Omniscio

OmniscioGuidesAuto mode is unavailable

Claude Code auto mode is unavailable

When auto mode stops, it is almost always the classifier rather than your code. This page takes the messages in that family one at a time: the exact sentence, what Claude Code was doing when it printed it, and the steps Anthropic documents for getting moving again. It also names the two phrasings people search for that Anthropic's documentation does not carry.

Last checked against Anthropic's documentation on October 23, 2026. By the Omniscio team. Updated .

Which message did you get?

Auto mode runs a second model, the classifier, over each action before it happens. Most of what goes wrong in auto mode is that classifier failing to answer, and Claude Code prints a different sentence for each way it failed. Find yours below. Reads, searches and edits inside your working directory skip the classifier altogether, so those keep working whatever the message says.

Auto mode is unavailable: the server returned no safety verdict

Auto mode is unavailable — the server returned no safety verdict for the last 10 responses, so Claude stopped. Send a message to try again, or switch out of auto mode.
The server-side auto mode classifier gave no verdict (timed out), so auto mode cannot determine the safety of <tool>.
Auto mode check unavailable

What it means. Under server-side classifier review, Claude Code does not send a separate classifier request. It asks the server to check the action as part of the session's own model requests, and the server's verdict decides the action. When a response ends before those review results arrive, or they arrive in a form Claude Code cannot read, there is no verdict. Auto mode then denies that one action rather than let it run unchecked.

One denial is survivable. Ten in a row is not, and that is the first message above: auto mode stops the turn. The third line is not an error at all. It is the spinner text you may see during the pause Claude Code takes before some denials, with a countdown beside it. Pressing Esc interrupts the turn there.

A denial names the kind of failure in brackets when Claude Code can work it out, which is where (timed out) comes from. The brackets do not always hold the kind of an outage: a denial may name (the conversation is too long for the check) instead, which has its own section below.

How to fix it

  1. Send another message. Claude picks the work back up and the count of responses with no verdict starts over.
  2. If the denial names (the conversation is too long for the check), run /compact at the prompt first. Until the conversation is shorter, another message gets the same denial.
  3. Press Ctrl+O to open the transcript viewer and read the full denial text. The row under the tool call is shortened.
  4. If the stop keeps coming back and your requests travel through an LLM gateway or proxy, ask whether it cuts streaming responses short or rewrites them. Both produce a missing verdict, and a gateway that passes responses through untouched fixes it.
  5. Set CLAUDE_CODE_AUTO_MODE_SERVER=0 before you start Claude Code. It then makes its own classifier requests instead of asking the server. On a direct connection to the Anthropic API this needs version 2.1.281 or later, so run claude update first. Anthropic calls the setting temporary.
  6. To approve the actions yourself, switch out of auto mode and work in a mode that asks.
  7. Run /status and read the Auto mode server row to see which path the classifier is taking. Enabled means the server's checks are deciding the session's actions; Disabled means the session has fallen back to Claude Code's own classifier requests, which is how you tell whether CLAUDE_CODE_AUTO_MODE_SERVER=0 took effect. Disabled does not mean auto mode is off.

Not reproduced. Reaching this needs Anthropic's classifier to stop answering, which we cannot arrange. Two older behaviours are worth knowing if you are on an older build: before 2.1.280 Claude Code denied each action with no verdict immediately and never stopped the turn, and before 2.1.302, on a direct connection to the Anthropic API, it did not compact a conversation that had grown too long for the server's check.

Sources: Anthropic, Claude Code error reference and Anthropic, permission modes.

Claude Code classifier unavailable: a model is temporarily unavailable

<model> is temporarily unavailable, so auto mode cannot determine the safety of <tool> right now. Wait a moment and then try this action again.
<model> is temporarily unavailable (rate-limited), so auto mode cannot determine the safety of <tool> right now
Not run · auto mode's check had no usable answer

What it means. This is the other half of the family, and the message names a model for that reason: the classifier can run on a model other than the one you chose with /model, so a model being unavailable to the classifier is not always the same as your session's model being unavailable. The classifier could not produce a decision, so auto mode did not approve the action.

