xAI

The /sessions command now opens the Agent Dashboard instead of a separate picker, and a new /goal slash command is available for setting objectives. Several fixes address output rendering (queued bash commands, minified JSON files no longer clipped at 2000 characters, background task preambles with newlines), terminal rendering improvements (text selection highlighting, Windows conhost drag-select support), and startup performance on large repositories.

User-run shell commands now display their complete output after finishing instead of silently dropping middle lines. Image pasting on macOS is now ~65× faster by reading the pasteboard directly instead of via osascript. Also added /minimal and /fullscreen commands to switch session modes, made ask_user_question tool configurable, fixed edit tool highlighting for multi-line strings, and improved terminal link handling.

Grok 4.5 is a new model built for coding, agentic tasks, and knowledge work.

MCP permission prompts now display the planned arguments so you can see what the tool will do before approving. Esc no longer cancels a running turn (use Ctrl+C instead), and double-Esc rewind now works while focused on scrollback. Fixed a crash shortly after launch caused by the telemetry exporter.

MCP tool output size limit is now configurable via environment variable, config.toml, or GrowthBook with the default unchanged. You can now add a local directory as a plugin marketplace source with grok plugin marketplace add, and auto permission mode prompts far less often on routine development commands.

Voice dictation indicator and stop button now remain visible and clickable during plan mode review. The new Worktree dialog expands to show long names and scrolls with a leading ellipsis when the terminal is narrow.

Voice dictation now works on Linux (requires pipewire, pulseaudio-utils or alsa-utils), and a new /auto slash command switches to classifier permission mode with the menu showing only the other mode. Shell state persists across tool calls in Cursor:Shell sessions, image edits use the higher-quality Imagine model, and git operations no longer cause long pauses from repeated full-repo scans.

Scrolling now feels smoother with better trackpad and wheel handling plus configurable speed and mode. Session search returns tighter multi-word results, handles filenames and plurals better, and the session picker now searches conversation content for queries of two or more characters. Resuming sessions no longer fails permanently when history files contain corrupted lines from interrupted writes.

Subscription upgrades are now detected automatically without restarting the CLI, and a new /docs slash command opens How-to Guides picker, browses web docs, or jumps directly to a guide by title. Bash permission prompts now offer a "Never allow" choice that persists the deny rule, reasoning effort menus are configurable per-model from the server and config.toml, and several UI and interaction bugs are fixed.

Media generation no longer hits per-session file limits, and image and video byte budgets have been increased. Voice language setting now lets you pick speech-to-text language in the Editor settings, and tab autocomplete suggests your next prompt as ghost text after each turn.

Tool call grouping can now be enabled to collapse consecutive read/search/list calls into single rows, and subagents now run in the background by default. Pressing Enter on an empty prompt sends the top queued follow-up immediately, permission prompts now correctly wrap long bash commands, and inline LaTeX math with padding spaces renders correctly.

Added remote_fetch option under [features] in config.toml to disable all backend catalog and settings fetches for air-gapped environments. Idle CPU and memory usage after long sessions or resume is now dramatically lower, and grep now stops early on broad searches. Announcements update live during sessions without restart, ask_user_question timeout is now configurable, thinking blocks display by default, and several UI and tool fixes improve usability.

Critical announcements now appear in a top banner during active sessions with a hide command. Pasting the same text again next to a paste chip now expands the chip into editable text instead of duplicating it, and paste preview shows a hint explaining how to expand the chip.

New /effort slash command changes reasoning effort on the active model. Managed connector links now include team ID in team sessions, chat titles from the gateway display in the sidebar, and skill descriptions recover correctly even with malformed YAML frontmatter. Plus improvements to text chip editing, terminal rendering, and voice session startup.

Grok iOS app 1.3.97 adds a Dictate to Grok shortcut that launches the app into dictation mode and sends messages by voice from the Shortcuts app, Action Button, or Siri.

Voice dictation STT is now available using Ctrl+Space or F8, with optional hold-to-talk on supported terminals. Background tasks and TODO lists now survive compaction and remain visible to the model. Command timeouts are configurable per-session, contextual hints can be toggled individually, subagent dialogs reliably show full transcripts, and network connections are more resilient to proxy or load-balancer drops.

New stream transforms let hosts hide, unwrap, or rewrite tool calls for display without affecting agent transcripts. Fixed sessions that remained stuck on the thinking indicator after the model finished responding, and recap requests no longer trigger context-length 400 errors on long conversations.

Pasting images into dashboard peek replies now works and survives turn cancellation. Prompt history is now saved even on fast Ctrl+C quit, fixing data loss on quick exits. Plus improvements to LaTeX rendering in tables, link clickability in /btw panels, tab alignment in prompts, and shell command cleanup.