Watch

grokkybara[bot] c68e39f604 Publish harness and TUI open-source

initial sync from the monorepo

2026-07-16 06:46:02 +01:00

224 KiB

Raw Blame History

Changelog

0.2.101 — 2026-07-13

Features

grok inspect now shows effective compatibility settings for Cursor, Claude, and Codex sessions.
New setting "Match display refresh rate" lets high-refresh displays run the TUI at native cadence.

Bug Fixes

Parked subagent status no longer duplicates or interleaves incorrectly in scrollback.
Status line during waits now shows elapsed time before the queued-message hint.
Queued messages sent with Enter now appear immediately instead of vanishing briefly.
Resume hint after quitting minimal mode now prints the correct grok --minimal --resume command.
Rate-limit messages now correctly direct API-key users to team plans instead of personal upgrades.

0.2.100 — 2026-07-13

Features

Session picker now discovers and resumes recent Claude Code, Codex, and Cursor sessions.
Welcome screen now offers a one-click resume nudge for recent Claude, Codex, or Cursor sessions.

Bug Fixes

Web fetch tool preserves full truncated page content as readable artifacts instead of discarding it.
Multiline mode now correctly sends the top queued message on empty Enter when a turn is running.
Queued commands no longer disappear or delay when pressing Enter twice quickly during a running turn.
Minimal mode text is now readable on dark terminals with proper contrast and highlighted user prompts.
Grok no longer crashes when printing resume hints after the terminal pane has closed.
Long-running turns with multiple waits now show updated status markers in the transcript instead of appearing stuck.
Claude and Cursor hooks are now correctly disabled at session start when disabled in config.

0.2.99 — 2026-07-12

Features

Multiline input now works on the agent dashboard the same way it does in regular sessions.
PageUp and PageDown now scroll the conversation while the prompt is focused.
Keyboard Shortcuts modal now follows Vim mode navigation keys when enabled.

0.2.98 — 2026-07-12

Breaking Changes

You can now pin authentication to API key or OIDC in config.toml; the unpinned method is no longer tried automatically.

Features

/context now shows token costs for skills and MCP servers.
env_key in config now accepts an array of environment variable names.
Linux middle-click paste from the primary selection now works; clipboard errors are handled more reliably.
/terminal-setup now shows your terminal's color support level and which themes are available.
grok setup --json prints your team's managed configuration without installing it.
Messages you type while the model waits on tasks now stay queued; pressing Enter twice sends them immediately by cancelling the current turn.
How-to Guides modal now shows a tip linking to Ask Grok above the footer shortcuts.
Subagent task and spawn_subagent tools now accept an optional model parameter in the CLI.
Keyboard Shortcuts modal now lists the paste key binding for images under the Input section.

Bug Fixes

A pre_tool_use deny now feeds the reason back so the model can retry instead of cancelling the turn.
Plan mode now strictly rejects edits outside the plan file, even under always-approve.
Web search and X search no longer fail when both a local function tool and the backend hosted tool are active.
Content-filter refusals from providers now show an explanation instead of ending silently with no output.
SQLite databases no longer cause bus errors on network filesystems such as NFS.
Resuming a session that is already open now focuses the existing view instead of creating duplicates.
Turn completion markers in scrollback now read "Worked for …" instead of "Turn completed in …".
/btw loading spinner now animates correctly when the main session is idle.
Mid-turn wait spinners now correctly show "Waiting on task output…" instead of Thinking.
Scrollbar thumb is now visible in the oscura-midnight theme.
Status messages for background work now end with a period.
Editing queued prompts no longer freezes the terminal or duplicates text into the composer.

0.2.97 — 2026-07-11

Features

Headless JSON output now includes token usage and cost per prompt and session.
SDK turns now expose detailed token usage and cost information via Turn.usage.
Double-click or Enter on a previous user message now lets you edit and resubmit it directly from the transcript.
Text selection in scrollback now works better when starting on chrome, gaps, or while scrolling.
Shell commands using rg (ripgrep) no longer require permission prompts by default.
Voice mode is now available for API-key sessions.
New environment variables allow tuning scroll and draw cadence for high-refresh displays.

Bug Fixes

Background tasks started by the model in headless mode are now killed on exit instead of leaking.
Agent process leaks on failed spawns and missing-stdio teardown are now prevented.
Parked turn markers no longer appear after interjections and now count down as background tasks finish.
The /context tool definitions line no longer shows the cryptic disclaimer suffix.
Terminal release waits boundedly for the process to exit and repeated waits share the drain grace.
Fixed a crash when the agent asked a question on narrow terminals.
Fixed misleading keyboard hints in the /mcps panel in minimal mode.
Clipboard copy now reports success correctly when using iTerm2 over SSH.
Fixed scroll/input conflicts when plan approval appeared over an open edit block.
Fixed frequent MCP and skills reloads that could freeze sessions on devboxes.
MCP servers using HTTP (such as streamable HTTP MCP servers) now automatically recover from disconnects.
Next reset time in /usage now shows in your local timezone instead of Pacific Time.

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.

0.2.95 — 2026-07-09

Features

Teams can now ship default allowed commands via managed_config.toml (user deny rules still win).
Mid-turn interjections now appear as normal user prompts (❯) instead of a separate cyan block.

Bug Fixes

IME text input in Otty no longer attaches unrelated clipboard images on every character.
Rewind now fully removes the selected turn from both scrollback and the model's conversation history.
Queued prompts now abort long blocking waits instead of waiting for the full timeout.
File links and media now work for worktree sessions under ~/.grok/worktrees/.
Collapsed Read/Edit tool rows now show only the filename instead of long absolute paths.
Clipboard copies on Wayland now succeed even when the terminal loses focus mid-copy.
User messages queued behind an auto-wake turn are no longer lost when the user presses Ctrl+C.
Slash completion now shows sibling skills that share a frontmatter name and correctly sizes wrapped descriptions.
Single tool calls that belong to a verb group now collapse into an aggregated header row.
Fixed sessions that became permanently stuck after tool-use history corruption.
/always-approve and /auto now toggle their mode on and off when run repeatedly.
Terminal command cards on grok.com now correctly settle after foreground bash tasks.
Copy failure toast now recommends trying /minimal for native terminal rendering.

Performance

File watching on Linux now uses far fewer system resources for large projects with many dependencies.

0.2.94 — 2026-07-09

Features

/sessions now opens the Agent Dashboard instead of a separate picker.
New /goal slash command** is now available when the workspace supports it.
grok inspect now lists skills from [skills].paths and correctly labels bundled vs user skills.
--minimal and --fullscreen choices are now remembered for future plain grok launches.

Bug Fixes

Queued bash commands promoted at turn end now render their output instead of disappearing.
Xcode / Foundation ACP clients can now drive grok agent stdio without silent parse drops on session/* calls.
read_file now returns full single-line content (minified JSON, large dumps) instead of silently clipping at 2000 characters.
Background task command preambles with newlines now render on separate lines instead of collapsing.
Text selections now highlight uniformly even over inline code, links, and syntax-colored spans.
grok --minimal now supports native drag-select on classic Windows conhost terminals.
Skill tokens such as /pr-workflow are now highlighted teal when used mid-sentence.
Fixed a crash when a filtered list shrinks while the filter is active.
Scroll lines and scroll speed settings now support fine unit-step adjustments.
Project-specific Claude plugins are no longer visible outside their project directory.
First prompt no longer stalls for many seconds on large repositories while the filesystem watcher starts.

0.2.93 — 2026-07-08

Breaking Changes

Esc no longer cancels a running turn; use Ctrl+C instead. Double-Esc rewind now works while focused on scrollback.

Features

MCP permission prompts now show the planned arguments so you can judge what the tool will actually do.
The "Managed by grok.com" link in the Extensions modal is now clickable and underlined.
Dragging inside rendered markdown tables now selects whole cells or rectangular ranges and copies as TSV.
Shift+Tab now goes straight to Plan mode when the plan-mode tip is showing.

Bug Fixes

grok --minimal now aligns the prompt, status bar, and messages flush-left with the welcome card.
/plugins no longer lists never-installed Claude marketplace entries and now groups plugins by their real source.
Successful image compression no longer leaves a permanent line in the transcript.
--no-ask-user now also disables ask_user_question for subagents.
--no-ask-user now also disables ask_user_question for subagents.
Fixed a crash shortly after launch on some systems caused by the telemetry exporter.

0.2.92 — 2026-07-08

Features

/minimal and /fullscreen commands let you switch the current session between minimal and fullscreen modes.
ask_user_question tool can now be enabled or disabled via config.toml, environment variables, or remote flags while defaulting to on.

Bug Fixes

User-run shell commands now display their complete output after finishing instead of silently dropping middle lines.
Edit tool output now correctly highlights multi-line strings and scopes that previously spilled across hunks.
Always allow grants for MCP, web_fetch and bash now take effect immediately in auto mode without re-prompting.
Cmd/Ctrl+click on bare http(s) links now opens only once on Warp terminals.
Cmd/Ctrl+click now works on imagine media paths and URLs that wrap across multiple terminal rows.
grok update on Windows no longer fails when a previous.old executable is still running.

Performance

Pasting images on macOS is now ~65× faster by reading the pasteboard directly instead of via osascript.

0.2.91 — 2026-07-07

Bug Fixes

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

0.2.90 — 2026-07-07

Features

New /minimal and /fullscreen slash commands let you switch the current session between minimal and fullscreen modes without quitting.
Session titles from /rename now appear on the prompt box border after resume.
grok models banner now correctly reports per-model API keys and deployment keys.
MCP tool output size limit is now configurable via environment variable, config.toml, or remote settings (default unchanged).
Chat conversations listed in the unified sidebar can now be renamed or deleted from the desktop app.
You can now add a local directory as a plugin marketplace source with grok plugin marketplace add.
Auto permission mode now prompts far less often on routine development commands.
Short media paths the model prints (images/1.jpg) are now clickable and open the file.
Preview now prefers common dev ports like 8080 when multiple HTTP servers are detected.

Bug Fixes

Model list now refreshes after upgrading from a free to a paid subscription tier.
Extensions modal now shows clearer enable/disable and install hints that match what each key actually does.
Folder trust no longer prompts for or scans the entire home directory when it is a git repo.
Code blocks no longer lose their background shading on the final line of unterminated fences.
Plan mode now activates immediately when toggled during an active turn instead of waiting for the next prompt.
Clicking back into the terminal window now focuses the prompt immediately when a permission or plan-approval panel is waiting.
Paths the model emits inside quotes no longer end with literal backslash-n characters and cause file-not-found errors.
Next reset time shown by /usage is now correct during daylight saving time.
The ask-user-question tool now waits up to 30 minutes by default before timing out.
The free-usage paywall no longer offers a Try Again button.
Inline images no longer bleed through when entering or leaving the fullscreen subagent view.
The [Open Image] button under generated media is now colored like a link.
Preview routing now auto-selects well-known ports like 8080 even with framework signals on obscure ports.
Login screen now centers the authentication URL when it fits on one line.
Enter key now queues follow-ups by default while the agent waits on task output.

0.2.89 — 2026-07-07

Features

Voice dictation now works on Linux (requires pipewire, pulseaudio-utils or alsa-utils).
New /auto slash command switches to classifier permission mode; the menu now shows only the other mode.
--effort and --reasoning-effort are now interchangeable CLI flags for setting reasoning effort.
Image edits now use the higher-quality Imagine model for better output.

Bug Fixes

Try Again on the free-usage paywall now correctly resubmits after rate-limit retries.
Cursor now respects your terminal's default blink style instead of always blinking.
Skill commands in scrollback now highlight only the command name, not the arguments.
Plan files now default to.grok/plan.md to match Grok conventions.
LaTeX math renders correctly for display equations and complex subscripts.
Queue hint in the terminal no longer shows incorrect bold text on part of the message.

Performance

Git operations like rebase no longer cause long pauses from repeated full-repo scans.

0.2.88 — 2026-07-06

Features

Scrolling feels smoother with better trackpad and wheel handling plus configurable speed and mode.
Session search now returns tighter multi-word results and handles filenames and plurals better.
Session picker now always searches conversation content for queries of two or more characters.
Tool call grouping is now enabled by default, folding consecutive reads and searches into single rows.
Plugins tab now supports u to update the selected plugin and shows non-blocking success feedback.
Reasoning effort used for a session is now recorded in summary.json and conversation history.

Bug Fixes

Session content search now correctly indexes messages containing escape sequences like newlines and quotes.
Formatted links now keep their link color when wrapped in bold, italic, or strikethrough.
Resuming sessions no longer fails permanently when history files contain corrupted lines from interrupted writes.

0.2.87 — 2026-07-05

Features

Subscription upgrades are now detected automatically without restarting the CLI.
Bash permission prompts now offer a "Never allow" choice that persists the deny rule.
New /docs slash command opens How-to Guides picker, browses web docs, or jumps directly to a guide by title.
Per-model reasoning effort menus are now configurable from the server and config.toml without a client release.
Finished thinking blocks now fold into grouped tool-call rows when group_tool_verbs is enabled.

Bug Fixes

--minimal mode now always uses your terminal's own colors so text stays readable on any background.
Invalid fields in [model.*] config blocks no longer cause the whole model to disappear from the picker.
File tools no longer target paths with literal trailing newlines or whitespace from model output.
Fixed interactions in no-freeform question modals so clicks below the last option no longer enter input mode.
Background tasks now correctly wake the agent after a cancelled blocking wait instead of staying idle.
Copying quoted text from rendered responses no longer includes the quote bar prefix in pretty mode.

0.2.86 — 2026-07-04

Features

Voice language setting now lets you pick speech-to-text language (including System) in the Editor settings.
Tab autocomplete now suggests your next prompt as ghost text after each turn.
/usage (and /cost) is now hidden for free and X Basic personal accounts.
Media generation no longer hits per-session file limits; image and video byte budgets increased.

Bug Fixes

--minimal flag now shows in grok --help.
Session resume notifications no longer appear when a workspace boots for the first time.
Claude-style Bash(cmd:*) permission rules are now correctly translated to prefix matches.

0.2.85 — 2026-07-03

Features

Pressing Enter on an empty prompt now sends the top queued follow-up immediately while a turn is running.
Tool call grouping can now be enabled via config.toml or settings to collapse consecutive read/search/list calls.
Consecutive tool calls of the same kind can now be folded into a single row when group_tool_verbs is enabled.
Subagent conversations now receive the same type-specific instructions in gateway chat as in the CLI.
Scheduled automation tasks now show their header panel correctly in gateway chat sessions.
Promotional announcements now appear with clickable CTAs and can be non-dismissible.
Subagents now run in the background by default unless explicitly set to false.

Bug Fixes

Permission prompts now correctly wrap long bash commands while preserving structure and quotes.
Claude Code settings with permissions.defaultMode are now correctly honored.
Project skills and commands are now discovered even when their directories are gitignored.
Inline LaTeX math with padding spaces now renders correctly instead of showing raw dollar signs.
Manual /recap now works on the same turn and long auto recaps are hidden from view while still saved.
Always allow for common commands like ls and git status now remembers just the command instead of extra arguments.

0.2.84 — 2026-07-03

Features

Announcements now update live during active sessions without restart or /new.
Hiding an announcement no longer suppresses later criticals; new ones reappear automatically.
run_terminal_cmd now requires a one-sentence description rationale in every invocation.
ask_user_question timeout policy is now configurable in config.toml and /settings.
Ask-Question timeout can now be toggled from /settings (Agent & Approval).
Thinking/reasoning blocks are now shown by default while the model is working.
Critical announcements now show a red title with a clickable [hide] button and aligned message.
Added remote_fetch option under [features] in config.toml to disable all backend catalog and settings fetches for air-gapped environments.

Bug Fixes

Images pasted or read from GIF, BMP or TIFF files are now automatically converted so they work with image generation.
Queue panel now shows action buttons on hover and the status bar displays a compact done/total task count.
Hook matchers now correctly see the real MCP tool name instead of the internal dispatcher name.
Copy now succeeds when running inside containers even when the terminal brand cannot be detected.
Tool result previews no longer paint opaque panels in grok --minimal.
grok wrap now correctly handles quoted strings and shell aliases.
Text selection settings now correctly honor explicit keep_text_selection values even when legacy keys remain.
Fixed a freeze that could occur when editing and sending the last message in the queue.
Fixed a startup crash on minimal Linux systems lacking system CA certificates.

Performance

Grep now stops early on broad searches, returning faster results with far less memory use.
Idle CPU and memory usage after long sessions or resume is now dramatically lower.

0.2.83 — 2026-07-02

Features

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.
Paste preview now shows a hint explaining how to expand the chip.

0.2.82 — 2026-07-02

Features

Managed connectors links now include the team ID when opening from a team session.
AGENTS.md files are now discovered and shown for workspace/hub sessions.
Chat conversation titles from the gateway are now shown in the sidebar.
New /effort slash command changes reasoning effort on the active model.
Double-click a pasted text chip to expand it into editable text.

Bug Fixes

Skill descriptions are now recovered correctly even when frontmatter YAML is malformed.
[Esc] hint in /btw panels now stays visible even on narrow terminals.
Background monitors now wake the agent on natural exit the same way bash tasks do.
Long option labels in question prompts are now always visible instead of disappearing when unfocused.
Pasted text preview now appears immediately after inserting a paste chip.
Hex color codes now render as colored dots with no extra space.
Pressing voice on the welcome screen now starts a new session.

0.2.81 — 2026-07-01

Features

Chat sessions no longer send workspace binding hints that belong to the backend.
New stream transforms let hosts hide, unwrap, or rewrite tool calls for display without affecting agent transcripts.
cancel() now accepts a timeout so a stuck turn cannot hang the session forever.
Run tool blocks now show the model's description as the main title when provided.
Hex color codes in prose now render as colored dots on truecolor terminals.
New setting "Show thinking blocks" controls whether agent reasoning is visible in the scrollback.
Spinners now show the description or short command of the task being waited on when available.

Bug Fixes

Fixed sessions that remained stuck on the thinking indicator after the model finished responding.
Mermaid diagrams now correctly display angle brackets and symbols instead of literal HTML entities in labels.
Recap requests no longer trigger context-length 400 errors on long conversations.

Performance

Grep searches now time out after 20 seconds by default (60s on WSL) instead of always waiting 60s.

0.2.80 — 2026-07-01

Features

Command timeouts can now be configured per-session with a foreground-only ceiling.
Background tasks and TODO lists now survive compaction and remain visible to the model.
Voice dictation STT feature: uses Ctrl+Space or F8, with optional hold-to-talk on supported terminals.
Contextual hints can now be toggled individually for undo, plan mode, and image input.

Bug Fixes

Subagent dialogs now reliably show full transcripts on open and reopen.
Recap blocks now copy only the summary body, not the header label.
Vim navigation keys now type into dashboard prompts; modals properly handle Esc/Left.

Performance

Network connections are now more resilient to proxy/LB drops.

0.2.79 — 2026-06-30

Features

Contextual hints now show shortcuts like plan mode or clipboard paste when relevant.
Graceful shutdowns now allow interrupted turns to resume with a configurable pause budget.
Grok.com chat sessions now integrate fully with the gateway bridge for model catalog and resume.

Bug Fixes

Question prompts now time out after 6 minutes instead of blocking forever.
Fixed a crash that could occur during conversation integrity repairs while a turn was active.

Performance

Compaction can now run part of its work in the background before it blocks the session.

0.2.78 — 2026-06-30

Features

Chat sessions show the grok.com model catalog in the picker.

Bug Fixes

Tabs pasted into the prompt now align correctly with proper cursor positioning.
Pasting images into dashboard peek replies now works and survives turn cancellation.
Links in /btw panels are now clickable and highlight on hover.
Prompt history is now saved even on fast Ctrl+C quit.
Stuck scrollback text selection can now be cleared with Esc or any non-drag input.
LaTeX math now renders inside markdown tables in the TUI.
Background shell commands started by the agent are now cleaned up when the CLI exits.

Performance

grok update downloads have a longer timeout.

0.2.77 — 2026-06-30

Features

Pasting images from the local clipboard now works when running commands through grok wrap.
Turn status spinner now shows what the agent is waiting on (response, subagent, task output, etc.).
Double-click word selection is now a discoverable option in the Text selection setting and stays in sync with highlight behavior.

Bug Fixes

Credit limit errors now show clearer upgrade or buy-credits messaging based on billing type.