By default that model is Claude Sonnet 5, and Anthropic documents the sessions that get a different one: a session Anthropic has configured a classifier model for, a session running on Sonnet 4.6, and a session that sets ANTHROPIC_DEFAULT_SONNET_MODEL. There is also a fallback. When Sonnet 5, or the model named in ANTHROPIC_DEFAULT_SONNET_MODEL, is unavailable from your provider or excluded by your organization's model policy, the classifier falls back to another model: your session's own, or an Opus model in a Fable session.

That bracket holds the kind of failure, and this is the message Anthropic lists the kinds for: (rate-limited), (overloaded), (server error), (timed out) and (connection failed). When none of them fits, the sentence appears with nothing in brackets, and more than one failure at once produces that form.

In an interactive session you may never see the sentence itself. Claude Code puts a dim row under the tool call instead, which is the third line above, and Ctrl+O opens the transcript viewer where the full message is. Before 2.1.296 that row was a red error.

How to fix it

  1. Wait a few seconds and try the action again. This failure is usually short lived, and Claude often retries on its own.
  2. Carry on with work that does not need the classifier. Reads, searches and edits in your working directory all keep working through this.
  3. If the message returns every time and names a model, check that the API or cloud provider your session runs on serves that model. On Amazon Bedrock, a message that keeps repeating can mean your account cannot invoke the named model at all, which repeats until the grant arrives.
  4. If you set CLAUDE_CODE_AUTO_MODE_MODEL yourself, this is the moment to check it. Point it at a model your provider serves, or unset it and let Claude Code use its own choice, which is Claude Sonnet 5 by default.

When the classifier answers, but not usefully

Auto mode could not evaluate this action and is blocking it for safety — run with --debug for details

What it means. The classifier answered in a form Claude Code could not read. Unlike the message above, this one is Claude Code's own parse failing rather than the model being away, and retrying the action normally succeeds on the next attempt. Run claude --debug and do it again to see the detail in the log.

Not reproduced. Both of these need the classifier to fail on demand, which we cannot arrange from a test machine. Sources: Anthropic, Claude Code error reference and Anthropic, the default classifier model.

The one denial where retrying does not help

Auto mode could not evaluate this action and is blocking it for safety — a safety check separate from auto mode blocked this request because of earlier conversation content — it isn't about the action itself — run with --debug for details
Agent aborted: auto mode classifier request refused by the safety safeguard in headless mode

What it means. A safety check that sits beside auto mode, rather than inside it, refused the classifier's own request, because of something already in the conversation. It is not a judgment about the action you asked for. The second line is what a background subagent receives in a non-interactive run, where there is no prompt to fall back to.

Claude Code tells Claude exactly that, and tells it to move on to other work rather than retry. These denials also stay out of the counters that make auto mode fall back, so a handful of them will not push your session out of auto mode, and in a non-interactive run they do not stop the run at all. Before 2.1.225 they did count toward those thresholds and returned the same wording as a real classifier block, so an older build can look worse here than it is.

How to fix it

  1. Do not retry. The same conversation content triggers the same filter again, which is the whole reason this one reads differently.
  2. Switch permission modes so you can approve the action from a prompt yourself.
  3. Start a fresh conversation without the content that triggered it.

Not reproduced. Source: Anthropic, Claude Code error reference.

Too long for the check: the classifier transcript exceeded the context window

Auto mode classifier transcript exceeded context window — falling back to manual approval (try /compact to reduce conversation size)
<tool> was not reviewed and did not run: the conversation is too long for auto mode's classifier. Claude Code will try to compact the conversation before the next request; if this action is still needed after that, issue it again and it will be reviewed normally.
Agent aborted: auto mode classifier transcript exceeded context window in headless mode

What it means. The classifier is sent a slice of the conversation along with the pending action, and that slice has outgrown the classifier's own context window. This is separate from your session running out of room, and it can happen while your own context still has space left.

Three lines, and they arrive in an order worth knowing. The second message comes first: the action was never reviewed, so it did not run, and Claude Code compacts before the next request. That line is addressed to Claude rather than to you. If the compaction succeeds there is nothing to do. If it fails, actions the classifier still cannot check get one of the other two lines. The first message is what an interactive session shows instead: auto mode hands the action to an ordinary permission prompt so you can approve or deny it yourself, and the third message is that same fallback reaching a subagent in a non-interactive run. Either way the run itself carries on.

