Claude Code 2.1.301 adds a maximum wait setting for unattended retries and fixes several failures that could mislead an automated workflow. Anthropic’s October 15 release notes describe changes to worktree cleanup, failed searches and skipped workflow steps, alongside new tracing fields and terminal fixes. The practical focus is distinguishing a completed action from an interrupted or failed one.
On this page
Unattended retries gain a usage-limit wait boundary
The official v2.1.301 release adds CLAUDE_CODE_RETRY_WATCHDOG_MAX_USAGE_LIMIT_WAIT_MS. Anthropic says unattended retry mode can use it to end a request when a usage limit will reset too far in the future, instead of continuing to wait.
The setting belongs to the retry watchdog. The release notes do not establish a universal recommended value for every workflow, so teams should use the documented behavior for their own runtime rather than copy an assumed timeout. A request ending at a limit is still an unfinished task that the surrounding system must account for.
Search failures and skipped steps become more visible
Anthropic says Grep and Glob previously could report no matches when their search program failed before returning results. Version 2.1.301 reports a search error in that case. An empty result and an unsuccessful search require different follow-up: the latter does not establish that a file or symbol is absent.
The release also fixes workflow agents sometimes skipping a step while the run still reported success. For a workflow that checks files, changes code and runs validation, a success label therefore still needs to be assessed against the saved changes and actual check results. The notes identify the fix; they are not evidence that every third-party workflow has been validated.
Worktree cleanup preserves a consequential distinction
Another fix concerns quitting a session whose WorktreeCreate hook created a worktree. The notes say Claude Code could remove that worktree without asking even when it might contain unpushed commits. The fix is relevant to teams using custom worktree hooks; it does not replace normal review of uncommitted changes and unpushed history before cleanup.
For the broader hook model, see the Claude Code hooks guide.
Tracing gets more detail about permission decisions
The release adds classifier_verdict and permission_mode to OpenTelemetry tool_decision events. With OTEL_LOG_TOOL_DETAILS=1, classifier_category also identifies the built-in rule matched by a classifier block. These fields help distinguish classifier decisions from permission-rule decisions.
Anthropic also reports fixes for sign-in refresh interruptions and a misleading login message caused by a full disk. Those are separate failure causes: a request to sign in again should not be assumed to identify every underlying storage or session problem.
