on the server-side classifier. When it wobbled on 09/28, the main issue gathered 27 comments in two days, and several others with the same error were opened the same days (#97766, #97812, #97868). Anthropic's status page logged elevated errors on Claude Code on 09/29, from 14:21 to 16:27 UTC, but as of 09/30 there was no incident naming auto mode.

What happens if I just keep retrying?

Claude Code waits longer between each attempt and, after 10 responses in a row with no verdict, ends the turn. During the wait the spinner shows Auto mode check unavailable with a countdown, and Esc interrupts.

The stop message, as reported in the same issue:

Auto mode is unavailable: the server returned no safety verdict for 10 responses in a row, so Bash was not run and this turn has ended.

Watch for half-finished work. A user reported in #97854 that the stop landed in the middle of a task with git add, and the edited file was left unstaged with no warning. After a stop like that, check:

git status
git diff --stat

How do I fix the "gave no verdict" error in Claude Code?

Try them in the order below. The first two handle server wobbles; the last two are for when the error won't go away.

1. Send another message

Any message from you resets the 10-response counter. Since the failure is server-side and intermittent, the next attempt often gets a verdict. That's what the author of #97854 saw: the pwd that kept failing went through on the next message, with nothing changed.

2. Leave auto mode with Shift+Tab

In the terminal, Shift+Tab switches mode. From auto, the first press goes to manual mode, and the next goes to acceptEdits. In manual, you approve each command; nothing goes through the classifier. To start the session outside auto:

claude --permission-mode default
# or, to approve file edits without asking:
claude --permission-mode acceptEdits

In acceptEdits, edits and common filesystem commands (mkdir, mv, cp) run without asking; the rest of Bash still asks for approval. To make manual the default for every session on the machine, use ~/.claude/settings.json:

{
  "permissions": {
    "defaultMode": "default"
  }
}

3. Need auto mode? Use the local classifier

Set the variable before starting Claude Code. It makes Claude Code use its own classifier requests instead of asking the server for the check:

export CLAUDE_CODE_AUTO_MODE_SERVER=0
claude

Two warnings from the classifier billing documentation: on a direct connection to the Anthropic API the variable requires version 2.1.281 or newer, and on accounts where those requests are billed they count as token usage. The docs call the variable temporary, so don't treat it as a permanent setting.

4. Always failing? Check your proxy and gateway

If the error never goes away and you use an LLM gateway or proxy, that's the likely culprit. The docs say the verdict gets lost when the gateway cuts the streaming response short or rewrites the safeguard_results field. Run:

/status

The Auto mode server row shows Enabled while the server decides the session's actions and Disabled once the session has fallen back to the local classifier. Ask the gateway admin to pass requests and responses through without changing fields.

Which path fits each case?

Situation	Likely cause	What to do
Fails sometimes, then works	Server-side classifier wobble	Send another message
Turn stopped after 10 failures	No verdict from the server for minutes	Shift+Tab to manual and git status
Need to run unattended	Server unstable and you want to keep auto	CLAUDE_CODE_AUTO_MODE_SERVER=0
Always fails, at any time	Gateway or proxy cutting the response	/status and review the gateway
Notice that auto mode is unavailable in the session	Auto mode turned off server-side, for the account, or by disableAutoMode	Start a new session later

The last case has its own rule: when the server turns auto mode off or rejects it for your account, the session stays without auto mode until it ends. Don't keep pushing that session; start another one later.

Can I keep this from blocking my work again?

You can limit the damage. If you run long tasks without watching the terminal, keep defaultMode on manual or acceptEdits and turn on auto only when you'll be watching. If you prefer auto all the time, keep CLAUDE_CODE_AUTO_MODE_SERVER=0 handy for wobbly days. And after any turn ended by the 10-response limit, run git status before moving on.