0.2.76 — 2026-06-30

Features

Auto permission mode is now added to the top of Shift+Tab cycles and enabled by default in settings.
grok agent stdio now checks for updates in the background like other modes.

Performance

Idle sessions no longer send repeated empty frames to the terminal, reducing CPU usage in the terminal emulator.

0.2.75 — 2026-06-29

Features

Prompt history (Up arrow / Ctrl+R) now shows only the current session's prompts, with the newest selected at the bottom.

0.2.74 — 2026-06-29

Features

Esc now cancels a running turn immediately; double-Esc clears prompt or opens rewind when idle.
grok wrap now shows copy success over SSH and suggests native drag-select when paste fails.

Bug Fixes

Clipboard copy now succeeds reliably on Wayland and KDE desktops instead of showing false positives.

0.2.73 — 2026-06-28

Features

Keep text selection highlight setting added so drag selections stay visible until dismissed.

Bug Fixes

Doubled lines after tab switches or focus changes in tmux or editor terminals are now healed.
Clipboard copy now only shows success when the pasteboard actually received the text via a trusted path.

0.2.72 — 2026-06-28

Bug Fixes

No longer triggers browser login at startup when an API key is already configured for inference.

0.2.71 — 2026-06-27

Bug Fixes

Fixed grok agent stdio hangs on Windows when used with persistent clients such as VS Code.

0.2.70 — 2026-06-27

Breaking Changes

Added grok wrap to run any command with local clipboard support.

Features

Ctrl+4 now toggles the prompt queue on local macOS VS Code terminals.

Bug Fixes

Session recaps (/recap and return-from-away) now show the full summary instead of being cut off mid-sentence.
Vim mode now focuses the prompt when you press / on a brand-new empty session.
Fixed grok agent stdio startup hangs on Windows when used with persistent clients such as VS Code or grok-desktop.
/mcps list no longer shows stale disabled entries when managed gateway tools are enabled.
Mermaid diagrams opened via [Open Image] now render at higher resolution instead of terminal size.
Pressing r in scrollback no longer accidentally rewinds the session.
Shortcuts cheatsheet now shows Ctrl+X on terminals that cannot deliver Ctrl+.
Folder trust prompts no longer re-appear for every standalone worktree clone.
Reasoning effort no longer silently resets from a user-chosen value after catalog refreshes.
Fixed clipboard copy inside editor terminals nested in tmux by emitting plain OSC 52.

0.2.69 — 2026-06-26

Features

The agent dashboard now shows each agent's model and mode in the peek panel, lets you cycle modes with Shift+Tab, collapses the Inactive section by default, and hides older idle agents behind a "N more" row.
Tool usage cards for search, directory listing, file deletion and glob now render as distinct typed cards instead of generic MCP entries.
The keyboard shortcuts help now shows richer descriptions and correctly scrolls wrapped text in the detail view.
You can now pass --json-schema to grok -p and receive a validated JSON object instead of free text.
Ctrl+L now interjects mid-turn in VS Code, Cursor, Windsurf, and Zed terminals.

Bug Fixes

Local plugins installed from your home directory are now automatically refreshed when you start a session, so new agents or skills added to the source appear immediately.
The /context command now reports the same number of tool definitions that are actually sent to the model.
In vim mode the agent dashboard peek no longer steals keyboard focus from the list, so j and k keep moving between agents.
/sessions on the agent dashboard no longer freezes the interface.
Dashboard now focuses the overview list immediately when agents exist.

0.2.68 — 2026-06-26

Features

MCP servers from host integrations can now be added, replaced, or removed without restarting the session.
Agent-run terminal commands now set GROK_AGENT=1 so host tools can tell them apart from interactive shells.

Bug Fixes

Attached images are now saved to real disk paths so the model can read them in any terminal.
/resume now selects the correct model when a saved model name is ambiguous.
Slash and completion menus no longer crash if the terminal is resized while open.

0.2.67 — 2026-06-25

Features

Added --json-schema flag for headless mode to constrain model output to a supplied JSON Schema.
Idle detection can now ignore background tasks when the env flag is set (off by default).

Bug Fixes

Preview panes no longer hibernate while actively viewed or polled.
Manual /rename now persists correctly and appears in /session-info even after auto title generation or resume.

Performance

Find and grep now transparently use faster bfs and ugrep binaries when present in the harness.

0.2.66 — 2026-06-25

Features

Custom sandbox profiles can now kernel-deny specific files and directories for reads/writes.
Marketplace plugins in subdirectories of a git repo can now be installed and loaded correctly.
Folder trust prompt now appears before starting a session when the feature is enabled.
Preview panes no longer hibernate while actively viewed.
Keyboard shortcuts help now expands inline for individual entries instead of only sections.
Idle detection can now ignore background tasks when the env flag is set (off by default).
Sandbox deny lists now accept glob patterns like /*.pem in addition to exact paths.

Bug Fixes

Local MCP servers now auto-recover after disconnects or session expiry.
OIDC sessions with XAI_API_KEY present no longer lose refresh on idle.
Inline video previews now show an install command only when the package manager is on PATH.
list_dir now reliably shows all immediate child directories even inside large monorepos.
Clicking a model in the dashboard /model dropdown no longer opens the wrong session.
Strikethrough now only applies to ~~double tildes~~; single ~~tildes~~ render literally.
Session cycling with Ctrl+[ / ] now switches from the session you are currently viewing.
Prompt history (Up / Ctrl+R) now shows the complete recent list instead of a scrambled partial one.
Authentication now correctly prefers the session method when both API key and cached token are present.
xychart-beta diagrams with category labels now render correctly as images.

0.2.65 — 2026-06-24

Features

grok -w --ref now creates worktrees based on the specified ref instead of HEAD.

Bug Fixes

Unidentified Windows consoles are now treated as Windows Terminal for capability decisions.
Esc in the dashboard input now moves focus to the list without clearing your typed draft.
Copying a tool header now copies just the path or command, not the Read/Run label.
Execute activity lines and headers no longer repeat a redundant cd into the session directory.
Inline video previews now show an install hint instead of a spinner when ffmpeg is missing.

Performance

Headless and stdio sessions no longer start unnecessary filesystem watchers, saving CPU and IO.
Scrolling feels more responsive in VS Code, Cursor, and Windsurf integrated terminals.

0.2.64 — 2026-06-24

Features

Dashboard now displays the current directory and branch; click or press Ctrl+L to change location, or Ctrl+W to dispatch new agents into fresh git worktrees.
/recap now appears as a collapsible tool-style block with a loading spinner while generating.

Bug Fixes

Dashboard arrow keys open agent details and exit overlays; closing an agent now selects the neighboring row.
/usage command and credit warnings are now hidden for API-key authentication.
MCP servers from your user config no longer appear labeled as project-scoped when running from your home directory.

0.2.63 — 2026-06-23

Bug Fixes

Fixed hook matchers so pipe-list and alias patterns no longer silently over-match unrelated tool names.

0.2.62 — 2026-06-23

Features

Hosts can now register hooks over the agent connection instead of only on-disk files.
Prompt and /usage warnings now correctly reflect prepaid credits and auto top-up status.
Desktop clients can now detect when a terminal is busy running a foreground process.
TODO list remains visible to the model after compaction so it can continue working on pending items.
/recap is now available by default — it generates a quick summary of your current session so you can catch up on what's happened so far.

Bug Fixes

MCP server connections no longer time out during slow cold starts of stdio servers that download dependencies.
File paths containing spaces (e.g. macOS app bundles) are now correctly turned into clickable hyperlinks in the terminal.
Resume now correctly picks the most recently active session instead of one that only had metadata updates.
/goal slash command now appears in the menu on the welcome screen before any prompt is sent when the feature is enabled.
Session picker no longer shows a stale row highlight when keyboard focus moves to the search bar.
Usage percentages in /usage and warnings now match backend flooring and show pay-as-you-go limits when applicable.
Team accounts can now list sessions after re-login; previously returned 403 on conversations API.

0.2.61 — 2026-06-22

Features

Closing a terminal tab with a running process now shows a confirmation dialog instead of killing it immediately.
/usage now shows prepaid credits balance and auto top-up status.
Clipboard copy on Wayland now also tries wl-copy; per-leg outcomes are now logged for diagnostics.
Goal mode toggles and limits can now be set in config.toml under the [features] table.
All /goal options (toggles, limits, role models) are now configurable together in a [goal] table.
Clipboard copies from VS Code over SSH now warn when non-ASCII text may be garbled.

Bug Fixes

Focus reports no longer leak as literal text when split across reads over SSH.
--disable-web-search now honored in grok -p and grok agent; auxiliary model routing respects catalog overrides.
Focus events now fire correctly for SSH-split focus reports.
Boolean tool flags now accept "true"/"false"/"yes"/"no"/1/0 strings and numbers in addition to native booleans.
Session last-active timestamps and message counts no longer regress under concurrent writers.
iTerm2 now always uses text/metadata image fallback instead of broken OSC 1337 overlays.
Model switches no longer leave the prompt queue stuck after a reconnect.
Closing a terminal tab with a running process no longer shows a confirmation dialog.
Custom agent profiles now correctly use the harness required by their pinned model.
Subagents under custom profiles now adopt the correct harness from the parent's model.
Changelog and release-notes modals now scroll with the mouse wheel and arrow keys.

0.2.60 — 2026-06-21

Features

/resume now shows sessions from your current working directory's repo at the top of the list.
Too-wide Mermaid diagrams now show a hint below the fallback box pointing to the Open Image button.
Cancel behavior for running subagents can now be set to always stop or always continue in config.toml.

Bug Fixes

Compaction no longer hangs indefinitely when the summarizer stream stalls after the server has finished.
Slash command completion now shows consistent suggestions and remembers recently used commands.
Queued prompts now reappear reliably after deleting the last item and re-queuing.
Headless sessions no longer produce authentication error noise from unauthenticated MCP servers.
Mermaid flowchart labels with long identifiers are now kept whole instead of being cut mid-word.
Cmd+Backspace now deletes only from the cursor to the start of the line instead of clearing the whole prompt.
Inline Mermaid previews now break long identifiers at word boundaries instead of mid-segment.
Signed git commits no longer corrupt the TUI by letting pinentry draw over the screen.
Arrow keys now move the prompt cursor or open history while a /btw answer panel is visible.
Long option descriptions in question prompts now expand fully when the row is focused.

Performance

Large MCP tool results are now truncated inline and saved to disk to avoid unnecessary context compaction.

0.2.59 — 2026-06-19

Bug Fixes

Session recaps no longer display doubled labels and manual recap now correctly suppresses the next automatic recap.

0.2.58 — 2026-06-19

Bug Fixes

Terminal command output files are now capped at 5 GB during execution and truncated to 64 MB after the process exits.
Interjection messages now display the actual user text instead of a generic header.
The legacy agent command is now kept in sync with grok after running grok update.
Headless (grok -p) runs now wait for background tasks and subagents to finish before exiting.

0.2.57

Features

Improved resilience to network blips during long responses by resuming instead of failing the turn.
grok plugin install <name> now resolves plugins from registered marketplaces instead of only local paths.

Bug Fixes

Fixed cases where long-running conversation compaction could hang indefinitely.
Notification hooks now fire only for real user-attention events and no longer trigger constantly during tool use.
Fixed literal display of HTML entities such as < and > in responses and tool output.
Typing [ in the pager prompt no longer appears delayed.
Copy now tries all available Linux clipboard tools so paste works reliably in more terminals.

0.2.56

Features

resume_from now continues a finished sub-agent in place instead of forking a new conversation.
grok sessions delete command now lets you permanently remove a session from the CLI.

Bug Fixes

MCP server connections no longer get torn down during rapid config reloads.
Stale leader processes are now cleaned up when leader mode is disabled via config or remote settings.
Sandbox profile is now preserved when resuming sessions so commands continue to work as before.
list_dir now shows more relevant files when a large directory appears early in alphabetical order.
Cancel button in turn status always shows [stop]; queue pane highlight now follows theme changes.
grok quit no longer hangs when background git or network tasks are slow.
The token count shown after auto-compaction now matches the context bar exactly.
The git branch icon now renders correctly in iTerm2 without a Nerd Font.
list_dir now gives clearer guidance when a directory is too large, using the actual tool names available in your session.
Ctrl+Enter now sends the prompt when the agent is idle (same behavior as Enter).
resume_from now correctly continues a sub-agent in the same working directory it was using before.
Files with non-ASCII names (e.g. Chinese) no longer crash the session when plan mode checks for markdown.
Session lists (welcome screen, /resume, grok sessions list) are now sorted by the same activity time shown in the UI.
Fixed bash tool failures when models send numeric arguments such as timeout as JSON strings instead of numbers.
Prevented crashes during bash command output streaming when building progress frames.
Disabled inline image rendering on iTerm2 terminals where scrollback overlays cannot be supported.

Performance

Fast tools like grep now show as completed immediately even when other tools in the same round are still running.
Long sessions that display inline images no longer grow to multi-GB memory usage.

0.2.55

Features

Added option to fully disable the hunk tracker via --hunk-tracker-mode, GROK_HUNK_TRACKER, or config.

Bug Fixes

Windows install scripts now download and run cleanly via irm | iex without spurious BOM errors.
Tables and wide content no longer leave stray characters next to timestamps in the scrollback.
Mermaid diagrams now render node labels cleanly without HTML tags or raw markdown syntax.
MCP servers using HTTP now recover automatically after temporary connection drops instead of becoming permanently unavailable.
Very long sessions can now scroll all the way to the bottom of the conversation history.

0.2.54

Features

Rewind now works end-to-end across conversation and file state with proper CAS handling.

Bug Fixes

Git branch icons now render correctly on Windows without Nerd Fonts.
Mermaid diagrams now render inline without the model suggesting external viewers.
MCP connection errors now show the actual failure reason to the model.
MCP servers with noisy stdout no longer disconnect unexpectedly.
Usage warnings now always display "Usage left: N%" instead of varying between "Free credits left" and "Credits left".
Window title no longer flashes or oscillates during permission prompts while the terminal window is focused.

Performance

Fixed pager freezes and 100% CPU usage when rendering very long agent reasoning outputs with thousands of styled spans.

0.2.53

Bug Fixes

Minor bug fixes.

0.2.52

Features

Tool auto-approval (YOLO) state is now tracked end-to-end in server-side agent sessions.
ER diagrams now render as entity boxes with attributes and relationships in the TUI.
New "Respect manual folds" setting keeps hand-expanded blocks stable while content streams in.
Ctrl+X now stops running turns or closes sessions from inside the agent detail view.
Grok can now export usage metrics and events to your own OpenTelemetry collector when enabled.
WezTerm users now receive guidance when Shift+Enter fails because kitty keyboard protocol is disabled.
Long-running sessions now tell the model when the local calendar date changes past midnight.
Agent Dashboard now works without leader mode and shows local idle sessions from disk.

Bug Fixes

Fixed oversized session replay logs that prevented large sessions from loading.
MCP server connections no longer flood reconnects on repeated stream errors.
ZDR and team upload flags are now populated immediately on login instead of only after background refresh.
Mermaid PNG export now handles quoted cardinalities in class diagrams and readable ER rows on dark theme.
Skill catalog no longer shows duplicate "Use when:" labels and check-work skill now prompts the model to read its instructions.
Compaction now rejects overly-short summaries that would discard real conversation state.
Background tasks no longer emit spurious failure messages when a session is resumed.
Fixed Windows path handling so external tools and model prompts receive clean paths without ?\ prefixes.
Images and media no longer remain visible when switching from an agent view to the dashboard.
Clipboard paste (Ctrl+V) now works for images on pure Wayland sessions.
Modals such as /sessions no longer crash on narrow terminals.
ptyctl resize now correctly notifies the child process.
Concurrent updates to the same version no longer fail with permission or EEXIST errors.
Mermaid diagrams containing CJK or other non-Latin text now render correctly instead of tofu boxes.
grok dashboard now reliably opens the dashboard instead of silently falling through to a normal session.
Sessions no longer remain blocked forever after a transient model catalog outage during reconnect.
Cancel no longer leaves the interface stuck on "Cancelling…" after lost responses during reconnects.
Forked sessions now retain the parent's full pre-compaction transcripts instead of only the compacted summary.
web_fetch errors on GitHub hosts now recommend using the gh CLI when internal access is blocked.
MCP server connections no longer hang when stdio servers emit undecodable lines.
Ctrl+C cancels now complete in under 50 ms instead of blocking for seconds.
Repeated varied edit failures on one file no longer trigger doom-loop warnings or terminations.

Performance

Compaction now reuses cached prompt prefix instead of full prefill.

0.2.51

Breaking Changes

grok mcp add now accepts positional arguments (e.g. grok mcp add filesystem -- npx...), supports --scope project, and adds -e/-H flags for env/headers.

Features

Mermaid flowcharts now render subgraph blocks as titled frames with correct internal and cross-boundary edges.
Class diagrams in Mermaid now render as proper UML boxes with attributes, methods and inheritance arrows instead of raw source.
Permission prompts now accept a double-click on an option to submit it, matching the existing Enter and number-key shortcuts.
New /code-review slash command now ships with the CLI and is always available.

Bug Fixes

Plan mode exit reminders no longer appear after the model has already started implementing the plan.
Expanded thinking blocks in scrollback now remain expanded when the agent finishes them.
grok update no longer downloads the same binary twice when multiple updaters or leader checks run concurrently.
Background task IDs after /compact are now shown verbatim so the model can reference them correctly in later tool calls.
Typing / while scrollback is focused now focuses the prompt and opens the slash-command dropdown.
Dashboard empty state is now a single hint line; dispatch and peek placeholders appear only when unfocused.
Fixed memory leaks that could cause the CLI to use tens of gigabytes during long sessions with many tool calls.
Login on SSH or headless machines now tells you when the browser cannot be opened automatically and shows the URL to visit manually.
Fixed git clone failures on Windows when the CLI tries to clone marketplace plugins into ~/.grok.

Performance

Large code blocks inside lists no longer cause multi-second UI stalls while streaming responses.

0.2.50

Features

Mermaid flowcharts now render edge crossings clearly instead of fusing unrelated connections.

Bug Fixes

Sequence diagrams with activate, autonumber, par, and more now render instead of showing parse errors.
MCP servers menu and slash commands now work when starting grok outside a project directory.
Ctrl+W in the prompt now deletes whole words like bash instead of stopping at punctuation.
Login no longer quits when an authentication code contains the letter q.

0.2.49

Features

Marketplace plugin listings now show skills, MCP servers, and commands when the catalog is published.
Mermaid flowcharts now render with fewer avoidable edge crossings.
stateDiagram mermaid blocks now render as Unicode diagrams instead of source fallback.

Bug Fixes

Skill reloads no longer corrupt active tool calls or produce duplicate results in the conversation.
grok --resume now correctly finds the real session instead of failing on empty image-only folders.
Pasted images and relative paths now use the correct directory when resuming a session created elsewhere.
Mermaid flowcharts now correctly render node groups, arrow endings, self-loops and line styles.
Fixed "unknown session id" errors that occurred after the leader process crashed or was killed.
Pasted images now survive interjections and queue edits instead of being dropped.
Managed MCP connectors (Slack, Linear, etc.) now appear correctly when using leader mode.

0.2.47

Features

stateDiagram Mermaid blocks now render as diagrams instead of source fallback.

Bug Fixes

Pasted images now survive interjections and queue edits instead of being dropped.
Managed MCP connectors (Slack, Linear, etc.) now appear correctly when using leader mode.

0.2.46

Features

Mermaid flowcharts now render with fewer avoidable edge crossings.

Bug Fixes

Fixed grok --resume failing on empty image-only session folders left by cross-directory pastes.
Fixed pasted images and relative paths using the wrong directory after cross-cwd resume.
Fixed Mermaid flowcharts that silently rendered wrong diagrams for & groups, circle/cross endings and self-loops.
Fixed zsh tab-completion for subcommands after the optional prompt argument was added.
Fixed "unknown session id" errors after the leader process crashed or was killed.
Fixed repeated auto-compaction attempts when the session is credit-blocked or auth is non-refreshable.

Performance

Parallel tool calls on the same path (multiple greps etc.) now execute concurrently.

0.2.45

Features

Mermaid diagrams now render to images when you click Open in a code block (on by default).

Bug Fixes

Fixed rare conversation corruption when skills changed while a tool call was still running.
Fixed grok --resume failing on empty image-only session folders left by cross-directory pastes.
Fixed pasted images and relative paths using the wrong directory after resuming a session from another folder.
Welcome screen logo no longer renders as invalid characters on legacy Windows command prompts and PowerShell.
Fixed "unknown session id" errors that occurred after the leader process crashed or was killed.

