#### Server-side classifier review
The whole hunk
from line 36, old and new numbered
/
lines
from line 36
3636* Tools that require user interaction: the built-in `AskUserQuestion` tool and MCP tools marked [`requiresUserInteraction`](/docs/en/mcp#require-approval-for-a-specific-tool)
3737* `rm` and `rmdir` removals targeting a [critical path](#critical-paths), which no allow rule or `PreToolUse` hook `"allow"` approves
3838* The [cross-session messaging safeguards](#skip-all-checks-with-bypasspermissions-mode)
39* Reads outside the working directories while [`permissions.blockReadsOutsideWorkingDirectories`](/docs/en/settings-reference#permissions-blockreadsoutsideworkingdirectories) is on: recognized file-reading Bash commands and any [unsandboxed retry](/docs/en/sandboxing#the-unsandboxed-retry-escape-hatch) that needs approval to run outside the sandbox prompt even in auto mode and `bypassPermissions` mode. Requires Claude Code v2.1.257 or later
39* Reads outside the working directories while [`permissions.blockReadsOutsideWorkingDirectories`](/docs/en/settings-reference#permissions-blockreadsoutsideworkingdirectories) is on: recognized file-reading Bash commands prompt even in auto mode and `bypassPermissions` mode, and so does any [unsandboxed retry](/docs/en/sandboxing#the-unsandboxed-retry-escape-hatch) that needs approval to run outside the sandbox. Requires Claude Code v2.1.257 or later.
4040
41 A command the shell parser can't trace, such as one that changes directory more than once or runs a subshell, prompts the same way even when it names no outside path. This prompt doesn't apply when the command runs in the [sandbox](/docs/en/sandboxing) and the sandbox enforces the block.
42
4143## Common setups
4244
4345Permission modes decide whether Claude asks before an action, and the [Bash sandbox](/docs/en/sandboxing) and outer [isolation boundaries](/docs/en/sandbox-environments) decide what an action can reach once it runs. Each row below pairs a goal with the flags or settings that get you there and the isolation it needs, as a starting point. [Available modes](#available-modes) lists what runs without a prompt in each mode.
from line 53
5153| Run in CI with an exact allowlist | `claude -p "run the test suite" --permission-mode dontAsk --allowedTools "Bash(npm test)" "Read"` | None beyond what your CI runner provides | [Cloud sessions](/docs/en/claude-code-on-the-web) ignore `dontAsk` from settings files |
5254| Run fully unattended inside a container | `claude -p "<prompt>" --dangerously-skip-permissions` | Required: a container, VM, or the [sandbox runtime](/docs/en/sandbox-environments#sandbox-runtime); on Linux and macOS, run it as a [non-root user](#skip-all-checks-with-bypasspermissions-mode) | Cloud sessions ignore this mode from settings files. In this `-p` run, the [few calls that would still prompt](#skip-all-checks-with-bypasspermissions-mode) are denied instead |
5355
54The Bash sandbox and auto mode work independently and combine, except in plan mode, where [auto-allow doesn't widen approvals](/docs/en/sandboxing#sandbox-modes). For the full interaction, see [How sandboxing relates to permissions and permission modes](/docs/en/sandboxing#how-sandboxing-relates-to-permissions-and-permission-modes) and [How isolation relates to permission modes](/docs/en/sandbox-environments#how-isolation-relates-to-permission-modes).
56The Bash sandbox and auto mode work independently and combine, with the exceptions listed under [Sandbox modes](/docs/en/sandboxing#sandbox-modes). For the full interaction, see [How sandboxing relates to permissions and permission modes](/docs/en/sandboxing#how-sandboxing-relates-to-permissions-and-permission-modes) and [How isolation relates to permission modes](/docs/en/sandbox-environments#how-isolation-relates-to-permission-modes).
5557
5658<h2 id="which-mode-a-session-starts-in">
5759 Which mode a session starts in
from line 310
308310
309311In v2.1.158 through v2.1.206, auto mode was off on these providers until you set `CLAUDE_CODE_ENABLE_AUTO_MODE=1`, and Claude Code ignored `defaultMode: "auto"` on these providers unless the variable was also set. The variable is still accepted for compatibility and has no effect from v2.1.207 onward.