How to fix it

  1. In an interactive session, approve or deny the action in the prompt that appears.
  2. Nothing is needed from you when the compaction a not-reviewed message announces succeeds. If you then see one of the other two messages, run /compact yourself.
  3. If you turned auto-compact off, turn it back on so Claude Code compacts on its own next time.
  4. Run /compact so later actions fit inside the classifier's window again.

Not reproduced. Sources: Anthropic, Claude Code error reference and the same reference, the shorter entry.

Auto mode is not available to this session at all

auto mode disabled by settings
Auto mode isn't the default for your organization · Shift+Tab to switch

What it means. The first line is the plain reading of the words, and it is a different problem from the classifier failing. Auto mode is unavailable to the session itself, so there is nothing to retry. Anthropic lists four reasons it happens: a settings file turning it off, a model that does not support it, Anthropic having turned it off on its side, or the server having rejected auto mode for your account.

That first line is what a running session shows when an administrator's managed setting reaches it: auto mode leaves the session, and it also leaves the Shift+Tab cycle. Before 2.1.251 a session that was already running kept auto mode until it ended. When either of the two server answers is the reason, auto mode stays off until the session ends, so starting a new session later is the way back. And where the built-in default, the flag or a settings file asks for auto mode and the session cannot have it, Claude Code starts the session in Manual mode rather than failing.

The second line is not about being unavailable

Read the second line separately, because it does not mean auto mode is off. Auto mode isn't the default for your organization · Shift+Tab to switch appears in organizations with Anthropic's HIPAA configuration applied, where the built-in auto default does not apply and a session starts in Manual mode when nothing else chooses its starting mode. Auto mode works there and stays available: press Shift+Tab to switch to it, or make it the starting mode by passing --permission-mode auto, or by setting permissions.defaultMode to auto in your own settings or your organization's managed settings.

How to check which reason you have

  1. Look for disableAutoMode in your settings files. Set to disable, in managed settings it is an administrator's decision and no setting of yours overrides it.
  2. Check your model against the supported list for your provider, because the lists differ. On the Anthropic API and Claude Platform on AWS it is Claude Opus 4.6 or later, Sonnet 4.6 or later, Haiku 5.5, or a Fable model. On Amazon Bedrock, Google Cloud's Agent Platform, Microsoft Foundry and signed-in Claude apps gateway sessions it is narrower: Sonnet 5 or later, Opus 4.7 or later, Haiku 5.5, and the Fable models. Sonnet 4.5, Opus 4.5, Haiku 4.5 and the claude-3 models are not supported for auto mode on any provider.
  3. Put defaultMode in ~/.claude/settings.json, not in a project's .claude/settings.json or .claude/settings.local.json. Anthropic notes that auto does not take effect from those two files, and that a session starting in Manual mode with no error at all usually means the setting is in one of them.
  4. Run claude auto-mode config to print what the classifier actually uses once your settings are applied, and claude auto-mode defaults for the built-in rules.

Not reproduced. Reaching either message needs an organization-level setting or a server-side decision we cannot make from a test machine. Sources: Anthropic, permission modes, the HIPAA section and Anthropic, auto mode configuration.

Two phrasings people search for that Anthropic does not document

Two of the searches that lead here are wordings we could not find in Anthropic's documentation. We checked rather than guessed: we downloaded all 228 pages listed in Anthropic's own documentation index and searched every one of them for the exact phrase. The same search finds Auto mode is unavailable in two of those pages, so it does find things.

Permission mode is unavailable. That exact sentence is in none of those pages. If you are looking at it, the message beside it is the thing to read: auto mode disabled by settings is the documented wording for an administrator turning auto mode off, and it is the closest Anthropic has to a permission mode being unavailable. Which modes exist and how to move between them is Anthropic's own permission modes page, and the refusals you get when an action is turned down are on the error messages guide.

Auto-mode setup scan failed. The /auto-mode-setup command exists and it does scan: it reads your project's CLAUDE.md, README.md, config files and git remotes, your autoMode and permissions.allow settings, and the hosts, buckets and command names from recent sessions in that project, with two optional extra scans. Anthropic documents one failure message for it, and it is not about scanning. Type the command and you get Unknown command: /auto-mode-setup when it is missing, which happens on a first session after an install or upgrade, on a version below the minimum for your plan, and before 2.1.304 in a session that cannot fetch feature flags. In Claude Code's web environment the command says it isn't available there. If you have seen a scan fail with different words, Anthropic does not document it, and we will not invent a cause for it here.