0.2.44

Features

K/J keys now snap the viewport to the top of previous or next assistant responses.
J/K (vim mode) now navigate between assistant responses in scrollback.
sequenceDiagram mermaid blocks now render as Unicode lifeline diagrams instead of source fallback.

Bug Fixes

Interjecting while editing a queued prompt no longer strands the composer or blocks the queue.
Mid-turn interjections now appear as separate user messages instead of being appended to tool results.
Project MCP config touches no longer trigger repeated reload storms.

Performance

Inference requests recover faster from silent engine stalls instead of waiting the full idle timeout.

0.2.43

Bug Fixes

ask_user_question tool can now be enabled in allowlists without requiring plan-mode tools.
Shift+Tab mode cycling (Normal → Plan → Auto-Approve) works again in the agent view.
Ctrl+C now cancels a blocking grok update cleanly instead of leaving an orphaned download repainting the terminal.

0.2.42

Bug Fixes

ask_user_question tool can now be enabled in allowlists without requiring plan-mode tools.
MCP servers provided at session start now persist across config hot-reloads.

0.2.41

Features

Compaction completion message now shows the before → after token reduction instead of only the final count.

Bug Fixes

Fixed token count after compaction so the displayed number no longer jumps back up on the next model response.
Fixed plugin skill loading when a manifest lists skill directories directly instead of a parent skills/ folder.

Performance

Fixed memory context injection on resume so the prompt prefix stays byte-stable and KV cache is preserved.

0.2.40

Features

grok --debug now produces per-session log files under ~/.grok/debug/ even with a leader process.

Bug Fixes

Doom-loop warnings now correctly describe cycles and distinct edit failures instead of claiming identical arguments.
Model list changes from config or cache now appear in already-connected TUI and IDE clients without restart.

0.2.39

Features

run_terminal_cmd can stream live stdout/stderr chunks when the workspace flag is enabled.
/session-info now displays the current turn index.
Server-synced and bundled skills are now discovered from launcher-injected directories.
Background & operator is now allowed by default in terminal commands.

Bug Fixes

Resumed subagents no longer loop forever during auto-compaction on large context windows.
Background task descriptions and & rejection messages now correctly name the real parameters.
Doom-loop detection no longer falsely triggers on distinct failing tool calls.

0.2.38

Features

Watching status line now appears when background monitors, loops, or subagents can wake the agent.

Bug Fixes

Default model selection now correctly chooses the intended entry when multiple models share a slug.
Minor bug fixes

0.2.37

Features

MCP tool result queries now list only command-line tools actually present on your system.
grok update now restarts any older running leader so all clients get the new binary.
Long-running bash commands that hit the timeout are now moved to the background by default instead of killed.

Bug Fixes

Subagents now correctly receive web_search and x_search tools from the parent session.

0.2.36

Features

Large MCP tool results are now saved with the correct extension and the model receives better hints for querying them.

Bug Fixes

Fixed false-positive doom-loop terminations when many parallel tool calls fail together in one batch.
Fixed a crash that could occur during auto-compaction when resuming a session containing reasoning content.

0.2.35

0.2.34

Features

grok login now defaults to device code flow, which works reliably in SSH, WSL, VPN, and browser-restricted environments.

Bug Fixes

Fixed a hang during auth refresh.

0.2.33

Bug Fixes

Fixed duplicate turn output when attaching a second client to an active leader session.
Fixed Send now on queued prompts.

0.2.32

Features

Slash commands from project plugins now appear correctly in every open conversation after a plugin change.

Bug Fixes

Prompts submitted rapidly now stay in correct submission order in the queue.

Performance

Grep searches on large repositories are now substantially faster and no longer hit the 60-second timeout.

0.2.31

Bug Fixes

Marketplace skills without proper descriptions are now hidden from listings instead of flooding the model with tables.
Prompts submitted rapidly now stay in correct submission order in the queue.

Performance

Grep searches on large repositories are now substantially faster and no longer hit the 60-second timeout.

0.2.30

Features

A new plugin install suggestion appears above the prompt when you type a known marketplace plugin name or domain.

Bug Fixes

Trace uploads and remote session restores now succeed with a deployment key and no browser login.
Resumed sessions no longer pad the sticky prompt with empty rows; cancelling a turn now keeps the rest of the prompt queue intact.
Cancelling a running prompt no longer leaves the interface stuck on the cancelling spinner.

0.2.29

Bug Fixes

/rewind before a compaction boundary no longer leaves later prompts in context.

Performance

Resuming large sessions is now substantially faster with no data loss.

0.2.28

Bug Fixes

Images read via read_file are now downscaled even when small in bytes but large in pixels.

0.2.27

Features

Image and video generation tools now include the saved filename and session folder in their output.

Bug Fixes

Monitor output no longer appears as raw XML in the conversation view during leader sessions.
Windows commands containing & are no longer incorrectly rejected by run_terminal_cmd.
Python -c save-to-file reminder now suggests correct commands on Windows.

0.2.26

Bug Fixes

Large pasted content no longer triggers context-window errors or breaks compaction and memory flush.
API-key users can now run grok agent --leader without forced interactive login or timeouts.
Compaction no longer retries endlessly on credit, size, or auth failures; shows a clear message instead.
Windows PowerShell and cmd.exe no longer falsely reject commands containing &.
web_fetch no longer crashes the CLI on pages whose root element matches a cleaning selector.

0.2.25

Bug Fixes

Session titles now generate reliably even for very long initial messages.

0.2.24

Bug Fixes

Minor bug fixes

0.2.23

Features

Leader sessions can now be viewed and controlled from multiple clients with a live dashboard.
Sessions can now be deleted directly from the /resume history picker.

Bug Fixes

MCP plugin servers with bundled OAuth client IDs now authenticate correctly.

0.2.22

Bug Fixes

Authentication errors with static API keys now surface a clear error instead of hanging the turn.

0.2.21

Features

allowed_models in config.toml now restricts which models appear in the picker and /model command.

Bug Fixes

Code navigation now returns correct results for secondary project windows with different working directories.

0.2.20

Bug Fixes

MCP servers declared in both a plugin's.mcp.json and plugin.json are now registered instead of dropped.
Git operations now correctly target the repository for each session's working directory.

0.2.19

Features

Monitors now appear labeled in background-task reminders after compaction and can be terminated by name.

Bug Fixes

Reading images with text-only models no longer triggers repeated 400 errors that brick the session.

0.2.18

Features

Official xAI plugin marketplace now appears automatically in the Marketplace tab on first launch.
Image and video generation now use api.x.ai directly for all users.
New image-to-video and reference-to-video tools are now available for generating videos from images.
New imagine skill provides prompt-craft and workflow guidance for image generation and editing tools.

Bug Fixes

image_edit now correctly resolves pasted or attached images referenced as [Image #N].
Background subagent completions are no longer reported twice when the agent is idle.
Subagents now use the same model as the parent session by default.

0.2.17

Features

Image and video generation tools now emit structured paths so the pager renders media without regex scraping.
Compaction summaries now use a more detailed structure that improves recovery after context reset.
image_gen can now be enabled via the harness model using [features] in config.toml or the GROK_IMAGE_GEN_HARNESS env var.
Improved config refresh on new sessions from the shell.

Bug Fixes

--restore-code no longer detaches the source repository when resuming a forked-worktree session from a different directory.
Read tool string coercion bug fixes.
ICO images pasted or read from disk are now automatically converted to PNG before being sent to the model.

0.2.16

Features

New segments compaction mode writes per-segment markdown files that the model can read to recover pre-compaction detail.
Claude and Cursor compatibility scanning (skills, rules, AGENTS.md) can now be toggled individually via env vars or config.toml.
grok inspect now shows the resolved on/off state and source for every Claude/Cursor compatibility toggle.
Cursor MCP servers and hooks are now discovered and can be disabled independently via GROK_CURSOR_MCPS_ENABLED / GROK_CURSOR_HOOKS_ENABLED.

Bug Fixes

Streaming tool output (bash/write_file) now renders completely in the pager instead of only the latest chunk.
Streaming bash tool output now appears correctly in the pager scrollback.
Routing a native tool (e.g. scheduler_create) through use_tool now gives a clear corrective error instead of an unrecoverable loop.
"Starting session..." spinner no longer gets stuck when zero MCP servers are configured.
Subagents now use the correct harness after switching models mid-session.
Fixed long startup delays when an external auth provider binary hangs or fails.
Subagent conversations no longer receive unrelated monitor events or background task completions from the parent.
The /loop command now accepts natural-language intervals instead of always defaulting to 10 minutes.
Fixed blank output on completed bash or code-execution cards after shell restart or reconnect.

Performance

Large pasted images no longer bust the prompt cache or exceed the 50 MiB request limit.

0.2.15

Features

Permission prompts now remember your last choice across tools and let you configure the first-prompt default in config.toml.

0.2.14

Features

Generated images and videos can now be opened directly from the terminal UI via buttons or clicks.
Background tasks panel now groups items, supports collapsible sections, and has clearer styling for monitors and loops.

Bug Fixes

Session titles are now generated reliably using a fixed default model.
--permission-mode now correctly overrides the permission_mode setting from config.toml when launching sessions.

0.2.13

Bug Fixes

Miscellaneous bug fixes

0.2.12

Features

Computer connection status now shows a connecting pill during terminal session initialization.
/check and subagents now read and follow full AGENTS.md rules from the repo.

Bug Fixes

--max-turns now correctly counts tool-use cycles instead of total messages.
@-mention file search now works again for local agent sessions.
Rendered images, files, and citations now replay correctly in chunk-mode history.
/context now displays the correct auto-compact threshold for the active model instead of always 85%.
Model responses are no longer silently dropped when the gateway emits legacy channel values.
Prompt responses no longer resolve before the turn's final output chunks reach the client.

0.2.11

Bug Fixes

Minor bug fixes

0.2.10

Features

/check has been renamed to /check-work; old command continues to work during transition.

Bug Fixes

Images smaller than 8×8 pixels are now rejected with a clear message instead of producing blocky results.

0.2.9

Features

Added --device-code as alias for device authentication and improved headless auth error messages.

0.2.8

Features

New /login slash command lets you re-authenticate from within a session without quitting.
Compaction summaries now include the full transcript path so the model can reference prior details.
Cursor skills and rules are now discovered alongside Grok and Claude directories.

Bug Fixes

Fixed monitor tool schema to show the correct 10-hour default timeout.
Fixed a panic that could occur when installing marketplace plugins.

0.2.7

Features

Image generation and image editing can now be toggled independently via [features] in config.toml.

Bug Fixes

Background tasks started inside subagents now continue running after the subagent session ends.

0.2.6

Bug Fixes

Background tasks started inside subagents now continue running after the subagent session ends.
Image description now reliably uses the grok-build model instead of falling back to the active session model.

0.2.5

Bug Fixes

Drag-and-drop and pasting images or files now works correctly on Windows.

0.2.4

Features

image_gen now uses the higher-quality grok-imagine-image-quality model.

Bug Fixes

read_file now correctly passes embedded base64 images to the model as vision tokens instead of truncated fragments.

0.2.3

Features

Memory system: /remember command, note modal with raw/enhanced preview, x.ai/memory/rewrite ACP extension, Ctrl+F fullscreen toggle for /memory modal.
Agent configuration: /config-agents modal with agents, personas, and defaults.
Goal classifier: end-to-end goal tracking with subagent-powered classification.

0.2.2

0.2.1

Bug Fixes

Pasting or dropping images now succeeds for truncated, CRC-corrupt, or tiny files instead of failing silently.

0.2.0

Performance

Large chat sessions now use substantially less memory and run faster during forks, rewinds, and compaction.

0.1.220-alpha.4

Features

Version output now shows [alpha] or [stable] in --version, banners, /session-info, and grok inspect.
New "always-approve" option appears first in permission dialogs. Bash commands now resolve reliably on NixOS and Homebrew installs.

Bug Fixes

Auxiliary model calls such as session titles and image descriptions now succeed for API-key users.

0.1.220-alpha.3

Bug Fixes

Pasting or dropping files from Finder and handling multiple images now works reliably on macOS.
Clipboard copy and paste now works reliably on Linux Wayland desktops.

0.1.220-alpha.2

Bug Fixes

Web and X search tools now activate correctly after switching to a model that supports backend search.

0.1.220-alpha.1

Features

Release notes are now fetched from the CDN instead of being bundled inside the binary.

Bug Fixes

--restore-code now stashes dirty changes before restoring and shows clear success or failure messages.
Shell background tasks now create their output file immediately and AwaitShell exits early on pattern match.

0.1.220

Features

Model selection now respects API-key permissions and defaults auxiliary calls like summaries to your current model.
Release notes command now loads the latest notes from the network instead of bundled files.
Version output now shows [alpha] or [stable] labels based on the release-channel pointer.
Permission prompts now offer an always-approve option; shell commands resolve more reliably across installs.

Bug Fixes

--restore-code now safely stashes your uncommitted changes before restoring and shows clear success or failure messages.
Pasting or dropping files from Finder now works reliably and images no longer lose their numbers or chips.
Background shell output is now available immediately after starting and AwaitShell exits early when a pattern matches.
Copy to clipboard now works reliably on Linux Wayland desktops using wl-copy or xclip fallbacks.
Web and X search now activate correctly after switching to a backend-search model without restarting the session.

0.1.219-alpha.1

Bug Fixes

read_file now always returns the file content instead of a duplicate-read message after compaction.

0.1.219

Bug Fixes

read_file tool now always returns file content instead of a duplicate-read message after compaction.

0.1.218-alpha.1

Features

Project instruction files now recognize.claude/CLAUDE.md and.claude/CLAUDE.local.md.

Bug Fixes

run_terminal_command on Windows PowerShell no longer suggests unavailable Unix utilities.

0.1.218

Features

Project instruction files named.claude/CLAUDE.md are now discovered automatically.

Bug Fixes

Windows PowerShell users no longer see model suggestions for Unix commands that do not exist.

0.1.217-alpha.5

Features

New /export slash command and grok export CLI command let you save or copy conversation transcripts as Markdown.

Bug Fixes

Image pasting now correctly prefers images from the clipboard and keeps dragged screenshots visible instead of losing them.
Video generation no longer forces short 5-second clips; longer durations now work as requested.

0.1.217-alpha.4

Features

New grok plugin commands let you manage plugins and marketplaces directly from the terminal.

Bug Fixes

Image pasting now correctly prefers images from the clipboard and keeps dragged screenshots visible.
Subagents now correctly use bring-your-own-key credentials when the parent model specifies them.
MCP server connections no longer fail with initialization errors under concurrent tool calls.
Signal sync is more resilient to transient authentication hiccups.
Authentication recovery after subscription changes no longer causes duplicate IdP calls.

0.1.217-alpha.3

Features

Bug Fixes

Image description now works correctly when using custom models with bring-your-own-key configs.
Subagents now correctly show the parent model's display name instead of the internal routing slug.

0.1.217-alpha.2

Bug Fixes

Model picker now correctly shows all entries including aliases when multiple models share the same slug.

0.1.217-alpha.1

Features

Bug Fixes

Tool calls now appear in the output as soon as they start instead of waiting until they finish.

0.1.217

Features

Todo list reminders now appear at turn end when pending tasks remain after compaction.
Laziness detection can now be enabled per-model to automatically nudge idle agents.
New grok plugin commands let you manage plugins and marketplaces directly from the terminal.
New /export command and grok export let you save or copy conversation transcripts as Markdown.

Bug Fixes

Pasting images from the clipboard or dragged screenshots now works reliably without temporary paths or duplicates.
Subagents now correctly use per-model API keys instead of falling back to OAuth when both are present.
Model picker now shows all models including aliases like "auto" when multiple entries share a slug.
Tool calls now appear in the scrollback immediately when dispatched.
Cursor image and PDF handling now processes attachments reliably without errors.
Signal tracking for sessions is now more resilient to temporary login hiccups.
Subscription detection now correctly respects access gates from the server.
Image description now works reliably with custom models and bring-your-own-key configs.
Subagents now show the correct parent model name instead of an internal routing slug.
Authentication recovery after subscription changes no longer causes duplicate login attempts.
Video generation no longer forces short ~5s clips when no duration is specified.
Fixed excessive authentication errors that could occur during token refresh in background operations.

0.1.216-alpha.1

Features

Grok now reminds itself to take action in very long conversations while still obeying instructions like "just plan".

0.1.216

Features

Grok now reminds itself to execute plans in long conversations while strictly following instructions like "just plan".

Bug Fixes

Image pasting on Linux now works reliably when the clipboard contains images.

0.1.215-alpha.1

Features

Default models now use plan mode tools (enter/exit plan mode, ask user) unless overridden.

Bug Fixes

Skills now resolve relative links to bundled files like reference docs inside the skill directory.
Subagents no longer cause the parent session to lose write or execute tools.

0.1.215

Features

Default agent behavior now uses plan mode tools for models without an explicit agent_type.

Bug Fixes

Relative links inside skill files now resolve to absolute paths automatically.
Subagents no longer cause the parent session to lose access to tools.

0.1.214-alpha.1

0.1.214

Features

Bug Fixes

Image and video generation no longer causes the model to repeat internal instructions about inline display.
Sessions now resume reliably even if the original model is removed from the catalog.
Fixed crashes when tool output or prompts contain multibyte characters like CJK or emoji.
Web search now correctly switches between local and hosted tools when changing models.
Custom agent profiles from clients are now respected for default models.

0.1.213-alpha.7

Features

New settings pane lets users configure appearance, themes, and behavior directly in the terminal UI.
New image_edit tool supports reference-based editing using uploaded photos or data URLs.
Goal mode now surfaces verification-blocked pauses with human-readable reasons and resume guidance.
Added grok logout subcommand to sign out from the terminal without launching the TUI.
Backend search is now enabled by default for web_search and x_search.
Added per-MCP-server expose_image_base64 option so raw image data remains available in tool output.
Improved credit-limit experience for max-tier users with an inline scrollback card instead of a modal.
Automatically detects mid-session subscription upgrades after a credit-limit error.

Bug Fixes

Goal mode now auto-pauses on repeated failures or doom loops and resumes cleanly without errors.
Fixed sessions in directories with very long or CJK names that previously failed with "File name too long".
Fixed /btw errors on Anthropic opus models caused by orphaned tool calls and temperature conflicts.
Doom-loop halts now correctly report harness termination instead of misleading "user cancelled" messages.
API-key users no longer see unavailable models such as grok-build that 404 on selection.
Image and video generation now recover better from network issues and large response bodies.
Improved error messages when hitting a subscription wall with an API key set, suggesting grok logout.
Removed --budget flag from /goal and made goal verification stricter.
Fixed a crash on exit when using MCP servers that require token refresh.
Fixed follow-up errors after backend search and prevented unsupported tools from reaching certain models.

0.1.213-alpha.6

Performance

Large sessions now upload and process without long stalls or excessive memory use.

0.1.213-alpha.5

Features

Web search and X search can now execute server-side with full session persistence when GROK_BACKEND_SEARCH is enabled.
read_file can now append configurable rule reminders for matching.cursor/rules files when enabled.
read_file now extracts text from PowerPoint (.pptx) files in addition to PDF and images.

Bug Fixes

run_terminal_cmd now reports clearer "exit: killed (reason)" headers for processes terminated by timeout or cancellation.
Fixed model errors when switching to non-reasoning models after previously setting reasoning effort.
Compaction summaries no longer cause the model to echo the original summarization instructions after resume.

0.1.213-alpha.4

Features

New /goal slash command for tracking multi-step coding goals with live status and automatic pausing on Ctrl+C.
Goal mode is now opt-in via config.toml or the GROK_GOAL environment variable.
Commands in workspace/proxy mode now preserve environment variables, cwd, functions, and aliases across calls.
Background tasks now report termination signals, reject self-killing patterns, and support unbounded timeouts.
Grok CLI now runs natively on Intel Macs via x86_64 builds.

Bug Fixes

Goal mode prompts now work on all platforms and correctly place source code in your project directory.
Background tasks now correctly trigger auto-wake even when a blocking wait times out.
Goal status bar now shows the correct deliverable number matching the detail modal.
Task tool now accepts model-emitted paths containing stray quotes or a leading tilde.
Subagent task output no longer appears twice in the scrollback or leaves trailing system reminders.
Failed background subagents now report errors and appear in task lists instead of silently disappearing.
Subscription purchases now correctly refresh authentication so API calls succeed after the paywall lifts.

