Claude Code 2.1.235 fixes a Shift+Tab that auto-granted edit access

The Aug 18 point release tightens permission dialogs, stops a language-server reconnect from busting your prompt cache, and trims CPU on background reviews.

Nowline AUG 19 6:00 AM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • A stray Shift+Tab was granting session-wide edits

    In the permission prompt's comment field, Shift+Tab approved the pending edit and granted session-wide edit permission instead of just closing the field — a silent auto-approve of everything after. 2.1.235 fixes it; update before your next long agent run.

  • Dialogs now match what you're actually granting

    Permission dialog text and the 'don't ask again' option now always reflect the true scope of a grant, and 'don't ask again' is withheld when the contents can't be fully shown. Fewer nasty surprises about what you just approved.

  • A language-server blip no longer busts your prompt cache

    Whole-prompt-cache invalidation was firing whenever a language server disconnected or reconnected mid-session, silently re-billing you for a full cache miss. Now fixed — cheaper, faster turns in editor-connected sessions.

  • Background reviews stop eating memory and CPU

    Long-running cloud sessions like /ultrareview and /autofix-pr no longer re-scan and re-render their whole event stream on every update. Leave them running and expect noticeably lower CPU and memory.

  • grep -m with context finally prints the right lines

    The native macOS/Linux embedded grep now handles -m N alongside -A/-C correctly and fails fast on pathological patterns instead of exhausting memory — fewer searches that wedge mid-task.