Sources: the search described above, plus Anthropic, generating auto mode environment entries and its unknown-command section.

What does not help

The advice that circulates for these messages is mostly about the wrong failure, so it is worth saying which of the usual moves Anthropic does not list.

  • Restarting Claude Code. Not among the documented steps for a classifier that stopped answering. Send a message instead, which is the documented way to restart the count of responses with no verdict, or switch out of auto mode.
  • Reinstalling. Nothing in this family is a broken install. The failure is a request that went unanswered, and a fresh copy of Claude Code makes the same request.
  • Switching your session model when the message names a model. The classifier runs on Claude Sonnet 5 by default rather than on your session's model, so switching yours usually changes nothing about which model has to answer. A name in the message is the classifier's, and CLAUDE_CODE_AUTO_MODE_MODEL is the setting that moves it. Two cases break the rule, and both are in the classifier section: a session on Sonnet 4.6, and a session where the classifier's model is unavailable from your provider or excluded by your organization, which falls back to your session's own model, or to an Opus model in a Fable session. There is also one place where switching your session model is the fix rather than a distraction, and that is an unsupported session model in the availability section.
  • Waiting it out when the message names a settings file. A settings refusal is not a pass, and it does not lift on its own. disableAutoMode in managed settings is an administrator's decision.
  • Retrying a safety-safeguard denial. The only message in this family that tells Claude not to retry, because the trigger is content already in the conversation rather than the action.
  • Assuming the classifier outage is yours to fix. A server-wide failure reaches everyone at once, so check status.claude.com before you change anything. Your own side of it is a gateway or proxy that rewrites responses, or an AWS account that cannot invoke the classifier's model on Amazon Bedrock.

Sources

Every message and every step on this page rests on one of these, read on October 23, 2026. Anthropic's error reference changes with nearly every Claude Code release and keeps notes such as "Before v2.1.280", so read it again before you trust a copy of this page from a few months back.

Common questions

Why does Claude Code say auto mode is unavailable?

Two different problems get described that way, and only one of them prints those words. The words belong to the server-side classifier giving no verdict for ten responses in a row, which stops the turn. The other problem is auto mode not being available to the session at all: there the session either starts in Manual mode with no message, or shows auto mode disabled by settings. Which message you have is how you tell the two apart.

Do I need to restart Claude Code when the classifier stops answering?

No. Anthropic's error reference says to send another message, which restarts the count of responses with no verdict, or to switch out of auto mode. A restart is not one of the documented steps.

Can I run the classifier myself instead of the server?

Yes. Set CLAUDE_CODE_AUTO_MODE_SERVER=0 before you start Claude Code and it makes its own classifier requests. On a direct connection to the Anthropic API that needs Claude Code 2.1.281 or later, and Anthropic marks the setting as temporary.

What does "the server returned no safety verdict for the last 10 responses" mean?

Under server-side classifier review, Claude Code asks the server to check an action as part of the session's own model requests. When a response ends before the review results arrive, or they arrive in a form Claude Code cannot read, there is no verdict, and auto mode denies the action rather than run it unchecked. Ten of those in a row stop the turn.

Does "permission mode is unavailable" come from Claude Code?

We could not find that sentence anywhere in Anthropic's documentation. We downloaded all 228 pages listed in Anthropic's own documentation index and searched every one of them for the exact phrase. The closest documented wording is "Auto mode is unavailable", and the closest thing to a permission mode being unavailable is a settings file setting disableAutoMode to disable.

Why does the message name a model?

Because the classifier can run on a model other than the one you picked with /model. By default that is Claude Sonnet 5, with a documented fallback when it is unavailable from your provider or excluded by your organization's model policy: your session's own model, or an Opus model in a Fable session. A message naming a model means that model could not answer the classifier's request, which is a different failure from the server giving no verdict.

When several sessions stop at once

The messages on this page end a turn rather than crash anything, which is exactly the failure that hides when you are running a group of sessions. Omniscio, the desktop app we make for running many Claude Code sessions at once, gives every session one of eleven statuses and shows it in the sidebar, and collects the sessions that need a decision from you in an inbox at the top of that sidebar. The session statuses page in the library lists them, the features page covers the rest of the app, and the download page has the installers.