Performance

MCP tool output is now truncated earlier to reduce wasted context when models rarely follow up.

0.1.213-alpha.3

Bug Fixes

Credit status now updates live after topping up or changing spending limits when at your usage cap.
Paywall now lifts automatically after subscribing via grok login or external purchase without needing to re-login.
Images from /imagine no longer render twice in the scrollback.

0.1.213-alpha.2

Bug Fixes

Fixed terminal UI corruption on macOS that occurred when system malloc messages interleaved with the interface.

0.1.213-alpha.1

Features

Marketplace plugins from URLs can now be pinned to a specific commit SHA for tamper protection.
Terminal output now shows both the start and end of long command results instead of only the tail.
/imagine now works reliably after session reload and shows images directly in tool results.
Images and PDFs read with the file tool now appear immediately instead of in a later message.

Bug Fixes

Plugins and Marketplace reloads now show a clear loading indicator instead of misleading per-item badges.
Search tool now gives the model better guidance for finding MCP tools by server and action name.
Windows users now get a more reliable shell (PowerShell preferred) that avoids path mangling.
Model list no longer gets stuck on only grok-build after waking from sleep on macOS.
Custom xai_api_base_url settings in config.toml now correctly route session commands for enterprise deployments.

0.1.213

Features

/imagine now renders generated images inline in tool results.
Images from read_file now appear inline inside tool results instead of a follow-up message.
Goal mode (opt-in) now tracks objectives and deliverables with live progress updates via the UpdateGoal tool.
New /goal slash command lets you set objectives, track deliverables, pause/resume, and receive live progress updates during long tasks.
New Settings pane (consolidated) for configuring appearance, themes, and general options directly from the UI.
Goal mode is now opt-in via [features] goal = true in config.toml or the GROK_GOAL environment variable.
Rewind (Ctrl+Z) now works in proxy/hub mode by routing file-state tracking through the remote workspace.
Paywall now lifts automatically after subscribing via grok login or external purchase without needing to re-login.
Commands in workspace/proxy mode now preserve environment variables, cwd, functions, and aliases across calls.
New /goal slash command for tracking multi-step coding goals with live status and automatic pausing on Ctrl+C.
Grok CLI now runs natively on Intel Macs via x86_64 builds.
New image_edit tool supports editing images from reference photos.
Goal mode now supports a "verification blocked" pause state with a user-visible reason.
read_file now extracts text from PowerPoint (.pptx) files in addition to PDF and images.
Goal mode verifiers now act as reviewers and subagent resumes use the latest ID.
New grok logout command clears your cached login session directly from the terminal.
Web search and X search now run server-side by default for improved results.
MCP image tools can now expose raw base64 data so agents can forward images via file tools.
Max-tier users now see an inline message with pay-as-you-go options when hitting credit limits.

Bug Fixes

Windows shell detection now prefers PowerShell and prevents Git Bash path mangling.
Fixed model list getting stuck on only grok-build after waking from sleep on macOS by refreshing the catalog on every auth token update.
Fixed cross-platform prompt text and ensured goal-mode workers write source code in the project workspace rather than temporary directories.
Enterprise config fix: setting xai_api_base_url in config.toml now correctly routes session commands without also needing to set cli_chat_proxy_base_url.
Fixed hub disconnects so the CLI degrades gracefully instead of panicking when the Computer Hub WebSocket drops.
Doom-loop terminations now show "Agent was unable to make progress — turn ended in Xs." instead of the misleading user-cancel message.
Fixed terminal UI corruption on macOS that occurred when system malloc messages interleaved with the interface.
Images from /imagine no longer render twice in the scrollback.
Goal status bar now shows the correct deliverable number matching the detail modal.
Task tool now accepts model-emitted paths containing stray quotes or a leading tilde.
Subagent task output no longer appears twice in the scrollback or leaves trailing system reminders.
Background tasks now report termination signals, reject self-killing patterns, and support unbounded timeouts.
Failed background subagents now report errors and appear in task lists instead of silently disappearing.
Subscription purchases now correctly refresh authentication so API calls succeed after the paywall lifts.
run_terminal_cmd now reports clearer "exit: killed (reason)" headers for processes terminated by timeout or cancellation.
Fixed model errors when switching to non-reasoning models after previously setting reasoning effort.
Compaction summaries no longer cause the model to echo the original summarization instructions after resume.
Fixed /btw 400 errors on certain Anthropic models by cleaning up mid-turn tool state.
API key users no longer see unavailable models like grok-build in the model list.
Image and video generation now handles slow or interrupted network responses more reliably without generic decode errors.
Fixed error messages when hitting a subscription wall with an API key set, now suggesting grok logout.
Fixed a crash on exit when using MCP servers that require token refresh.
Reduced 400 errors on background flush/dream operations for thinking models.
Fixed follow-up search errors and ensured backend search only activates on supported models.
Subscription tier upgrades are now detected automatically after a credit limit error without needing to log out.

Performance

Large sessions now upload and process without long stalls or excessive memory use.

0.1.212-alpha.5

Bug Fixes

Fixed /btw side questions failing with 400 errors on models that use extended thinking.

0.1.212-alpha.4

Breaking Changes

Subagents now always start fresh and must be fully briefed via the prompt instead of inheriting parent conversation history.

Features

Memories now avoid storing low-value tool counts and file lists while skipping near-duplicate entries via semantic similarity checks.

0.1.212-alpha.3

Features

Added configuration options for the auto-compaction threshold on a per-model or global basis.
Shows 'Logged in with API key' indicator on the welcome screen when using an API key.

Bug Fixes

Improved error messages when calling use_tool with an invalid or built-in tool name.
Fixed context window incorrectly dropping to 128k for some models, preventing mid-session cascading compactions.
Fixed incorrect context windows for models referenced by multiple catalog entries with different keys.

0.1.212-alpha.2

Features

Linkify URLs and file paths across all TUI output
Improve /sessions modal and /new /fork banners
Improve review plan UX in the TUI

Bug Fixes

Use multipart upload for S3 payloads exceeding 8 MiB

0.1.212-alpha.1

Features

web_fetch now downloads and saves images and videos to disk (with magic-byte checks) instead of rejecting them.

0.1.212

Breaking Changes

Subagents no longer support fork_context; they always start fresh and must be fully briefed via the prompt.

Features

Images and videos now display directly in the scrollback with click-to-play support.
Skills now show their author in the extensions list and search results.
web_fetch now downloads and saves images and videos to disk instead of rejecting them.
Added configuration options for the auto-compaction threshold on a per-model or global basis.
Shows 'Logged in with API key' indicator on the welcome screen when using an API key.

Bug Fixes

Long bash task output no longer inflates context; full logs remain available via disk pointers.
Auth file locking is now more robust against hung processes and concurrent refreshes.
Concurrent logins and refreshes now wait efficiently without timeouts or retry storms.
Improved error messages when calling use_tool with an invalid or built-in tool name.
Fixed context window incorrectly dropping to 128k for some models, preventing mid-session cascading compactions.
Fixed incorrect context windows for models referenced by multiple catalog entries with different keys.
Improved memory quality by removing noisy auto-saved details and avoiding duplicate stored memories.
Fixed /btw errors on models that use extended thinking by excluding internal reasoning from side-question requests.

Performance

search_tool now instantly finds exact tool names instead of falling back to fuzzy search.
search_tool now better matches compound tool names like grafana-ai__SearchDashboards.

0.1.211-alpha.3

Features

New browser tools for tab and network inspection.
Browser tools security hardening with cookie mode and domain allowlisting.
Plugin install shorthand now accepts owner/repo format.
Memory modal now renders markdown with mouse interaction and scrollbar.
Keyboard shortcut dialog (Ctrl+.) redesigned for discoverability.
Ask-question modal UX with double-click selection and copy support.
Loading spinner added to image viewer overlay.
PowerShell install scripts added for Windows users.

Bug Fixes

Fixed diff line numbers to show correct absolute positions.
Fixed tasks stuck in killing state in tasks pane.
Privacy opt-out now persists across token refresh.
Fixed infinite restart loop during auto-update failures.
Gitignored files now readable by default.
Fixed auth refresh storm on expired tokens.
/privacy opt-out now works for personal OIDC users.
Model catalog now retries after sleep/resume.
Fixed terminal rendering leak on exit.

0.1.211-alpha.2

Bug Fixes

Login now handles corrupt authentication files by backing them up instead of prompting repeatedly.
Fixed loops getting stuck when resuming sessions.
Stopped duplicate notifications for background tasks and subagents waited on directly.
Loop indicators now disappear properly when subagents shut down.

0.1.211

Features

Added browser tools for navigation, screenshots, and network inspection via Chrome.
Background tasks and subagents now automatically notify the agent when they finish.
Browser tools now use safer cookie handling and domain restrictions by default.

Bug Fixes

Skills now reliably apply persona instructions to subagents via shared definitions.
Notifications from background tasks now render correctly without duplication.
Login now recovers from corrupted auth files instead of failing repeatedly.
Loop tasks no longer get stuck when resuming previous sessions.
Directly waited background tasks no longer trigger extra notifications.
Loop indicators are now properly removed when subagents shut down.
run_terminal_cmd tool name restored; background parameter rename preserved.
Model picker now recovers automatically after sleep/resume or temporary network issues instead of staying stuck on Grok Build.
MCP server connections now succeed for servers that enforce audience binding on OAuth tokens.
read_file now accesses gitignored files by default (grep/list_dir/search_replace still block them unless configured).
--no-ask-user flag now correctly disables the ask_user_question tool in both TUI and headless modes.

0.1.210-alpha.6

Bug Fixes

File writes on Windows now automatically retry if briefly locked by editors or antivirus.
Login process now waits up to 10 minutes instead of 5 and shows a clearer timeout message.
Terminal interface no longer gets garbled by subprocesses like git or npm.
Plan mode no longer activates unexpectedly from tool descriptions containing similar phrases.
Authentication now recovers immediately when using multiple terminals after token refreshes.

0.1.210-alpha.5

0.1.210-alpha.4

Bug Fixes

grok sessions search now finds local and remote sessions correctly, and grok -r <id> restores remote sessions reliably.

0.1.210-alpha.3

0.1.210-alpha.2

Features

Tool permission prompts and lists now clearly show server name followed by action.

0.1.210-alpha.1

Bug Fixes

Model list no longer collapses to default after long inactivity or re-login.
Session restore no longer crashes on Windows after system reboot.
Fixed authentication failures after token refresh errors.
Completed background tasks now remain queryable via get_task_output.

Performance

Interactive sessions now initialize MCP servers progressively for faster startup.

0.1.210

Features

Faster startup for interactive sessions with progressive MCP loading.
MCP tools now show as 'Server: Action' in permission screens.

Bug Fixes

Full model list now survives long inactivity, logins, and token refreshes.
Fixed crash restoring sessions on Windows after system reboot.
Windows file saves now retry brief editor or AV locks automatically.
Completed background tasks now remain queryable via get_task_output.
Fixed terminal garbage during auto-updates on macOS.
grok sessions search now finds local and remote sessions correctly, and grok -r <id> restores remote sessions reliably.
Login timeout increased to 10 minutes with clearer error message.
Fixed pager getting stuck in plan mode due to tool titles mentioning plan commands.

0.1.209

Features

Read file now renders PDF pages as images with optional page range.
PDF files now render as images when read by the AI assistant.
PDF files can now be read as text by specifying pages='text' in the read tool.

Bug Fixes

Subagents now show as cancelled when you reject dangerous commands like rm -rf.
Model selection no longer resets on login or token refresh.
Background tasks now wait for completion when block=true is requested.
Subagent summaries now show accurate tool calls and turns.
AI now sees completed background bash tasks between turns.
Sessions with content now appear in desktop sidebar after restart.
AI warns about duplicate background tasks before starting new ones.
Resumed sessions now inform the AI about any background tasks still running from before.
Background tasks automatically resume the conversation when they finish, even if idle.
Python background scripts now stream output in real-time without buffering delays.

0.1.208-alpha.2

Features

PDF files can now be read to extract and display rendered page images.
PDF files now render as page images when using the read_file tool.
Extract plain text from PDFs using pages="text" or format="text" in read_file.

Bug Fixes

Subagents cancelled by permission prompts now correctly reported to parent agents.
Custom model selections no longer reset on login, logout, or token refresh.
Blocking task output requests now wait reliably for completion.
Subagent stats in task output and session cards now accurate.
Background task completions now promptly visible to model between turns.
Saved sessions now appear correctly in desktop sidebar after restart.
Resumed sessions now inform model about previously running background tasks.
Model automatically wakes up when background tasks complete.

0.1.208-alpha.1

Features

Shell command suggestions from history and PATH now appear as ghost text.
File and directory completion now available in shell suggestions.
AI shell command suggestions enabled via GROK_SUGGESTIONS_AI=1 environment variable.
Plan approval status now clickable; new Abandon option exits plan mode.

Bug Fixes

Fixed restoring sessions from cloud without local cache.
Fixed image errors when loading old sessions with bad images.
Skill commands now display cleanly in history, titles, and search.
Cloud sessions now report correct OS, shell, paths, and git status.

0.1.208

Features

Shell command suggestions now appear as dim italic ghost text after the cursor.
File and directory path completion now available in shell suggestions.
AI-powered shell suggestions configurable via [suggestions] or env vars.
Plan status now clickable; new Abandon plan button and d key.

Bug Fixes

Remote session restores no longer fail with 401 auth errors.
Corrupt images in history are now skipped or replaced with placeholders.
Skill commands now display cleanly in history, search, and titles.
Cloud sessions now use correct remote OS, shell, paths, and git status.
Bundled skills like /check and /best-of-n now work in TUI sessions.

0.1.207

Features

New /check-work (/verify, /check) and /best-of-n (/bon) slash commands now work in headless mode.

Bug Fixes

Fixed session workspace setup so git, shell, and file operations no longer fail.

0.1.206-alpha.1

Features

Simple mode now persists across sessions in config.toml.

Bug Fixes

Git and bash commands now work reliably in the desktop app on Windows.
Slash command autocomplete no longer duplicates unique skill entries.
Git branch changes now update the status bar after shell commands.

0.1.206

Features

Simple mode now saved in config.toml across restarts.

Bug Fixes

Fixed git and bash commands failing in Grok Desktop on Windows.
Fixed duplicate slash commands in autocomplete dropdown.

0.1.205-alpha.1

Features

New /check-work and /best-of-n commands** for self-verification and parallel task solving.

Bug Fixes

Fixed authentication failures during file uploads.
Fixed terminal display corruption from mouse events and subprocess output.
Fixed sessions stuck on read-only filesystem after login.

0.1.205

Features

New /check-work and /best-of-n commands for self-verification and parallel solving.

Bug Fixes

Fixed authentication failures during file and artifact uploads.
Fixed terminal garbling from mouse events and child process output.

0.1.204-alpha.3

Features

Remote tools from hub-registered workspace servers are now discoverable and callable in sessions.
Worktree folders now use readable names like 'projects-my-repo' instead of session IDs.
Session lists now show AI-generated titles, git branches, repo names, and worktree labels.
grok -w my-worktree creates a custom-named worktree folder.
Session picker now groups entries by repository with headers and indents.
Remote workspaces now handle multiple sessions over one connection.
/usage manage now opens billing page; grok login enforces tiers correctly after refresh.

Bug Fixes

Fixed AI-suggested paths missing leading slash or quotes in tools like grep.
Fixed file upload failures during large batch transfers over unstable connections.

0.1.204-alpha.2

Bug Fixes

Error messages for model not found or auth failures now show available models, auth type, version, and login fix instructions.

0.1.204-alpha.1

Features

Subagents now enabled by default, fixing skills like /implement and /execute-plan.

Bug Fixes

Legacy web login no longer treated as first-party xAI account for sharing and billing; prompts grok login upgrade.
Billing display now includes pay-as-you-go status and monthly limits with better error handling.
Copy auth URLs now works in Docker containers and displays cleanly for external providers.

Performance

Large tool outputs shortened when sent to Grok, full text saved to file in session folder.

0.1.204

Breaking Changes

Legacy authentication removed; run grok login to upgrade to OAuth.

Features

Hub tools like read_file now discoverable and callable in pager sessions.
Worktree directories now named readably (e.g., projects-my-repo) instead of random IDs.
Session lists show AI-generated titles, branches, repos, and worktree labels.
Session picker now groups entries by repository for easier navigation.
Remote workspace servers supported via --hub-workspace-mode remote=<id>.
Subagents now enabled by default, fixing /implement and /execute-plan skills.
/usage manage opens billing page; /usage or show displays credits.

Bug Fixes

Legacy web login no longer treated as premium xAI account; run grok login instead.
Session picker removes duplicate blank sessions from same folder.
Sessions from deleted worktrees still appear in picker after restart.
Credit bar now shows pay-as-you-go status and monthly usage limits.
Auth URLs copy correctly in Docker containers and display cleanly for external providers.
Model/auth errors now show available models, auth type, version, and login fixes.
Tool path resolution fixed for absolute paths missing leading slash.
Fixed upload failures from HTTP/2 connection degradation.
Memory search now reliably includes global and workspace MEMORY.md files.

Performance

Large tool results shortened in chat; full text saved to file with path link.

0.1.203-alpha.2

Bug Fixes

/btw slash command now works correctly with all backends including Anthropic.

0.1.203-alpha.1

Features

New grok-build-v2 agent available in config.toml for prompt/tool testing.

Bug Fixes

Fixed duplicate MCP and skill reminders when resuming sessions.
Plan review overlay now always shown, even in always-approve mode.
Fixed empty model list and unknown model after login or logout.
Fixed unwanted login prompts during brief network issues at startup.
Fixed grok sessions list auth failures.

0.1.203

Features

New grok-build-v2 agent option in config.toml for testing prompt and tool changes.
Session picker /load deep search now shows spinner, content matches, and previews on expand.

Bug Fixes

Resumed sessions no longer duplicate MCP server and skill reminders.
Plan mode exits now always show interactive approval, even in always-approve mode.
Model picker now updates correctly after login/logout without empty lists or 'unknown'.
grok sessions list now works reliably after login.
/btw slash command now supports all backends like Anthropic Messages API.

0.1.202-alpha.1

Features

Added /usage command to view account usage and billing information.
Integrated Computer Hub client for remote computer-use sessions.
Added --hub-url flag to configure Computer Hub endpoint.
Implemented /rewind with inline overlay, message picker, and live dimming preview.
Added /memory slash command with browse and toggle routing.
Memory save notifications in scrollback with remember/forget/recall instructions.
Added grok memory clear CLI command with scope selection and confirmation.
Added /memory on|off to toggle memory mid-session.
Vision support for base64 images extracted from user queries.
Added --allow/--deny CLI flags with PolicyDeny variant and deny rule UX improvements.
Configurable parallel verifiers with improved verification prompt.
MCP connection progress UI with system reminders and progressive strategy fix.

Bug Fixes

Fixed session resume defaulting to legacy auth instead of OIDC.
Reject WebLogin tokens and force interactive login; fixed devbox auth.
Normalized skill names by converting underscores to hyphens.
Fixed auth revalidation to always check token validity.
Fixed memory search scoring, modal UX, and pager display.

0.1.202

Features

Add /usage command and billing extension fix
Computer Hub integration into xai-grok-shell
--hub-url flag and thread hub config to shell
Implement /rewind with inline overlay, picker, and live dimming preview
Stream tool-call argument deltas via the buffered path
LocalRegistry cloneable handle with shared dispatch across workspace
Add /memory slash command with browse and toggle routing
Memory save notifications in scrollback with remember, forget, and recall instructions
grok memory clear command with scope and confirmation
Support /memory on|off to toggle memory mid-session
Extract base64 images from user queries as vision tokens
--allow/--deny CLI flags with PolicyDeny variant and deny rule UX improvements
Configurable parallel verifiers with improved prompt
MCP connection progress UI with system reminders and progressive strategy fix
require_plan_approval config flag in pager
Best-of-N mode for grok-build
Anthropic Messages backend with streaming reasoning traces end-to-end

Bug Fixes

Fix terminal tool calls when agent runs via scheduled task on Windows
Fix resume defaulting to legacy auth instead of OIDC
Reject WebLogin tokens and force interactive login; fix devbox auth
Normalize underscores to hyphens in skill names
Fix auth to always revalidate
Fix memory search scoring, modal UX, and pager display

0.1.201-alpha.1

Features

New --self-verify flag starts automatic work-checking after each response.