310312
313#### Server-side classifier review
314
315On Amazon Bedrock, Google Cloud's Agent Platform, and Microsoft Foundry, Claude Code reviews auto mode actions with its own classifier requests by default. To have the platform's server-side classifier review [the actions that go to the classifier](#how-the-classifier-evaluates-actions) as part of the session's model requests instead, set [`CLAUDE_CODE_AUTO_MODE_SERVER=1`](/docs/en/env-vars). Where the platform runs the classifier, its verdicts decide those actions; where it doesn't, Claude Code falls back to its own classifier requests. In v2.1.271 and v2.1.272, asking the platform was the default on these providers.
316
311317### What the classifier blocks by default
312318
313319The classifier trusts your working directory and the remotes that were configured for it when the session started. A remote added or repointed during the session with `git remote add` or `git remote set-url` isn't trusted, and everything else is treated as external until you [configure trusted infrastructure](/docs/en/auto-mode-config). Before v2.1.200, remotes added mid-session were also trusted.
from line 405
399405* Sending data to the trusted domains, buckets, and services you list in [`environment`](/docs/en/auto-mode-config#define-trusted-infrastructure). This covers data flow only, not destructive or credential operations on the same infrastructure
400406* [Claude in Chrome](/docs/en/chrome) navigation to a trusted internal domain, localhost, or a URL you named
401407
402Sandbox network access requests are routed through the classifier rather than allowed by default. As of v2.1.198, the classifier reuses its verdict for a network host and port instead of re-running on every connection:
408Sandboxed commands don't get network access by default. Claude names the hosts a command needs on the command itself, the classifier reviews them with the command, and an approved list opens those hosts for that one command alone. [Per-command allowed domains](/docs/en/sandboxing#per-command-allowed-domains-in-auto-mode) covers what a list can and can't open and what happens when a command reaches for an unlisted host.
403409
404* An allow is reused until new content enters the conversation, at which point that host is checked again
405* Claude Code v2.1.234 and later reuse a deny caused by the conversation outgrowing the classifier's context window until new content enters the conversation, or until [compaction](/docs/en/costs#reduce-token-usage) shrinks what the classifier reads. Claude Code then checks the host again
406* A deny that the classifier reached by evaluating the request lasts for the turn in the interactive CLI. In [non-interactive mode](/docs/en/headless) and Agent SDK sessions, Claude Code reuses that deny for the rest of the run, because those sessions have no turn boundary
407* Changing your permission mode or rules drops all cached verdicts
408
409410Run `claude auto-mode defaults` to print the full rule lists as JSON. If routine actions get blocked, an administrator can add trusted repos, buckets, and services via the `autoMode.environment` setting: see [Configure auto mode](/docs/en/auto-mode-config).
410411
411412Pushing to any branch of the repository you're working in and creating a pull request that matches your request run without a prompt, unless the push or pull request falls under the [blocked list](#what-the-classifier-blocks-by-default), such as secrets or sensitive data leaving the repository, or a pull request that targets a different repository or organization. To require a human checkpoint before these commands while staying in auto mode, add `permissions.ask` rules, which match the command [as written](/docs/en/permissions#bash-rule-limits): see [Common boundaries](/docs/en/auto-mode-config#common-boundaries).
from line 448
447448 <Accordion title="How the classifier evaluates actions">
448449 Each action goes through a fixed decision order. The first matching step wins:
449450
450 1. Actions matching your [allow, ask, or deny rules](/docs/en/permissions#manage-permissions) resolve immediately. Writes to [protected paths](#protected-paths) route to the classifier even when an allow rule matches, and so do `rm` and `rmdir` removals targeting a [critical path](#critical-paths) in Claude Code v2.1.218 and later. MCP tools marked [`requiresUserInteraction`](/docs/en/mcp#require-approval-for-a-specific-tool) prompt you directly even when an allow rule matches, and so do connector tools [your organization set to `ask`](/docs/en/mcp#organization-controls-on-connector-tools) in sessions where that setting reaches Claude Code. Ask rules that match on a command's content, such as `Bash(git push *)`, fall back to a permission prompt
451 1. Actions matching your [allow, ask, or deny rules](/docs/en/permissions#manage-permissions) resolve immediately, with these exceptions:
452 * Writes to [protected paths](#protected-paths) route to the classifier even when an allow rule matches, and so do `rm` and `rmdir` removals targeting a [critical path](#critical-paths) in Claude Code v2.1.218 and later
453 * MCP tools marked [`requiresUserInteraction`](/docs/en/mcp#require-approval-for-a-specific-tool) prompt you directly even when an allow rule matches, and so do connector tools [your organization set to `ask`](/docs/en/mcp#organization-controls-on-connector-tools) in sessions where that setting reaches Claude Code
