Claude Code 2.1.301 landed on npm on October 15 at 14:34 UTC, a day after 2.1.300. It’s a large maintenance release, roughly fifty entries, most of them fixes.
What’s new
CLAUDE_CODE_RETRY_WATCHDOG_MAX_USAGE_LIMIT_WAIT_MS puts a ceiling on unattended retry mode. If a request hits a usage limit whose reset is too far away, it now ends instead of waiting. For scheduled runs, that’s the line that matters most here.
Vim mode gets the % motion: jump to the matching bracket, on its own or with d, c and y.
Two new OpenTelemetry fields on tool_decision events. classifier_verdict and permission_mode separate auto mode classifier decisions from permission rules. With OTEL_LOG_TOOL_DETAILS=1, classifier_category names the built-in rule a classifier block matched. Anyone reading the Claude Platform traces now gets the missing half.
Mods gain rows on their Input: a multi-line text field in the terminal where Enter submits and Shift+Enter, or a backslash before Enter, starts a new line.
The fixes worth noticing
Workflow agents sometimes skipped their step while the run still reported success, triggered by the harness note saying their task text carries no user authority. A silent failure inside an automation is the worst kind.
Switching to a model with a smaller context window could spend usage credits on a summary by the previous model without asking. That’s fixed, and the behavior improved alongside it: a conversation too long for the new model is now deliberately summarized by the old one first.
Behind an LLM gateway whose responses name a different model than the one requested, conversations were being compacted over and over, on a model nobody picked.
Also: the fable alias now retries once on the previous Fable model when an organization lacks access to the newest one, rather than erroring. The Read tool no longer reads .xlsm, .parquet, .epub and .heic as text. Quitting no longer silently removes a worktree created by a WorktreeCreate hook that may hold unpushed commits.
The pattern is automation
Three of the most visible fixes concern runs nobody is watching: the skipped workflow step that still reported success, the unasked-for billing on a model switch, the endless compaction behind a gateway. Add the new cutoff for the retry watchdog. The release reads like the cleanup after a week in which Claude Code mostly ran unattended.