Bug Fixes

Reasoning effort settings no longer incorrectly apply to unsupported models like grok-build.
Web search, image gen, and video gen tools now avoid 401 errors during auth provider gaps.
/share command now works in token's final 5 minutes and guides API key users to login.
/billing command now works reliably in token's final 5 minutes.
Deny rules in config now block actions even with --always-approve.
Skills now provide follow-up context automatically after use.

0.1.201

Features

New --self-verify flag starts sessions with automatic work-checking after each response.

Bug Fixes

Reasoning effort settings no longer incorrectly apply to models without support, like grok-build.
Web search, image generation, and video generation tools no longer fail with missing API key errors.
Deny rules in config.toml now block tool calls even with auto-approval enabled.

0.1.200-alpha.1

Features

grok-build model now defaults to Chat Completions API.

Bug Fixes

MCP tools now work more reliably by always fetching schemas before use.
MCP tool calls preserve all custom arguments without dropping extras.

0.1.200

Bug Fixes

MCP tools now work more reliably by enforcing schema lookup before calls.

0.1.197-alpha.5

Bug Fixes

MCP tools now always use correct parameters by requiring schema lookup first.

0.1.197-alpha.4

Features

/model command now lets you pick reasoning effort levels like high/medium/low for supported models.
Images in tool output now appear as pictures the AI can see, not long base64 strings.

Bug Fixes

Shell commands now work in Grok Desktop on Windows.
Image compression messages now explain if size, dimensions, or both triggered resizing.

0.1.197-alpha.1

Features

Agent profiles now support tool allow/deny lists, reasoning effort, turn limits, and reliable model overrides.
Custom models in config.toml now support both Bearer token and x-api-key authentication.
Built-in docx, xlsx, pptx skills now available for office document editing.
MCP tools now time out after 100 minutes by default instead of 1 minute.
MCP server reminders now include instructions to discover and use tools correctly.
New Verify mode (Shift+Tab) runs automatic checks and fixes after each agent turn.
Tool call arguments now stream progressively as generated by the model.

Bug Fixes

Agents auto-exit plan mode without prompts when started in always-approve mode.

0.1.196-alpha.6

Bug Fixes

Fixed interactive shell, command syntax, and git editor issues on Windows.

0.1.196-alpha.2

Features

New grok-build-orchestrator agent uses a lead model to coordinate subagents for builds.

Bug Fixes

Custom UI settings in config.toml no longer get lost when saving other changes.

0.1.196

Features

New grok import resumes Claude Code sessions with grok import <id/path>.
New grok-build-orchestrator coordinates subagents for complex builds.
/skills panel now includes toggles to enable or disable individual skills.

Bug Fixes

Discovers CLI tools from .zshrc/.bashrc/virtualenvs (e.g., tclips).
Custom UI settings like timestamps persist across config saves.
Model switches rebuild harness and hold prompts until complete.
Model list refreshes correctly after login or logout.
Subagents now execute tools in the specified working directory instead of always using the parent's.
Generated session titles now display correctly in /sessions list and terminal.
Fixed shell detection on Windows for interactive PTYs, command syntax, and git editors.
Fixed grok share failing with backend ZodError due to duplicate request headers.

Performance

Faster session startup on large repositories by backgrounding project info.

0.1.195-alpha.7

Features

New grok login --devbox option for signing in from headless devbox environments without browser access.
Skill file reads now display as 'Skill {name}' for cleaner scrollback view.

0.1.195-alpha.4

Features

Billing credits and usage now fetchable for display in pager/desktop.
New grok mcp doctor diagnoses MCP servers and suggests fixes.
Session picker now shows precise last activity times.

Bug Fixes

Fixed generation hangs on persistent authentication failures.

0.1.195-alpha.3

Bug Fixes

Image and video generation now works reliably after long idle periods.
Fixed crashes from large images read by the Read tool.
Improved Windows support for auto-updates, installation, and shell commands.

0.1.195-alpha.2

0.1.195-alpha.1

Bug Fixes

Terminal stays responsive when cancelling shell commands.
Fixed requests to Anthropic models from remote model lists.

0.1.195

Features

New grok mcp doctor diagnoses MCP server configuration and connectivity.
Subagents can now target specific directories without new worktrees.
Session picker now shows accurate last activity times.

Bug Fixes

Image and video generation now works reliably after long idle periods.
Large images read from files no longer crash API requests.
Terminal stays responsive when cancelling long-running shell commands.
Anthropic models from remote lists now route to correct API endpoints.
Fixed resume and fork crashes from old session files.
Fixed crashes from large images in tools and attachments.
Fixed generation hangs on persistent authentication failures.

0.1.194-alpha.1

Features

After plan approval, Grok suggests /implement for structured coding when available.

Bug Fixes

Model lists and pickers now show only models accessible via your GROK_CODE_XAI_API_KEY.

0.1.194

Features

Plan mode now recommends the /implement skill after user approval when available.

Bug Fixes

grok models and TUI picker now show team-specific models when GROK_CODE_XAI_API_KEY is set.

0.1.193-alpha.4

0.1.193-alpha.3

Features

Pressing Ctrl+C before any response now restores your prompt to the input box, ready to edit or resend.
Native Windows x86_64 support now available via install script and auto-updater.

Bug Fixes

Configured MCP tools now appear reliably in initial session context messages.

0.1.193-alpha.2

Bug Fixes

Fixed false-positive loop warnings for different shell commands during edit-test cycles.

0.1.193-alpha.1

Features

Native Windows support now available experimentally for running Grok without WSL.
Interactive menu for importing Claude settings lets you select items like permissions or MCP servers on welcome screen or /import-claude.

Bug Fixes

Image compression now skips small files and warns if large images can't be resized under 3.75 MB limit.
Feedback errors now appear in the session instead of failing silently.
Clearer login error messages now suggest grok login, API key env var, or config.toml.

0.1.193

Features

Native Windows support (opt-in) lets you run Grok CLI directly without WSL, with MCP server compatibility.
New interactive import for Claude settings (Ctrl-I or /import-claude) lets you pick permissions, env vars, MCP servers, hooks, or paths to migrate.
Cancel with Ctrl+C before any response to restore your unsent prompt (text and images) to the input box.
dontAsk permission mode now auto-rejects tools without prompting in headless sessions.

Bug Fixes

Large images (>3.75MB after decoding) now compress to JPEG efficiently, keeping originals if needed.
Authentication errors now clearly suggest grok login, env var, or config.toml api_key.
Fixed false-positive loop warnings for different shell commands during edit-test cycles.
Configured MCP servers now appear reliably in the initial session prompt.

0.1.192-alpha.3

Features

Import settings from .claude/ (permissions, env vars, MCP servers) into config.toml using startup prompt or /import-claude.
Claude migration now fully disables legacy .claude/ fallbacks after import.

Bug Fixes

Worktree errors now display the specific failure reason instead of generic messages.
Shell commands now appear immediately in the status bar during execution.

0.1.192-alpha.2

Breaking Changes

Temporarily removed the Auto option from the /model picker.

0.1.192-alpha.1

Features

Skills reload automatically when editing and saving SKILL.md files without restarting.

Bug Fixes

Slash commands like /loop now only appear when supported by the current model.
Fixed empty model hash display in status bar and feedback channels.
Fixed authentication errors for OIDC users with per-model API keys in config.toml.

0.1.192

Breaking Changes

Temporarily removed the Auto option from the /model picker.

Features

Import Claude settings from .claude/ into .grok/config.toml via welcome screen ('i') or /import-claude.
Skills reload automatically when editing/saving SKILL.md files without restarting grok.

Bug Fixes

Slash commands like /loop now only appear when the current model supports them.

0.1.191-alpha.6

Features

Skills reload automatically when you edit and save SKILL.md files without restarting.

Bug Fixes

Slash commands like /loop now only appear when supported by the current model.

0.1.191-alpha.5

0.1.191-alpha.4

0.1.191-alpha.3

Bug Fixes

Ctrl+C clears drafts without cancelling turns; press twice on empty prompt to cancel.

0.1.191-alpha.2

0.1.191-alpha.1

Bug Fixes

Context bar now reflects live token usage and matches /context.

0.1.191

Features

Improved grok login with automatic code delivery from browser consent page.

Bug Fixes

Context bar now accurately reflects live token usage matching /context.
Ctrl+C clears drafts without cancelling turns; press twice on empty prompt to cancel.

0.1.190-alpha.3

Features

New [[version_overrides]] sections apply config patches for specific CLI version ranges.

Bug Fixes

/context now updates correctly with token usage per conversation turn.

0.1.190-alpha.2

0.1.190-alpha.1

Features

New cli.minimum_version config enforces version floor and prompts updates.

0.1.190 0.1.189-alpha.4

0.1.189-alpha.3

0.1.189-alpha.2

Features

New slash command /feedback available in headless and stdio agent modes.
New error_reporting config enables Sentry independently of telemetry.

Bug Fixes

Fixed session cleanup race preventing errors in concurrent headless sessions.

0.1.189-alpha.1

0.1.189

Bug Fixes

Auth screen now has consistent styling and better spacing.
Fixed authentication errors when starting new sessions after long idle.
Prevents auth errors on welcome screen from background tasks during idle.

0.1.188-alpha.1

Bug Fixes

Fixed misleading auth errors for OIDC sessions with per-model API keys in config.toml.

0.1.188

Bug Fixes

Fixed misleading 'Invalid API key' errors for OIDC sessions with per-model credentials.

0.1.187

Features

Added reset button for tool permissions in desktop agent settings.
Added mid-turn interjection via Ctrl+Enter without canceling the turn.

Bug Fixes

Fixed OIDC session refresh when using per-model API keys in config.toml.
Fixed write denials on symlinked paths like /tmp after prior reads.

0.1.186-alpha.1

Performance

Reduced CPU spikes during git operations like pull and checkout.

0.1.186

Performance

Reduced CPU spikes from file watcher during git pull, checkout, and other operations.

0.1.185-alpha.2

0.1.185-alpha1

0.1.185 0.1.184-alpha.2

0.1.184-alpha.1

Bug Fixes

Large images now compressed client-side before API with user notification.
Plan mode reentry reminder now includes explicit exit instructions to prevent hangs.
Model unavailability on account changes now handled gracefully with auto-switch.
Session picker shows all sessions; remote fetches now have timeouts.
Credential prompts no longer corrupt TUI by detaching child processes from terminal.
Remote sessions now matched by repo URL across machines and protocols.
Fixed ghost text in input field from config warnings and log messages.

0.1.184

Bug Fixes

Large images now compressed client-side before API with user notification.
Model unavailability after account switch now handled with clear notifications.
Session picker now shows all sessions; remote loads have timeouts.
Credential prompts no longer corrupt TUI in terminal commands.
Remote sessions now matched by normalized repo URL across machines.
TUI input field no longer shows ghost text from config warnings and log messages.

0.1.183-alpha.1

Features

New grok trace command uploads or exports session data for debugging.
Slash command arguments now support fuzzy autocompletion with Tab.

0.1.183

Features

New grok trace command exports or uploads session data for debugging.

0.1.182-alpha.4

Bug Fixes

Model list refreshes correctly after account or team switch to prevent auth errors.
MCP tools no longer corrupt TUI display with credential prompts like GPG pinentry.

0.1.182-alpha.3

Features

grok models now lists available models and the default.
Permission prompts improved with numeric shortcuts and scope selector.

Bug Fixes

Same-file edits now serialize to prevent lost concurrent changes.

0.1.182-alpha.2

0.1.182-alpha.1

0.1.182

Features

grok models command now lists available models and authentication status.
Permission prompts now use numeric shortcuts, scope selector, and syntax highlighting.

Bug Fixes

Improved auth token refresh prevents 401 failures in long conversations.
Login checks now use shared client with 5s timeout to prevent hangs.
Concurrent edits to the same file now serialize to prevent lost changes.
Fixed TUI corruption from credential prompts in MCP tools like GPG pinentry.

0.1.181-alpha.2

Bug Fixes

Improved auth token refresh prevents failures during long conversations.
Added timeout to login checks prevents indefinite hangs.

0.1.181-alpha.1

Features

X Premium and Premium+ tiers now grant access to Grok Build.

Bug Fixes

Model picker now shows config alias names instead of internal slugs.

0.1.181

Bug Fixes

Model picker now displays config key names like grok-build for aliased models.

0.1.180-alpha.3

0.1.180-alpha.2

0.1.180-alpha.1

Features

New --restore-code flag restores original git commit on session resume.
Session resume warns if current git HEAD diverged from original commit.
Welcome screen shows SuperGrok upsell without active subscription.
/share slash command hidden when sharing disabled remotely.

Bug Fixes

/privacy opt-out now works for OAuth2 and CLI team tokens.
Esc unfocuses prompt; double-Esc cancels running turns.

0.1.180

Features

New --restore-code flag restores session commit when resuming with -w.
Session resume now warns if codebase HEAD diverged from original.
Welcome screen restricts actions and upsells SuperGrok without subscription.
/share slash command now hidden when sharing is disabled remotely.

Bug Fixes

/privacy opt-out now works for OAuth2 and CLI team tokens.
Esc unfocuses prompt during turns; double-Esc from scrollback cancels running turns.
Safety and policy errors no longer trigger unnecessary re-authentication.
list_dir cleans display paths for ., empty, and ./ targets.

0.1.179-alpha.5

Bug Fixes

Prevents disk accumulation from orphaned upload files and scratch directories in long sessions.

0.1.179-alpha.4

Bug Fixes

Prevents disk accumulation from orphaned upload files and scratch directories in long sessions.

0.1.179-alpha.3

Bug Fixes

Fixed conversation retries on empty model responses lacking content.
Improved bash commands in permission prompts with syntax highlighting, wrapping, and allow-always persistence.

0.1.179-alpha.2

0.1.179-alpha.1

0.1.179

Features

How-to guides bundled in binary, accessible via Ctrl+P and model-readable.
/logout slash command clears credentials and returns to login.
/privacy slash command toggles coding data retention opt-in/out.
SuperGrok upsell shown on welcome screen without subscription.
Live countdown and direct cancel added for /loop scheduled tasks.

0.1.178-alpha.1

Bug Fixes

Fixed authentication failures and web search disabling during token refresh periods.

0.1.178

Bug Fixes

Fixed authentication failures during token refresh periods preventing 401 errors.

0.1.177-alpha.1

Bug Fixes

New file creations now render as all-green diffs with 'Creating' prefix.

0.1.177

Bug Fixes

write_file tool outputs now render as proper diffs with 'Creating' prefix.
Authentication token refreshes now robust across shell, pager, and desktop.

0.1.176-alpha.1

Features

grok sessions search now finds sessions by content with scores and snippets.
Session picker now combines fuzzy matches with deep content search.

Bug Fixes

Scheduled tasks now require explicit user confirmation before cancellation.
Ctrl+V image paste now handles empty clipboard text and bracketed paste.

0.1.176

Features

grok sessions search now supports full-text queries across chat content and tools.
Session picker now combines fuzzy matching with deep content search results.
Unified logs now saved to ~/.grok/logs/unified.jsonl with Download Logs in desktop.

Bug Fixes

Scheduled tasks now require user confirmation before cancellation.
Image pasting now handles Ctrl+V and whitespace-only clipboard events.

0.1.175-alpha.1

Features

Configurable announcements and tips now load from local config files.
Live MCP server toggle enables enable/disable without session restart.
New slash command /session-info displays model, context usage, turns, and session title.

Bug Fixes

Auto-update restart now launches the correct new version.
Auth.json hot-reload fixed for grok login in other terminals.
Claude permission settings now merge rules across all files with correct precedence.

Performance

Status bar git info updates faster without synchronous spawns.

0.1.175

Features

New slash command /dream triggers manual memory consolidation.

Bug Fixes

Context bar now updates token count immediately after compaction.
Remote session restore with -w now targets worktree without source pollution.

0.1.174-alpha.2

Features

Configurable announcements and tips now load from local config files.
Live MCP server toggle enables/disables without session restart.

Bug Fixes

Auto-update restart now launches the correct new version.
Auth.json hot-reload fixed for grok login in other terminals.

Performance

Status bar git info updates instantly without synchronous spawns.

0.1.174-alpha.1

Features

Announcements and tips now load from local config files.
MCP servers can now be toggled live without restarting the session.

Bug Fixes

Auto-update restart now launches the correct new version.
Auth.json hot-reload now detects grok login from other terminals.
API error messages no longer include request body for privacy.
Status bar counting fixed, diff indentation preserved, and bash timeout increased to 10 hours.
Rate limit errors now show friendly messages with upgrade instructions.
External auth tokens now refresh correctly on 401 errors.

Performance

Status bar git info now updates without synchronous git spawns.

0.1.174

Features

New slash command /session-info displays model, context usage, turns, and session title.

0.1.173-alpha.2

Features

Configurable announcements and tips now load from local config files.
Live MCP server toggle enables/disables without session restart.

Bug Fixes

Auto-update restart now launches the correct new version.
Auth.json hot-reload fixed for grok login in other terminals.
Fixed status bar counting, diff indentation, and increased bash timeout to 10 hours.
Rate limit errors now show friendly messages with upgrade instructions.

Performance

Status bar git info updates faster without synchronous spawns.

0.1.173-alpha.1

Features

Announcements and tips now configurable from local files like config.toml.
Live MCP server toggle enables disable/re-enable without session restart.
New config options [features] managed_config and [endpoints] managed_config_url for enterprises.
Crash handler opt-in via [diagnostics] crash_handler.
/loop slash command and monitor tool for recurring tasks.
Vim-style quit commands :q, :q!, :wq, :wq! now exit TUI.

Bug Fixes

Auto-update restart now launches correct new version via symlink.
External grok login now hot-reloads auth without app restart.
Fixed hangs from broken streaming responses.
Stale background tasks now marked complete on session reload.
MCP servers loaded from ~/.claude.json and .mcp.json files.

Performance

Status bar git info renders faster using cached notifications.

0.1.173

Features

Live MCP server toggle enables disable/re-enable without session restart.
Blocked state now displayed in settings with profile images and manage button.
New config flags managed_config and managed_config_url for enterprise setups.
MCP servers now load from ~/.claude.json and.mcp.json files.
Opt-in crash handler captures SIGSEGV/SIGBUS dumps via config or env var.
New /loop slash command schedules recurring tasks with monitor tool.
Vim quit commands :q, :q!, :wq, :wq! now exit TUI.
Signal-safe crash handler wired into pager with opt-in config gate.

Bug Fixes

Auth refresh now handles re-auth, OIDC principal context, and stale tokens.
Stale background tasks now marked complete on session reload.

0.1.172-alpha.1

Features

Configurable announcements and tips now load from local config files.

Bug Fixes

Auto-updates now restart with the correct new version.
Auth.json hot-reload fixed for grok login in other terminals.

Performance

Status bar git info updates faster without synchronous spawns.

0.1.172

Features

Announcements and tips now configurable from local files like config.toml.

Bug Fixes

Auto-updates now restart with the correct new version displayed.
External grok login now hot-reloads auth without restarting.

Performance

Status bar git info renders faster using cached notifications.

0.1.171-alpha.5

Bug Fixes

Auto-update restart now launches the correct new version and displays updated version number.

0.1.171-alpha.4

Bug Fixes

Exit plan mode now shows full approve/revise/feedback options instead of simple y/n prompt.

0.1.171-alpha.3

Features

New grok ssh command enables clipboard copy from remote sessions in Apple Terminal.

0.1.171-alpha.2

0.1.171-alpha.1

Features

MCP servers can now be enabled or disabled from the UI.

0.1.170-alpha.2

Features

grok login --device-auth enables login from SSH, Docker, and other headless environments without port forwarding.
Skills tab added to the hooks/plugins modal.
Permission rules now support path globs like Edit(src/**/*.rs) and recursive ** patterns.
Slash command dropdown widened to show plugin names.
Ctrl+P command palette for quick access to commands and actions.
MCP integration tools are now discoverable and usable via search_tool and use_tool.
Terminal multiplexer clipboard support with automatic routing and diagnostics for tmux/zellij/screen.
/btw side-questions now appear as a compact inline panel above the prompt instead of a fullscreen overlay.
Command palette sections group commands by category for easier navigation.
'New Session in Worktree' command added to the palette.
Environment variables from .claude/settings.json are now loaded into sessions.
Ctrl+Shift+V / Cmd+Shift+V pastes text inline without triggering tool calls.
Marketplace now supports GitHub source type and a known_marketplaces.json registry.