454 * A shell command that carries [per-command allowed domains](/docs/en/sandboxing#per-command-allowed-domains-in-auto-mode) also routes to the classifier even when an allow rule matches, because a rule approves the command, not its hosts
455 * Ask rules that match on a command's content, such as `Bash(git push *)`, fall back to a permission prompt
451456 2. Read-only actions and file edits in your working directory are auto-approved, except writes to [protected paths](#protected-paths) and [the first read outside the working directories](#first-read-outside-the-working-directories), which prompts you
452457 3. Everything else goes to the classifier. The connector tools and `requiresUserInteraction` MCP tools that prompt you directly in step 1 never reach the classifier, so neither an org-required approval nor a consent step is auto-approved
453458 4. If the classifier blocks, Claude receives the reason and tries an alternative. In most sessions the reason names the rule the classifier matched, such as `[Data Exfiltration]`, rather than giving a written explanation; see [Review denials](/docs/en/auto-mode-config#review-denials)
from line 469
464469
465470 Claude Code also runs `git status` itself before a command that would discard uncommitted work, such as `git reset --hard` or `rm -rf`, and shows the classifier whether staged, modified, or untracked work is present. Claude Code reports untracked files in that check even when the repository's git configuration sets `status.showUntrackedFiles=no`.
466471
467 The classifier sees user messages, tool calls other than read-only lookups such as file reads and searches, and your CLAUDE.md content. Tool results are stripped, so hostile content in a file or web page can't manipulate it directly. You can annotate a call's result with a [PostToolUse hook's `classifierContext` field](/docs/en/hooks#annotate-a-result-for-the-auto-mode-classifier), which the classifier reads as application-provided context.
472 In the classifier requests sent by Claude Code itself, the classifier sees user messages, tool calls other than read-only lookups such as file reads and searches, and your CLAUDE.md content. Tool results are stripped from those requests, so hostile content in a file or web page can't manipulate the classifier directly.
468473
474 You can annotate a call's result with a [PostToolUse hook's `classifierContext` field](/docs/en/hooks#annotate-a-result-for-the-auto-mode-classifier), which the classifier reads as application-provided context. The field requires Claude Code v2.1.236 or later.
475
469476 A separate server-side probe scans incoming tool results and flags suspicious content before Claude reads it. For more on how these layers work together, see the [auto mode announcement](https://claude.com/blog/auto-mode) and the [engineering deep dive](https://www.anthropic.com/engineering/claude-code-auto-mode).
470477 </Accordion>
471478
from line 491
484491
485492 The session's first auto-mode request validates the Sonnet 5 default: if the request succeeds, Sonnet 5 stays the session's classifier model, and if it fails because the model isn't available, the session uses the fallback instead. After that validation settles, the classifier's model doesn't change for the session.
486493
487 On Enterprise plans and on accounts that use the Claude API, [Claude Platform on AWS](/docs/en/claude-platform-on-aws), Amazon Bedrock, Google Cloud's Agent Platform, or Microsoft Foundry, classifier calls count toward your token usage. Each check sends a portion of the transcript plus the pending action, adding a round-trip before execution. Reads and working-directory edits outside protected paths skip the classifier, so the overhead comes mainly from shell commands and network operations.
494 On Enterprise plans and on accounts that use the Claude API, [Claude Platform on AWS](/docs/en/claude-platform-on-aws), Amazon Bedrock, Google Cloud's Agent Platform, or Microsoft Foundry, classifier calls count toward your token usage. Each check sends a portion of the transcript plus the pending action, adding a round-trip before execution. Reads and working-directory edits outside protected paths skip the classifier, so the overhead comes mainly from shell commands and network operations. On Amazon Bedrock, Google Cloud's Agent Platform, and Microsoft Foundry, you can move the review into the session's model requests instead; see [Server-side classifier review](#server-side-classifier-review).
488495
489 The classifier reuses a sandbox network verdict for a host and port, so repeated connections to the same host don't each add a check. [What the classifier blocks by default](#what-the-classifier-blocks-by-default) describes how long an allow and a deny last.
496 Sandboxed network access adds no per-connection classifier requests. The classifier judges [the hosts a command names](/docs/en/sandboxing#per-command-allowed-domains-in-auto-mode) together with the command in one review, and Claude Code checks each connection against the approved list without calling the classifier again.
490497 </Accordion>
491498</AccordionGroup>
492499