v0.2.96

0.2.96 — 2026-07-10

Features

• System notifications now carry structured kind/title/body for better rendering.
• x.ai/pr/status now reports whether an open PR is in the merge queue.
• Compact mode now activates automatically on very small terminals.
• Up arrow on an empty prompt now browses prompt history; /history searches it.
• Stop hook runs now appear inline on the turn-completed line instead of a separate block.
• Subagent rows now fold into verb-group headers and the tasks pane shows live activity labels.
• Dashboard shortcuts now advertise ? instead of Ctrl+. on terminals that cannot deliver the latter.
• Double-clicking scrollback while Text selection is fold/nav now shows a tip offering Ctrl+Y to enable Word select.
• grok worktree ls now works as a short alias for grok worktree list.
• MCP tool output truncation can now be set per-repo in .grok/config.toml.
• Auto-send of queued follow-ups during task waits can now be enabled fleet-wide via remote settings.
• Welcome screen now offers one-click resume of a recent Claude Code session via ctrl+u.

Bug Fixes

• Vim l key now opens the selected agent detail view in the dashboard.
• Terminal commands with no args now run through a shell, matching the CLI.
• Agent teardown no longer crashes on slim Linux images that lack the ps command.
• Esc now dismisses an open /btw panel before backing out of a dashboard overlay.
• Resumed grok.com chats now use the conversation's last model instead of the gateway default.
• JetBrains terminals on Windows now default to minimal mode to avoid raw mouse-report leaks in the prompt.
• Skill token highlights now survive line wraps and the slash menu opens when typing / before existing text.
• Truncated or tiny images are now dropped before sending and previously poisoned sessions self-heal on restart.
• Session switch hints after /new or fork now show the working command in minimal mode.
• Progress bars in Ghostty and WezTerm now stop correctly for parked task waits.
• /effort now rejects levels the current model does not support instead of sending a bad value to the API.
• /recap on a fresh session now says "No messages yet" instead of failing.
• Monitor and system messages no longer appear as user prompts when resuming old sessions.
• /rewind completion now appears as a brief toast instead of a permanent transcript line.
• Auto recaps no longer appear under a newer user message when you start typing again.
• Authentication retries after token refresh no longer hang for minutes or days.
• Text selection tip no longer appears on the first double-click or on non-assistant blocks.
• Skill slash commands queued while a turn runs can now be sent immediately with Enter or the interject chord.
• Drag text selection now works inside the dashboard dispatch input box.
• Multi-line paste chips in dashboard inputs now support preview and expand like the main prompt.
• Live previews now always fetch the latest content without browser or CDN caching.