Bug Fixes

Login with cached credentials no longer fails when upgrading between auth methods.
Shift+1 in VSCode terminal now correctly triggers bash mode.
Single-line command display no longer shows extra newlines.

Performance

Streaming zstd compression for file uploads reduces transfer size.

0.1.170-alpha.1

0.1.170

Features

grok login --device-auth enables login from SSH, Docker, and other headless environments without port forwarding.
Skills tab added to the hooks/plugins modal.
Permission rules now support path globs like Edit(src/**/*.rs) and recursive ** patterns.
Slash command dropdown widened to show plugin names.
Ctrl+P command palette for quick access to commands and actions.
MCP integration tools are now discoverable and usable via search_tool and use_tool.
Terminal multiplexer clipboard support with automatic routing and diagnostics for tmux/zellij/screen.
/btw side-questions now appear as a compact inline panel above the prompt instead of a fullscreen overlay.
Command palette sections group commands by category for easier navigation.
'New Session in Worktree' command added to the palette.
Environment variables from .claude/settings.json are now loaded into sessions.
Ctrl+Shift+V / Cmd+Shift+V pastes text inline without triggering tool calls.
Marketplace now supports GitHub source type and a known_marketplaces.json registry.

Bug Fixes

Login with cached credentials no longer fails when upgrading between auth methods.
Shift+1 in VSCode terminal now correctly triggers bash mode.
Single-line command display no longer shows extra newlines.

Performance

Streaming zstd compression for file uploads reduces transfer size.

0.1.169-alpha.2

Features

Added grok login --device-auth flag for headless environments like SSH and Docker.

0.1.169 0.1.168-alpha.1

Features

/compact-mode slash command toggles denser UI layout.

0.1.168

Features

/compact-mode slash command toggles denser UI layout.

0.1.167-alpha.2

0.1.167-alpha.1

Bug Fixes

Automatic cleanup deletes stale session files older than 30 days.
Linux binaries now fully static for older distro compatibility.

0.1.167

0.1.162

Features

Marketplace plugin provenance tracked in InstalledRepo with source metadata and name collision warnings.
web_fetch tool enables web content fetching with domain permissions and proxy support.
video_gen tool integrated with remote settings gating and session-based authentication.
/feedback command activates dedicated mode for submitting session feedback to x.ai.
Git stash-all via ACP extension and desktop Changes panel button including untracked files.
Pager --oauth flag forces OAuth flow for deferred welcome-screen authentication.
Fractional bash timeouts like 30.5s now supported in toolset.bash config.toml.
.mcp.json support enables team-shared MCP servers via project root config with full precedence.
Auth info extension returns login method ID and email via x.ai/auth/info ACP method.
Branch diff section shows PR-style diffs vs default branch in desktop changes panel.
Jujutsu VCS support enables detection, operations, routing, and workspaces.
Startup announcements emits x.ai/announcements/refreshed ACP notification on agent init.
Auto-enroll updates defaults to enabling automatic updates without interactive prompt.
Subagent visualizations in pager now show persona, role, and model metadata.
Bundled agents discovered from ~/.grok/bundled/agents/ with lowest precedence after project/user/built-in.
Pager welcome screen displays rotated tip-of-the-day from RemoteSettings honoring config and env overrides.
image_gen/video_gen tools instruct model to display generated media inline via markdown.
Project plugins default disabled with enable via [plugins].enabled and added counts for UI.
Enterprise managed configs via ~/.grok/managed_config.toml/requirements.toml with grok inspect and grok setup.
web_fetch tool enabled via [features] web_fetch in config.toml alongside env/remote settings.
grok login defaults to OAuth instead of legacy relay; use --legacy or GROK_OAUTH_ENABLED=0 for old behavior.
Clipboard copy over SSH/tmux now reaches local terminal via OSC 52 alongside native clipboard writes.

Bug Fixes

Cross-CWD session resume now finds local sessions under any stored directory before remote restore.
Automatic cleanup removes old grok/grok-pager binaries post-update, keeping current + previous.
Preserves legacy credentials when clearing OAuth scope in auth.json during login.
Plan file path fix includes session-specific path in EnterPlanMode output for correct writes.
Rewind truncation fix skips synthetic system-reminder messages when counting user prompts.
User prompt history excludes subagent prompts preventing Ctrl+R pollution.
Path-not-found hints from remote settings now reach TUI and shell agent config.
Legacy WebLogin users regain managed MCP config fetching post-OIDC migration.
Skill discovery handles missing frontmatter using dir name and normalizes spaced names to hyphens.
web_fetch PDFs saved to disk with read_file guidance instead of garbled text.

Performance

Parallel tool dispatch accelerates multi-tool batches via concurrent execution.
Faster session memory staleness via 7-day half-life default and equalized session source weights.

0.1.161-alpha.3

Features

Interactive plan approval dialog enables approve/reject/feedback on exit_plan_mode before building.

Bug Fixes

Prevents model crashes on large files by capping hashline_read output at MAX_LINES_READ like read_file.
Automatic cleanup removes old grok/grok-pager binaries post-update, keeping only current and previous.

0.1.161-alpha.2

Features

Parallel tool dispatch by default accelerates multi-tool execution via batched skills and timing fixes.
Web fetch tool enables controlled web content retrieval with interactive domain permissions and config gating.

Bug Fixes

Fixed pager login screen for enterprise OIDC via centralized interactive auth detection.
Per-key error streak detection prevents cross-tool interference in doom loop termination.

0.1.161-alpha.1

Features

Force OAuth login from pager welcome screen using new --oauth startup flag.

0.1.160-alpha.8

Features

Per-hook enable/disable toggles via ~/.grok/disabled-hooks file and 'e' key with multi-line j/k navigation fix.
Bash-mode execute blocks auto-expand after completion to display output immediately.
Video generation tool supports xAI API with async polling, download and sequential MP4 naming.
Marketplace plugin installs route through git_install with provenance tracking.
Parallel tool dispatch via GROK_PARALLEL_TOOL_DISPATCH=1 runs concurrent Phase 2 execution.
VideoGenConfig threading gates video_gen tool via feature flag and session auth.
/feedback command enters teal-accent mode for fire-and-forget session feedback POST.
modelId in UserMessageChunk _meta enables frontend turn-to-model association.
Git stash-all via x.ai/git/stash ACP extension wired to desktop Changes panel.
Float timeout_secs in toolset.bash config.toml enables fractional second timeouts.

Bug Fixes

Silent refresh prevents re-auth on near-expiry tokens, fixes auth_type propagation and telemetry staleness.

0.1.160-alpha.7

Features

Mid-session hook reloading on trust/untrust/add/remove without restart.
Per-turn Stop hooks plus UserPromptSubmit event fire before processing with scrollback annotations.
OAuth login option via --oauth/--legacy flags, GROK_OAUTH_ENABLED env, feature flag (legacy default).

Bug Fixes

Cross-CWD session resume finds local sessions from worktrees via --resume/--load.

Performance

Sequential N.jpg filenames replace UUIDs in image_gen for massive token savings.

0.1.160-alpha.6

Features

Live auto-refresh of open hooks/plugins modal on registry changes like install + reload.
Adds live PR status display in desktop changes panel via new ACP x.ai/pr/status extension.
Adds domain allowlist and typed WebFetchOutput enum with DomainNotAllowed and CrossHostRedirect variants.

Bug Fixes

Prevents false-positive background '&' detection inside heredoc bodies like cat << EOF... &http.Request... EOF.
Restores web search functionality by reinjecting required proxy headers like x-grok-client-version.
Ensures config.toml base_url fully overrides default model api_base_url across all credential paths.

0.1.160-alpha.5

Features

GrokNight default theme with runtime color quantization ensures correct rendering across truecolor, 256-color, and 16-color terminals.
ACP endpoints for hooks/plugins listing enable pager modals to display loaded hooks and discovered plugins.
ACP action endpoints for hooks/plugins management support trust, install, reload, and update operations via pager.
Legacy relay auth flow via grok login --legacy restores pre-OIDC accounts.x.ai token exchange.

Bug Fixes

Terminal-width-constrained tables wrap cell text proportionally without overflow or column misalignment.
Auth fallback to legacy scope reads old accounts.x.ai tokens on devboxes provisioned via x setup.

0.1.160-alpha.4

Features

WebFetch tool enables secure URL fetching, HTML-to-markdown conversion, SSRF guards, and in-memory caching.

0.1.160-alpha.3

Breaking Changes

OIDC OAuth replaces relay auth; removed --auth-signin-url/exchange-code-url/redirect-target flags, run grok login to migrate.

Features

LSP integration injects diagnostics as reminders and exposes opt-in code intelligence tools via .grok/lsp.json.
Glob-aware permission rules support prefix/suffix wildcards, tool=* and refined bash/edit matching.
Claude settings.json compatibility loads legacy rules as fallback when native TOML absent.
Path-specific Read permissions match against file paths from read_file/list_dir tools.
Web search model override via CLI, config.toml or env defaults to grok-4-1-fast-reasoning.
Grep-specific permissions enable path-only matching for Claude-compatible rules.
Remote bash timeout toggle via remote settings auto-backgrounds foreground commands with local override.
Image generation tool creates Imagine API images saved to session/images/ folder.
Configurable web search disable from config.toml via disable_web_search = true, OR-ed with CLI flag.
Image generation tool gated behind a server-side flag or GROK_IMAGE_GEN=1 env var with session auth.
Dynamic nested skill discovery in .grok/skills/ subdirs with runtime system reminder announcements.

Bug Fixes

Quit confirmation responds immediately on second ^D without infinite delay.
Terminal restore disables raw mode before CSI sequences to prevent garbage output.
Session summary falls back to prompt prefix on tool parse failure avoiding crash.
Prevents YOLO auto-approval of plan reviews by switching ExitPlanMode to independent ext_method channel.

0.1.160-alpha.2

Features

Configurable permission policies enable automatic allow/deny rules via ~/.grok/config.toml before existing checks.
Prevents timeout doom loops by auto-backgrounding foreground commands exceeding default 120s without explicit timeout.
ACP git extensions enable desktop git info/branches/checkout without local shelling, supporting cloud workspaces.
Richer hook annotations add HTTP URL/status/response previews to scrollback for pre-tool-use summaries.

Bug Fixes

Robust session cancellation shares logic with 5s safety-net timer to prevent stuck working states.
Suppresses config warning for [desktop] section owned by grok-desktop using opaque serde sink field.
Prevents hook reload panics by wrapping registry in Arc to avoid RefCell borrows across awaits.

Performance

6x faster worktree creation skips redundant LFS smudge filters on BTRFS snapshots using GIT_LFS_SKIP_SMUDGE=1.

0.1.160-alpha.1

Features

New TUI slash commands /clear alias, /context usage breakdown, /version, and /login auth status.
Skills discovery expands to.agents/skills directories alongside.grok and.claude.

Performance

Fast btrfs worktree creation on rootless devboxes via gRPC delegation to explorer agent.

0.1.160

Breaking Changes

OIDC OAuth2 replaces legacy relay login for grok login; add --legacy flag for old flow and remove --auth-signin-url etc. CLI flags (migrate by deleting auth.json).
Adds grok login --legacy fallback; removes --auth-signin-url etc flags (use env/config defaults, no migration needed for most users).

Features

ACP git extensions (x.ai/git/info, /branches, /checkout) enable cloud-local git ops without subprocesses.
Detailed hook scrollback shows HTTP POST URLs, status codes, and response previews per executed hook.
Remote bash timeout config via remote settings auto_background_on_timeout overrides local config.toml.
ACP hooks/plugins listing (x.ai/hooks/list, /plugins/list) exposes loaded configs for pager modals.
Image generation tool via xAI Imagine API saves to session/images/ folder.
Management actions for hooks/plugins via x.ai/{hooks,plugins}/action with live modal refresh.
disable_web_search now honors config.toml value ORed with CLI flag.
WebFetch tool scaffold securely fetches URLs to markdown with SSRF protection.
Dynamic discovery of nested skills from.grok/skills/ subdirs at runtime via reminders.
PR status extension (x.ai/pr/status) fetches branch PR state, title, number, and URL via gh CLI.
Web fetch domain allowlist support with enum outputs for blocked domains and cross-host redirects.

Bug Fixes

Readable markdown tables in narrow terminals via proportional column shrinking and smart cell text wrapping.
Improved permission UX by reordering 'don't ask again' option first across TUI, pager, and desktop clients.
Reliable prompt cancellation via shared logic and 5s safety-net timer prevents stuck 'working' states.
Silent [desktop] config handling consumes grok-desktop section without spurious 'unrecognized key' warnings.
Panic-free hook reloads via Arc-wrapped registry prevents BorrowMutError during concurrent plugin updates.
Clean terminal restore disables raw mode before CSI sequences to avoid garbage on Ctrl+D exit.
Robust session summaries handle chat completion errors gracefully without panics.
Eliminates compiler unused import warnings by scoping test imports to cfg-gated function.
Prevents YOLO auto-approval of exit plans by switching to independent ext_method.
Fixes web search 426 errors from missing client version header in requests to the API proxy.
Config.toml base_url fully overrides default models' api_base_url for all credential sources.

Performance

6x faster worktree creation by skipping redundant LFS smudge filters on snapshot checkouts and resets.

0.1.159-alpha.11

Breaking Changes

Default model updated enables seamless checkpoint swaps; migrate by renaming [internal_models] to [models] in config.toml.

Features

New hooks-remove and hooks-untrust commands with shared helpers across shell and TUI.
Bash mode (! cmd) enables direct shell execution from pager, bypassing agent loop.
Subagents default_model config forces all subagents to one model, overriding others.

Performance

Faster exit-plan-mode by sending plan content inline, eliminating client readFile round-trip.
Longer inference idle timeout default raised to 10min for complex responses.

0.1.159-alpha.10

Breaking Changes

Default model updated enables seamless checkpoint swaps; migrate by renaming [internal_models] to [models] in config.toml.

0.1.159-alpha.9

Breaking Changes

Config validation warns on unknown keys with typed sections; migrate auto_update etc. to [cli], [models].

Features

Resolved model ID now shown in /session-info output and feedback submissions.
Model sees task completions as <system-reminder> tags in tool results, eliminating polling loops.
Ctrl+R fuzzy history search in pager prompt supports Up/Down navigation and mouse.

Bug Fixes

Consistent tool/parameter names in model-facing text and errors via TemplateRenderer resolution.
Single progress lines during remote session restore, eliminating duplicates.
Plan mode UI updates correctly on agent entry and across session switches.
Compaction succeeds for Responses API models by preserving encrypted reasoning.

0.1.159-alpha.8

Features

Question-answer panel in pager with GrokBuildPlanNoSubagents mode excluding subagent tools.
New hook events Stop, Notification, UserPromptSubmit, SubagentStart/Stop plus Claude-compatible toolUseId schema.
Resume completed subagents via task resume_from inheriting raw transcript, tool state, model; schema/docs/provenance/observability.
/clear slash command aliases /new to start a fresh session.

0.1.159-alpha.7

Features

Install and uninstall plugins via /plugins install <url/path> and /plugins uninstall <name> supporting git refs, subdirs, local symlinks, and multi-plugin repos.

Bug Fixes

Prevents delimiter deletion in hashline_edit replace by clarifying inclusive anchor/end_anchor range in tool docs.
Prevents 413 errors by compressing large read_file images to max 1024px JPEG under 768KB with progressive quality reduction.
Eliminates TUI model jump by patching leader initialize response to use client's default_model instead of agent's.

0.1.159-alpha.6

Features

Execute inline hooks and MCP servers from plugin manifests with shell /plugins add/remove and HTTP handler support.
Redesigned OAuth consent page uses icon cards; shell callback pages now styled with success/error feedback.
MCP catalogs expose scope_name for human-readable labels alongside scope/scope_id.
Discover skills from.claude/skills directories alongside.grok/skills.
TUI hook messages render inline as scrollback annotations with hashline aliases.

Bug Fixes

use_tool now normalizes double-encoded JSON strings in tool_input to objects for reliable MCP dispatch.
Hashline anchor docs/errors now specify LINE:HASH1:HASH2 format accurately.
Codex grep_files uses shared rg_path() matching grok_build pattern.
Safe truncation prevents mid-character cuts in subagent/sampling previews.
insert_after empty content now inserts blank line matching docs.
Config watcher deduplicates events per debounce batch for stability.

0.1.159-alpha.5

Features

Dev tool usage stats pane shows real-time activity breakdown, timeline, and inter-token latency.
Claude frontmatter parity parses allowed-tools lists/strings, model, and effort in skill frontmatter.
Auto-injected managed MCPs from grok.com for WebLogin, deduped with config.toml opt-out.
ACP worktree management adds list/show/gc/db methods with filters and dry-run.
ACP x.ai/auth/logout removes scopes from ~/.grok/auth.json.
Custom npm registry via config.toml or GROK_NPM_REGISTRY respects enterprise.npmrc.
CLI worktree commands use ACP with repo-wide session resolution for -w -r.
Repo-wide worktree session resume resolves locally across same-repo directories via ACP before remote fallback.
TUI worktree mode uses repo-wide ACP resolution for interactive and headless resume.

Bug Fixes

Legacy search_replace skips nearest-match hint computation matching confusable_hint gating.
Shorter run_terminal_cmd error messages revert verbose background & operator wording.
fsnotify forwards.git events to watch_git clients including HEAD/index changes.

0.1.159-alpha.4

Features

ACP worktree management adds list/show/gc/db methods with filters and dry-run support.

Bug Fixes

Legacy search_replace skips nearest-match hint for 0.4.10 clients matching confusable_hint pattern.

0.1.159-alpha.3

Features

Subagent worktrees preserved after completion with path in output; role-level fork and isolation defaults added.
Always-approve mode renames yolo flag/slash command with backward-compatible aliases preserved.
Help skill reads ~/.grok/config.toml to answer MCP server and model configuration queries.
Deployment keys supported on all API proxy endpoints including storage and sessions.
Public install-grok.sh script supports deployment keys and channels without VPN.
Subagent prompts use dedicated compact template with system role/persona.
SubagentSessionMetadata v1 enables GCS persistence with full provenance and registry extensions.

Bug Fixes

Subagents inherit parent file toolset resolving hashline vs standard configuration correctly.
Subagent forking fixed with system prompt injection, conversation loading, and provenance tracking.
Subagent personas override model/reasoning with full precedence chain and PromptMode::Extend for built-ins.
TUI no longer stuck after auto-compact by preserving original turn ID.
Blocks nested subagent spawning by limiting depth to 1.
Subagent permission prompts show tool details; hashline_edit now requires approval.
Forked subagent context normalized to System + BackgroundContext + Task for recency.
Subagents inherit parent yolo mode and OTEL trace; depth limited to prevent recursion.
Unified skill resolution rejects ambiguities across shell, TUI, backends with qualified names and alternatives.
Prevents TUI session errors on leader connect by skipping eager new session creation.
Multi-edit diff previews for hashline_edit tool parse details array in TUI.
Fixes subagent terminal stalls using spawn_local on LocalSet single-threaded runtimes.
Prunes orphaned kill_task/get_task_output after capability filtering removes providers.

Performance

Eliminates hyper DispatchGone errors with 2 idle connections and 90s pool timeout.

0.1.159-alpha.2

Features

Namespaced plugin skills and agents enable qualified resolution with deduping and plugin provenance.
/plugins slash commands add list, reload, trust with plugin config in config.toml.
Remote default_model resolution falls back silently if unavailable in user list.
Subagent permission attribution enriches TUI dialogs and events with child provenance.
Subagent lineage tree structures enable recursive depth-sorted display in TUI tasks panel.
Structured subagent details and lineage render diagnostics below selected TUI tasks rows.
TUI /plugins commands support live reload, qualified autocomplete, and plugin source display.
Subagent safety guards fallback to parent model on unknown config or fork context overflow.
Bundled /help skill extracts README.md to ~/.grok/ for slash command and model self-help.
Identifies worktree sessions in client sidebars via session_kind and source_workspace_dir metadata in summary.json.

Bug Fixes

Per-edit diff metadata prevents cascade from line shifts in hashline multi-edits.
Prevents SQLITE_BUSY errors during concurrent SQLite WAL setup by setting busy_timeout before journal_mode.

0.1.159-alpha.1

Features

Subagent persona support enables layered instructions from config.toml and.grok/personas/*.toml with fail-closed resolution.
Compaction-safe fork inheritance preserves parent prefix while summarizing only child-owned suffix.
Enhanced task tool guidance coaches model on fork_context, capability_mode, overrides, and personas.
External auth providers support custom login binaries with TUI/headless flows.
Configurable feedback/trace endpoints route enterprise telemetry independently.
Subagent capability enforcement filters disallowed built-in tools at spawn time.
--no-memory flag disables cross-session persistence overriding all other configs.
Subagent diagnostics in TUI tasks panel display fork source, capability mode, and persona.
Worktree isolation enables subagents to edit without affecting parent workspace.
A/B retry discards failed forks and replays prompt via new ACP extension method.
Synchronous user questions via client RequestPermission block agent until answered.
User approval for exit_plan_mode via RequestPermission blocks until build or revise.
Reasoning deltas streamed to thinking panel for real-time visibility.

Bug Fixes

Fork-safety filtering removes synthetic messages and truncates incomplete turns from subagent inherited context.
Compact ACP diff metadata returns only changed lines for scattered edits spanning over 80 lines.
Per-client terminal/FS routing prevents misrouting in leader mode across mixed clients.
BYOK model api_keys preserved during proactive auth refresh by skipping session token overwrite.
Unified turn_number across remote traces and backend storage enables reliable data joins post-rewind.
Clean error messages for session restore failures in non-git directories without backtraces.
Stable slash autocomplete prevents tick refreshes overwriting command suggestions.
Explicit AuthType gates refresh to prevent overwriting user-provided api_keys.

0.1.159

Breaking Changes

Default model updated; rename [internal_models] to [models] in config.toml.

Features

Server-controlled default model via remote settings with CLI/config fallback.
Subagent capability modes enforce toolset filtering at spawn time.
--no-memory flag disables cross-session persistence overriding all configs.
Subagent metadata shows fork source, capability, and persona in TUI.
Worktree isolation runs subagents in private git worktrees.
A/B retry handler replays failed dual forks with same prompt.
Permission dialogs attribute requests to child subagents.
Lineage tree builder groups subagents by parent for diagnostics.
Subagent detail pane shows diagnostics and child lineage tree.
Synchronous user questions intercept ask_user_question tool to block agent on client answers via RequestPermission.
TUI plugin management adds /plugins list/reload/trust with live hook reload and CLI --plugin-dir.
Interactive plan approval intercepts exit_plan_mode to block until client Build/Revise.
Question-answer panel adds pager UI for interactive permissions and subagent-free GrokBuildPlan agent.
Live reasoning deltas stream raw text chunks to thinking panel.
Dev tool stats pane docks below scrollback with timeline, ITL p50/p99, and category breakdowns.
Skills reject ambiguous short names with qualified alternatives listed.
Subagents use dedicated prompt template excluding persona catalog.
Subagent role/persona now in system prompt for durable behavioral identity.
Skills parse Claude frontmatter for allowed-tools, model, effort overrides.
Plugin install/uninstall commands support git repos, tags, subdirs, local paths.
Auto-injects managed MCPs from grok.com into CLI/TUI for WebLogin users.
ACP worktree management adds list/show/gc/db methods with filters and dry-run.
ACP logout method removes scopes from ~/.grok/auth.json.
Discovers skills from.claude/skills alongside.grok/skills across local, repo, and user directories.
Inline TUI scrollback annotations for hooks with ✓ success indicator and hashline_* Claude aliases.
New hook events Stop/Notification/UserPromptSubmit/SubagentStart/Stop with Claude-compatible PreToolUse schema.
Exposes resolvedModelId in feedback submissions and /session-info slash command.
Warns on unrecognized config.toml keys at startup using serde_ignored and typed sections.
/clear slash command alias starts new session like /new.
Full resume command shown on TUI exit including cd and --resume with session ID.
Git staging awareness in hunk-tracker with batch get-all-file-contents endpoint.
Direct bash mode (! cmd) in pager bypassing agent for shell execution.
Global subagent default_model config overrides all other model sources.

Bug Fixes

Preserves BYOK api_keys during proactive session token refresh.
Unified turn_number across traces and DB enables reliable remote data joins.
Clean errors on session restore without backtraces or panics.
Slash autocomplete no longer clobbered by tick-driven arg refreshes.
Prevents token refresh overwriting user-provided API keys via explicit AuthType tracking.
A/B forks strip reasoning from assistant messages to prevent CoT leakage.
Per-edit diff metadata captures regions directly avoiding cascade in multi-edits.
TUI no longer stuck streaming after auto-compact turn_id desync.
Prevents recursive subagent spawning by limiting depth to 1.
Subagent permission prompts now display full tool details and paths.
Forked subagents get normalized prompts as [System, BackgroundContext, Task].
Leader TUI avoids session conflicts by skipping eager new session.
TUI renders multi-edit diffs correctly for hashline_edit tools.
Subagent tool stalls fixed via spawn_local for terminal actor.
Prevents subagent toolset failures by pruning orphaned kill_task/get_task_output after capability filtering.
Skips nearest-match hint computation for legacy 0.4.10 search_replace clients.
Shortens run_terminal_cmd & error messages by removing redundant background clause.
Fixes deployment key auth in feedback/sampling clients and external endpoint headers.
Normalizes double-encoded tool_input strings to objects in use_tool dispatch.
Clarified hashline anchor formats to LINE:HASH1:HASH2 in tool descriptions, docs, and errors.
Codex grep_files uses shared rg_path matching grok_build and opencode patterns.
Char-boundary-safe truncation prevents invalid UTF-8 in subagent, sampling, and prompt previews.
insert_after empty content inserts blank line matching tool docs and replace behavior.
Deduplicated config watcher events within debounce batches prevents flaky rapid-write tests.
hashline_edit docs specify inclusive replace range preserving endpoint delimiters like }.
Compresses read_file images to JPEG max 1024px/768KB preventing oversized payload 413 errors.
Frictionless desktop auto-auth by refreshing auth_manager from disk in initialize().
Eliminates duplicate summary in BackgroundTaskStarted tool output.
Suppresses duplicate reminders after kill_task or completed get_task_output.
Prevents false non-git warnings from unexpected libgit2 errors.
Blocks unsafe pipelines by tree-sitter parsing all command segments.

Performance

Fewer DispatchGone errors via larger HTTP pool (2 idle connections, 90s timeout).
Inline plan content in exit_plan_mode avoids client readFile round-trip.

0.1.158-alpha.14

Features

Full-text search across past sessions via x.ai/session/search ACP method with FTS5 indexing and workspace filtering.
ACP resume session in worktree via x.ai/git/worktree/resume_session for programmatic clients.

Bug Fixes

Scoped subagent cancellation targets only current turn with TUI confirmation modal.
Prevents corrupted post-session diffs by awaiting hooks before root repo replication.
Fast AB cancellation with async hook awaits and 30s timeouts to avoid hangs.
Compact snippets for scattered edits using per-region views with gap markers over 80 lines.
Large session sharing via signed GCS URLs bypassing 413 payload limits.

Performance

Efficient subagent progress updates with coalescing, adaptive 1-5s polling, stale indicators, and elapsed time display.

0.1.158-alpha.12

Features

Live subagent progress tracking in TUI tasks panel via ACP list_running/get polling and push notifications.
Configurable bash tool params like timeouts injected from config.toml [toolset.bash] into GrokBuild actors.

Bug Fixes

Precise io::ErrorKind mapping from ACP file errors enables consistent NotFound and PermissionDenied handling.
Silent grok-pager installation hides internal details from install and auto-update user output.
Improved batch edit errors explicitly state atomic all-or-nothing semantics and retry guidance.

0.1.158-alpha.11

Features

Configurable hashline toolset with mutual exclusion validation, scheme parameters, and dynamic descriptions.
Multi-source Claude-compatible hooks from settings files and directories with project trust gating and unified /hooks commands.
Live subagent progress shows turns, tool calls, token usage, tools used, and errors while running.

Bug Fixes

SSE stream parsing handles flat error format from Grok proxy without deserialization failure.
Grep and list_dir outputs use display paths to prevent leaking internal worktree paths to model.

0.1.158-alpha.10

Features

Plan mode state machine integrates into session lifecycle with persistence and tool support.
Post-A/B session hooks via GROK_AB_POST_SESSION_HOOK capture worktree changes.

Bug Fixes

Prevents Linux Docker overlayfs hangs in A/B sessions using git-based worktree replication.
Eliminates false tool timeouts for long-running builds after fixing overlayfs hangs.
Fixes grep large-output gRPC errors by raising tools-server limit to 128 MiB.

0.1.158-alpha.8

Features

Enhanced announcement UX adds paging, prev/next commands, and GROK_DEV_ANNOUNCEMENTS override.
--disable-web-search flag omits web search tool from agent for benchmark isolation.
hashline_read tool outputs files with line anchors using chunk h=3 c=8 scheme.
Optional auth for feedback enables unauthenticated submissions with GROK_USER_METADATA.
Plan mode tools add enter_plan_mode, exit_plan_mode, and ask_user_question with notifications.
Auto-installs grok-pager alongside grok during internal and GitHub release updates.
Mid-session token refresh re-runs external auth or uses OIDC refresh_token on expiry.
hashline_edit tool supports replace/insert_after/write with anchor validation and overlap checks.
Tiered range warnings for hashline edits caution on medium/large multi-line rewrites.
Shifted-anchor recovery in hashline edits suggests retry anchors, reports ambiguities, with wider context.
Anchor-annotated grep injects stable anchors into ripgrep output for seamless edit workflows.
Hooks v0 system runs project scripts on pre/post-tool-use and session lifecycle events.
Independent feedback gating via GROK_FEEDBACK_ENABLED separates it from telemetry controls.
Leader CLI tooling supports discovery, targeting, profiling commands, dev spawn.

Bug Fixes

Bash run_terminal_cmd respects enabled_background=false by hiding is_background in schema and rejecting at runtime.
Structured CLI errors in tool-server with code, message, and retryable fields.
Async metadata in list_dir prevents blocking the executor on overlayfs-backed paths.
Overlay path rewriting in AcpSessionFs guards AB isolation against display path leaks.
ESC cancel recovery clears cancelling state on prompt-complete notifications.

Performance

Runtime CPU profiling foundation enables leader process profiling via control protocol and pprof.

0.1.158-alpha.7

Features

Managed MCPs flag resolves env>config>remote settings>default, disabled in headless mode.
Grok 4.20 default model with CLI/env/config/remote-settings overrides for catalog models.

Bug Fixes

Prevents queued prompts flushing together in TUI via turn IDs ignoring stale responses.
Fixed viewport height enables full prompt dropdown expansion in one frame.
Respects explicit yoloMode=false per-session overriding client defaults.
Aborts path walks on first timeout preventing 25-minute hangs on slow filesystems.
5-minute tool timeouts prevent registry lock hangs with execution tracing spans.

Performance

Tuned HTTP client for IC adds keepalives, disables Nagle, sets fast connect timeout.
Streams request bodies directly to IC without buffering on common non-rewrite path.

0.1.158-alpha.6

0.1.158-alpha.5

Breaking Changes

Context window requirement forces BYOK users to explicitly set it in config.toml, with migration via Serde error messages.

Features

Dynamic model context updates ensure the chat shell uses the latest limits from the backend for accurate operations.
Custom models endpoint lets users configure their own OpenAI-compatible proxy for enterprise needs.
Configurable AB turn timeout enables dynamic limits via remote settings without redeploying.
Automatic file re-reading after compaction injects recent file contents for immediate model context.
Persisted memory reminders maintain context in chat history for improved session continuity.
Timeout for AB comparisons cancels stalled sessions after a set time, improving reliability.
Configurable memory injection enables controlling first-turn searches for better session starts.

Bug Fixes

Alpha channel updates now correctly select the latest version, fixing issues with stale releases.
Non-blocking memory flushes ensure sessions progress without delays during idle operations.
Accurate prompt completion handling prevents errors from being misreported as cancellations.
Stable application startup fixes crashes from nested runtimes during initialization.
Correct A/B session context windows ensure auto-compaction uses the right thresholds.
Reliable TUI startup eliminates panics from nested runtimes.
Async file handling prevents session hangs on slow filesystems in AgentsMdTracker.
TUI layout fixes keep prompt anchored and status visible for smoother user interaction.

Performance

Faster SSE error detection reduces CPU overhead by avoiding full JSON parses on normal chunks.
Reused HTTP connections speed up turns by eliminating TLS handshakes between requests.
Single-pass request serialization cuts overhead for large conversations in streaming paths.

0.1.158-alpha.4

Features

Remote announcements surfaced from remote settings with tolerant parsing, periodic refresh, expiry filtering, and persistent hide/show state.
grok completions subcommand generates bash/zsh completion scripts, with fast-path exit before any network or auth warmup.
Running subagents preserved after compaction with IDs, types, and descriptions injected into the post-compaction system reminder.
External auth provider delegates login to a user-supplied binary for sandboxed VMs, CI, and air-gapped environments with automatic mid-session token refresh.
Unicode confusable resilience across search_replace, read_file, and doom-loop paths — smart quotes and em-dashes no longer cause silent edit failures.

Bug Fixes

use_tool works in all server contexts by dispatching through ToolCallContext instead of session-actor interception, fixing silent failures in grok-tools-server.
MCP tool calls use fresh auth by looking up the current client at call time, fixing silent 401s after OAuth token refresh.
Remote settings forwarded through proxy — telemetry, trace upload, tool search, and writeback flags were silently dropped and never reached the client.
Leader startup fetches remote settings so telemetry, doom loop, and other remote-settings-gated features work in leader mode.
AuthManager uses configured proxy URL instead of a hardcoded default proxy host for user-info fetches.
Stale todo spinners cleared at turn end via transient Plan notification that marks lingering in_progress items as completed for display.
Config.toml model overrides applied correctly — custom api_key, env_key, and base_url on built-in model keys were silently dropped by the enum's per-variant field layout.
AB comparison cancel cleanup now uses canonical teardown path, fixing overlay mount and FS isolation registration leaks.
Connect timeout for sampling clients (10s default, configurable) prevents frozen CLI on unreachable inference servers; TLS warmup now targets the sampling path.
Leader protocol backwards compatibility by defaulting the ready field when connecting to older leader binaries.
Custom skill path discovery now works end-to-end — paths added via x.ai/skills/add or config.toml are included in agent skill lookup.

0.1.158-alpha.2

Features

Configurable telemetry destinations via [telemetry] config section and env overrides, with a trace-upload kill switch for noisy GCS warnings.
Pull-on-miss session restore fetches remote sessions when not found locally, with remote-settings-gated writeback for TUI sessions.
Session restore with dedup archives materializes GCS-referenced patches, blobs, and untracked files before applying.
Binary file attachments persisted to session storage with content-hash dedup and surfaced as path hints for model context.
Subagent support gated behind --subagents flag, env var, config, or remote settings — disabled by default.
Non-git-repo startup warning with blocking quit/continue prompt, gated behind a server-side feature flag.
Line-numbered memory_get output matching read_file format, with config-backed search defaults from [memory.search].
Richer session auto-save captures tool-usage breakdown and file paths touched; shell commands excluded to avoid persisting secrets.
grok memory reindex and doctor CLI commands for index maintenance, plus access-frequency boost in hybrid search.
Cursor-based session reconnect skips already-seen replay events, forwarding only post-cursor updates as live.
Subagent support in web UI with clickable session cards, persistent spawn/finish notifications, and structured completion output.
Session restore progress tracking with phase-level events, elapsed timers, and explicit incomplete-vs-complete outcome differentiation.
Per-subagent model routing via [subagents.models] config and agent definition model field, with conditional override for heavy parent models.
Transparent stdio bridge reconnect replays cached initialize and session/load after leader restart so external clients resume immediately.
Leader/client version mismatch notification surfaces a TUI banner and headless log warning when client and leader binaries diverge.
OIDC manual paste fallback races stdin against the loopback server so remote VM users can paste auth codes directly, with [auth] config alias.
Session ID clipboard copy on /session-info, with a transient status banner for success and error feedback.
Telemetry defaults to off with centralized env > config > remote settings > default precedence for telemetry and trace uploads.
User-defined subagents via .grok/agents/*.md files with per-subagent config toggles and dynamic Task tool descriptions.

Bug Fixes

Doom-loop detection hardened with whitespace-normalized fingerprints, per-file failure tracking, nearest-match hints, and outcome-aware error streaks.
Restore no longer overwrites git identity — synthetic commits use scoped env vars instead of writing to repo-local git config.
Session token routing fixed for default models — grok login users no longer get 400 errors on proxy-routed models.
Richer API error diagnostics with redacted headers, request body previews, and response metadata in failure messages.
Worktree list preserves full IDs by computing dynamic column width instead of truncating to 16 characters.
Restored remote sessions create distinct local children with parent tracking, preventing identity reuse and duplicate restores on repeated grok -r.
Auto A/B testing works on non-macOS by treating absent worktree pool as passthrough instead of gate.
Worktree ID collision eliminated by switching from time-based UUID v7 prefix to random UUID v4.
Cancelled and restored sessions no longer hang by sending explicit shutdown commands before dropping session handles.
Memory search config and watcher correctly applied — [memory.search] was silently ignored and first-use sessions missed watcher startup.
Auto-continue prompts excluded from real user query counts in compaction, memory hooks, and session-end telemetry.
Deleted memory files removed from search results and appended content indexed immediately without watcher restart.
Doom-loop detection overhaul with polling-aware stagnation tracking, turn-scoped resets, targeted per-context warnings, and richer error classification.
A/B overlay path sanitization across system prompt, tool results, URL-encoded paths, error messages, and get_task_output command display.
MCP server retry on init failure by restoring HTTP config after handshake errors instead of permanently dying.
Sessions no longer hang on stalled inference — content-aware idle timeout distinguishes keepalive SSE events from real completion tokens.
Turn completion signaled after empty-response retry by emitting a fallback AgentMessageChunk when streaming events were lost.
Up-arrow no-op on empty history prevents entering history search mode when no previous prompts exist.
Safer conversation compaction by stripping orphaned tool results and falling back when validation fails.
Reliable prompt persistence by gating persist_ack on chat history acceptance and flush barrier completion.
Leader startup deadlock resolved by releasing the file lock before connecting so the leader can reach readiness.

Performance

Faster session restore by downloading codebase, memory, and state archives concurrently instead of sequentially.
Bounded hunk tracker memory — binary and >1MB files tracked without retaining content, fixing 1.2GB retention from large files.
Non-blocking telemetry uploads prevent prompt turns from stalling when the GCS proxy is unhealthy.
Reduced conversation query overhead with narrow targeted queries that avoid full O(n) conversation clones.
Faster session resume via selective prompt extraction that skips full deserialization of non-prompt update lines.
Bounded long-session memory via eager in-memory pruning of old tool results after each user turn.

0.1.158-alpha.1

Breaking Changes

Repeated failed edits now trigger reread guidance and search_replace NoMatchesFound responses changed shape to include file_path; update ACP/tool-output parsers.
Bounded hunk-tracker memory usage adds explicit file-content status views to the ACP response; clients should migrate from legacy content fields to baseline and current metadata.

Features

Configurable telemetry routing with custom event endpoints, Mixpanel controls, and remote kill switches for analytics and trace uploads.
Remote session restore on local miss by hydrating backend sessions into local storage and gating writeback sync with a server-side flag.
More complete remote restores by materializing deduplicated patches and blobs before replay, with partial-restore warnings for external users.
Binary file attachments are now supported by decoding blob resources to session storage with MIME metadata, size limits, and content-hash deduplication.
Subagent task execution adds child agent sessions with lifecycle tracking, progress rendering, and background-task management across the CLI stack.
Opt-in subagent spawning adds a --subagents flag plus env, config, and remote settings controls to gate the task tool.
Safer startup outside repositories with a feature-flagged non-git warning that lets users quit before losing git-backed tracking and rewind features.
More useful memory tooling with line-numbered memory_get, config-backed memory_search defaults, clearer memory docs, and richer auto-saved summaries that omit shell commands.
Memory maintenance commands with reindex/doctor workflows and search ranking that lightly boosts frequently retrieved memories.
Versioned tool contracts with preset-selected legacy behavior for run_terminal_cmd and read_file, plus contract metadata for integrators.
Subagent transcripts in web with persisted status cards, structured completion data, and reload-safe child session replay.
Per-subagent model routing with config and agent-definition overrides, plus correct inheritance from each parent session's live model.
Transparent stdio reconnect recovery by replaying cached initialize and session state after leader restarts, then notifying clients.
Remote-VM sign-in fallback by accepting pasted OIDC redirect URLs or tokens when the browser runs on another machine.
One-click session ID copying from session info, with a transient status banner and clipboard error reporting.
Telemetry stays off by default, with env, config, and remote settings precedence for telemetry and trace uploads.

Bug Fixes

Stronger repeated-edit detection by normalizing whitespace in search_replace fingerprints so indentation-only retries now trigger doom-loop warnings.
Faster edit recovery after misses with nearest-match hints in search_replace errors and tighter guidance to use minimal unique anchors.
Session restore no longer overwrites repo git identity by scoping synthetic commit author details to the restore subprocess only.
Error-loop termination is more reliable by tracking per-file tool failures separately and only resetting warnings after successful writes.
Longer repeating tool cycles are detected by raising doom-loop cycle length coverage and using deterministic fingerprint hashing.
Logged-in ACP clients can use default fast models again by routing session tokens through the chat API proxy instead of api.x.ai.
API failures are easier to diagnose with redacted request and response context in sampling errors and improved error propagation.
Full worktree IDs in listings prevent truncated UUIDs and hashes, with dynamic column sizing in worktree list output.
Restored sessions keep local lineage by creating a new local child session and tracking the remote parent for repeatable resumes.
Auto A/B works on Linux and other non-macOS systems by falling back to on-demand worktree creation when no pool exists.
Concurrent worktree creation avoids collisions by switching temporary worktree IDs from time-based UUID prefixes to random UUID v4 values.
Cancelled and restored sessions stop getting stuck by routing prompt-complete notifications correctly and sending explicit shutdowns during teardown.
More reliable cross-session memory by sharing one backend configuration, filtering synthetic prompts, fixing delete and append reindexing, and correcting telemetry counts.
Cursor-aware reconnect replay skips already processed updates and resumes session streams from the last seen event.
Fewer false doom-loop stops with smarter polling detection, per-turn resets, targeted warnings, and synthetic warning tags.
A/B comparison path isolation so prompts, tool results, and task output no longer expose overlay worktrees or wrapper commands.
Recoverable MCP startup failures by restoring HTTP configs for retries and surfacing readable MCP App handshake errors.
Clearer restore progress and failure states with phased events, elapsed times, and explicit incomplete-session-state reporting.
Hung streaming requests now fail fast by timing out keepalive-only SSE streams that stop producing real model content.
Leader startup no longer races client connects by binding the IPC socket early and gating ACP traffic on explicit readiness.
More reliable A/B filesystem isolation with a configurable non-overlay scratch base before falling back to tmpfs mounts.
Clients can distinguish auto-update restarts by propagating explicit leader shutdown reasons through the IPC protocol and reconnect state.
Completed turns no longer disappear downstream by emitting a fallback text chunk when streaming events were lost after retries.
Leader/client version mismatches are surfaced immediately with ACP notifications, TUI banners, and headless warnings after registration.
Prompt history search no longer opens empty by ignoring Up-arrow history mode when both queued prompts and saved history are absent.
Legacy list_dir compatibility for older clients, restoring empty-directory, error, and depth-threshold output parity.
Safer conversation compaction by stripping orphaned tool results and falling back when replayable history validation fails.
More reliable prompt persistence by acknowledging prompts only after chat history accepts them and the flush barrier completes.
Legacy task error parity for older clients, restoring exact not-found wording in get_task_output and kill_task.

Performance

Faster session restore downloads codebase, memory, and state archives concurrently while surfacing clearer restore strategy and warning summaries.
Prompt turns stay responsive by moving GCS telemetry uploads to fire-and-forget paths when retries or queue fallbacks occur.
Lower memory use in long sessions by replacing full conversation clones, selectively scanning updates, and pruning retained tool results.
Faster startup before code navigation by lazily building indexes only for web clients that advertise x.ai/codeNavigation.

0.1.158

Features

Remote announcements from remote settings with tolerant deserializer.
Shell completions via grok completions <shell>, leader CLI, and stdio reconnect replay.
User-defined subagents via.grok/agents/*.md with config toggles.
Post-compaction reminder includes running subagents with IDs, types, and poll/cancel instructions.
External auth provider binary enables login via custom commands in sandboxed/air-gapped environments.
Custom models_base_url auto-fetches OpenAI-compatible model list for enterprise proxies.
Guarded normalized fallback matching enables search_replace on confusable Unicode typography (gated by flag).
Full-text search across sessions via new x.ai/session/search ACP extension with FTS5 indexing.
Retained file context post-compaction by re-reading up to 5 recent files into history.
Persisted memory reminders upsert into conversation system prompt without duplicating transient injections.
Configurable AB turn timeout cancels stalled comparisons after wall-clock limit with cleanup and observability.
Configurable first-turn memory injection with dedicated thresholds and remote settings support.
Managed MCPs feature flag skips fetching by default in headless mode.
Paged announcement UX adds prev/next commands with reliable startup visibility.
Disables web search tool via --disable-web-search flag or GROK_DISABLE_WEB_SEARCH=1 for benchmarks.
hashline_read tool annotates file output with anchors like LINE:LOCAL:CONTEXT→CONTENT.
Grok 4.20 default model with CLI/env/config/remote-settings overrides for web search and summaries.
Unauthenticated feedback submission via optional tokens and GROK_USER_METADATA env.
Plan mode tools EnterPlanMode, ExitPlanMode, AskUserQuestion send structured notifications.
Auto-installs grok-pager alongside grok during internal and GitHub release updates.
Mid-session token refresh via OIDC grants or external auth binaries prevents expiry.
Hashline toolset enables anchor-stable file read/edit/grep with validation, recovery, ranges, and config integration.
Hooks system executes custom scripts for pre/post-tool and session events from ~/.grok/hooks/ with deny-wins trust controls.
Independent feedback flag gates /feedback and popups separately from telemetry via GROK_FEEDBACK_ENABLED.
Leader CLI adds grok leader list/info/profile/kill/dev for discovery and CPU profiling.
Plan mode state machine enables agent planning phase with enter/exit tools and session persistence.
Hooks from Claude settings loads from ~/.claude/settings.json and project .claude/settings.json.
Live subagent progress shows turns, tools, tokens, errors in get_task_output for running tasks.
Unified /hooks command supports list, trust, add <path> with fuzzy autocomplete.
ACP x.ai/subagent/list_running queries live progress for all running subagents of parent session.
Bash tool params from config.toml override schema defaults like timeout_secs for GrokBuild.
Live subagent progress in TUI tasks panel via ACP polling and push notifications.
ACP extension resumes sessions in worktrees via single call matching grok -w -r CLI flow.
Signed GCS uploads for shares bypass proxy limits to prevent 413 errors on large sessions.

Bug Fixes

Prompt persistence ack gated on chat history flush barrier completion.
Leader spawn deadlock resolved by releasing lock before connect.
Headless clipboard failures handled silently in session info.
Server-side flags forwarded via proxy for telemetry and other features.
Leader fetches settings enabling remote-settings-gated runtime features.
AuthManager respects configured proxy URL instead of hardcoding the default proxy host.
Clears stale todo spinners at turn end with transient Plan notification without mutating underlying state.
Config.toml model overrides now apply custom api_key/base_url to built-in models.
Prevents queued prompts flushing together via per-turn ID guarding race between notification and response.
Alpha updates pick max(stable, alpha) and harden channel admissibility across all installers.
Unicode confusables now detected and normalized in search_replace/read_file with diagnostics, reminders, and guarded fallback matching.
Fixes overlay mount and FS isolation leaks on AB comparison cancel by using canonical cleanup.
Adds connect timeout to sampling clients and warms TLS roots for first-chat cold-start latency.
Leader protocol compatibility by defaulting ServerMessage::Registered.ready to true.
Custom skills now discoverable via config.toml paths and x.ai/skills/add in system prompts.
Non-blocking memory flushes by spawning idle-timer tasks asynchronously.
Accurate turn end display by not defaulting missing prompt-complete stop reasons to cancelled.
Correct error display for retries by checking explicit Cancelled stop reason only.
Startup panic eliminated by replacing nested tokio runtime with direct await.
Correct auto-compact in AB forks by overriding context_window from target model registry.
Async FS operations in AgentsMdTracker prevent hangs on overlayfs-backed filesystems.
Fixed TUI viewport layout anchors prompt at bottom with reliable status bar.
run_terminal_cmd exported schema hides and rejects is_background when disabled via params.
Fixed viewport height enables single-frame dropdown resize without animation.
Respects explicit per-session yoloMode=false overriding client defaults in leader and TUI.
Aborts path walks on first filesystem timeout preventing 25-minute hangs on slow mounts.
5-minute tool execution timeout prevents registry hangs with diagnostic tracing spans.
Non-blocking list_dir uses tokio::fs::metadata to avoid executor hangs on overlayfs-backed paths.
A/B fork writes are guarded by display-to-overlay path rewriting in AcpSessionFs.
ESC cancel resets [cancelling] state via relaxed prompt-complete guard.
Linux A/B forks skip overlayfs and replicate via git diff to prevent container syscall hangs.
No tool timeouts allows long-running builds without false positives.
Matches tools server gRPC limits to 128 MiB preventing OutOfRange on large grep outputs.
SSE parsing handles flat errors from Grok proxy alongside OpenAI-standard nested format.
ACP file errors map to io::ErrorKind ensuring consistent NotFound/permission dispatch.
Display paths in grep/list_dir prevent internal worktree leaks to model in A/B forks.
Silent grok-pager installation hides implementation details from install/update output.
Improved hashline_edit batch errors state atomicity and retry-all guidance.
Scoped subagent cancellation targets only current-turn subagents.
Post-session hooks await before replication avoids root-repo diff corruption.
AB cancellation unblocked via async hooks and cleanup timeouts.
Per-region snippets for scattered edits cap output at ~40 lines instead of 10K+ for distant changes.

Performance

SSE error detection skips JSON parse on normal chunks via fast contains(""error"") guard.
Faster LLM turns by persisting HTTP/2 client with connection pooling across conversation turns.
Faster streaming completions by single-pass serialization replacing serde Value mutation.
Tuned reqwest client enables keepalives, nodelay, and timeouts for faster IC backend requests.
Streamed request bodies skip buffering and parsing for large non-rewritten payloads.

0.1.157-alpha.1

Features

Subagent support gated behind --subagents flag, GROK_SUBAGENTS env var, config.toml, or remote settings — disabled by default.

Bug Fixes

Cancelled and restored sessions no longer hang by sending explicit shutdown commands before dropping session handles.

Performance

Faster session restore by downloading codebase, memory, and state archives concurrently instead of sequentially.

0.1.157

Features

Remote session restore on miss fetches session data from backend when not found locally, with remote-settings-gated writeback.
Dedup-referenced codebase content materialised during restore by downloading GCS-backed patches and blobs before replay.
Binary file attachments decoded from base64, written to session storage with content-hash dedup, and surfaced as path hints to the model.
Subagent spawning and lifecycle management with coordinator tracking, TUI progress rendering, and task tool integration.
Non-git directory warning with blocking confirmation modal at startup, gated behind a server-side feature flag.
Line-numbered memory_get output matching read_file format, with memory_search defaults now respecting [memory.search] config.
Richer session-end auto-saves now include tool-usage breakdown and file paths touched; shell commands excluded to prevent credential leakage.
grok memory reindex and doctor CLI commands for index maintenance, plus access-frequency boosting for retrieved memory chunks.

Bug Fixes

Doom-loop detection hardened with whitespace-normalized fingerprints, per-file failure tracking, nearest-match hints, and error-streak termination.
Session restore no longer overwrites git identity — synthetic commits use subprocess env vars instead of repo-local config.
Session-authenticated users no longer get 400 errors on default models — credential routing checks actual token source, not advertisement method.
API error messages include structured context — request URL, relevant headers, and body preview surfaced on auth, payload, and server failures.
Worktree list displays full IDs with dynamic column width instead of truncating to 16 characters.
Memory search config honored — [memory.search] settings were silently ignored across all three backend construction paths.
Correct memory injection after compaction — first-turn context now uses the real user query, not the auto-continue prompt.
Deleted memory files no longer searchable — watcher sync removes stale chunks, and /memory append content is indexed immediately.

Performance

Reduced session memory pressure by replacing 8 full-conversation clone sites with narrow single-field actor queries.
Faster session resume via selective updates.jsonl scanning that skips full deserialization of irrelevant update types.
Bounded long-session memory by eagerly pruning old tool results from the retained conversation after each user turn.

0.1.156

Features

Per-session tip rotation via a persistent cursor so each launch shows the next tip in sequence instead of the same UTC-day tip.
Cross-devbox session restore via grok sessions list/search CLI subcommands and hardened grok -r with upload-ordering safety, cwd-scoped lookups, and staged-vs-unstaged correctness.

Bug Fixes

Scoped MCP connector selection so multiple connectors with the same URL but different auth scopes (Personal/Team/Org) route to the correct token.
GCS upload reliability with panic-catching upload tasks, full error-chain logging, and aggressive HTTP connection pool eviction to prevent stale-connection retries.
Doom loop detection for repetitive edits that were previously invisible because interleaved bash commands reset the detector.
Binary blob size cap enforced during archive building, preventing multi-GB binaries from attempting uploads that would timeout.
Graceful upload drain on shutdown so pending GCS uploads complete before exit instead of being silently abandoned.
Malformed tool-call JSON sanitization replaces invalid arguments with {} before sending to providers, preventing permanent 400 loops from broken model output.
MCP tool name validation skips tools with invalid characters (e.g. spaces) that cause Anthropic/OpenRouter 400 errors, instead of crashing the session.
Patch restore compatibility with older git by removing --allow-empty flag and pre-checking for diff content before calling git apply.

Performance

Streaming multipart upload for large files (>50 MB) bypasses proxy body limits and avoids loading entire files into memory.

0.1.155-alpha.6

Features

Tip of the day shown at startup, served dynamically from remote settings with local opt-out via [cli] show_tips = false.
OAuth2 Authorization Code + PKCE login flow replacing legacy relay auth, with server-controlled rollout and permanent legacy fallback.

Bug Fixes

Gitignore enforcement on read_file and search_replace to reduce accidental secret exposure — aligns with list_dir and grep behavior.
Session resume race fix — prevents --resume from creating a shadow session when LoadSession is in-flight via leader IPC.
Atomic summary writes via temp-file + rename — fixes EOF parse errors when A/B comparison startup races with summary updates.
Memory tool labels — memory_search and memory_get now render as "Memory Search" and "Memory Read" instead of generic labels.
HTTP/2 connection poisoning fix — rebuilds the sampling client on final retry to recover from stale connections after server GOAWAY/RST_STREAM.

0.1.155-alpha.5

Features

Tip of the day shown at startup from remote settings, rotated daily; opt out with [cli] show_tips = false.
Gitignore-aware file tools — read_file and search_replace now refuse access to ignored paths, matching grep and list_dir.

Bug Fixes

Session resume no longer races or silently creates a new session when leader IPC or replay fails.
Atomic summary.json writes prevent EOF parse errors during concurrent A/B comparison session loads.
Memory Search and Memory Read labels in TUI instead of generic Search/Read for memory tool calls.
HTTP/2 connection poisoning recovery via idle timeout eviction and fresh client rebuild on final retry attempt.
Bash command output appears immediately by flushing the replay buffer before turn end in ! prefix mode.

0.1.155-alpha.4

Features

End-to-end distributed tracing across clients and the API proxy via W3C _meta.traceparent propagation into a single distributed trace.

Bug Fixes

Structured error variants for read_file, list_dir, search_replace, and todo_write so callers can distinguish file-not-found, permission-denied, and duplicate-id failures.
Replay notifications now guaranteed before session/load response via forward_with_completion() drain; removes public acp_send_fire_and_forget from xai-acp-lib.
Writeback sync now includes cwd and title in backend metadata on every flush and rename, fixing null values in remote session listings.
Typed TaskNotFound errors for kill_task and get_task_output with known-task-ID hints, enabling Python-side structured error classification.
Terminal commands detached from controlling TTY via setsid() so subprocesses like ssh cannot steal input from the TUI.

0.1.155-alpha.3

Features

MCP Apps support for rendering interactive UI tools served via ui:// resources from MCP servers.
OpenID Connect login with PKCE, local callback server, and configuration via config.toml or environment variables.
On-demand MCP tool discovery via BM25 search and use_tool meta-dispatch for KV cache-stable definitions.
Kill background tasks by task_id via new x.ai/task/kill ACP extension method.
Integration tool names visible in TUI instead of generic search_tool/use_tool plumbing labels.

Bug Fixes

Rewind correctness overhaul — mutation-free preview, ghost message filtering, compaction-aware replay for all targets.
Memory flush and compaction no longer fail on orphaned tool_result messages by stripping tool blocks before summarization.
Slash command autocomplete enters argument phase instead of executing immediately, and option-key word traversal corrected.
Cancel cleanup awaits child process exit to reclaim memory before returning, preventing cascading OOM kills.

Performance

Parallel tool execution via shared resources and explicit ToolCallId, removing the sequential single-tool bottleneck.

0.1.155-alpha.2

Features

Config hot-reload watches auth, MCP servers, memory, skills, UI, and models for live changes without restart.
/memory slash command appends notes to workspace or global MEMORY.md with smart Markdown heading normalization.
cgroup v2 memory limits for spawned commands gracefully OOM-kill only the offending process, keeping the session alive.
Debugging technique retention across sessions by capturing API endpoints, CLI commands, and investigation workflows in memory flush.
File overwrite guard on search_replace prevents empty old_string from silently replacing existing file contents.
Full command preview in scrollback for long or multi-line bash commands that exceed the status bar.

Bug Fixes

Duplicate tool name validation prevents unreachable tools when two share the same client-facing name.
Large pasted text written to disk so the model's read_file fallback finds the content instead of file-not-found.
TUI reconnect no longer hangs permanently after leader restart — 30s timeout and guaranteed completion signal.
Duplicate leader exits cleanly with 30s lock timeout and early bail-out when an existing socket is detected.
A/B fork session traces now include repo_root and remote_url via shared git2 discovery logic.
Distinct fs tool error messages differentiate 'does not exist' from 'is a file, not a directory' to prevent agent loops.
Drain timeout for inherited pipes prevents cmd & without stdout redirect from blocking the actor loop indefinitely.
Leader log level defaults to info instead of debug, cutting log volume ~10x over long-running sessions.
Memory flush sanitization strips tool_result blocks before windowing to prevent orphaned tool_use_id 400 errors.
A/B comparison disk leak fix — correct overlay unmount ordering prevents 4.9 GB leak, and primaryModelId now reflects the actual variant model.
Revert memory flush sanitization that stripped tool messages and broke flush request formatting.

Performance

Faster first prompt on large repos by capping startup git_status at 2s instead of blocking 10-20s on index refresh.
Dedup blob existence pre-check skips redundant uploads when another agent has already written the same SHA256 content.
Session eviction on client disconnect drops SessionHandles to reclaim ~100-500MB per session when IPC clients leave.
mimalloc allocator on macOS Apple Silicon for alpha builds, fixing 15GB+ RSS from system allocator's unreturned pages.

0.1.155

Bug Fixes

HTTP/1.1 fallback for sampling requests when HTTP/2 connections fail with transport errors from GOAWAY or RST_STREAM frames.
Handle response.failed and error streaming events from the Responses API instead of silently dropping them.
Prevent file descriptor exhaustion on macOS by capping parallel walker threads and raising the soft FD limit at startup.

0.1.154-alpha.6

Features

Overlay-based A/B replication replaces fsnotify+hunk-tracker with overlayfs to capture all terminal-created files during A/B sessions, fixing lost-file bugs on Linux.
/flush slash command in the TUI triggers an on-demand memory flush to disk via the existing backend extension method.

0.1.154-alpha.5

Features

Tool execution tracing via call_id passthrough, follow-up messages, and structured error detail fields in CLI output.
Mandatory AGENTS.md discovery — runtime-discovered project instruction files must now be read before proceeding.
PTY notification routing to originating client via targetClientId, with cwd fallback and terminal UI polish.
--no-auto-update flag for leader mode to prevent update-triggered shutdowns in long-running sessions.
On-demand memory flush via /flush command and x.ai/memory/flush extension with concurrent-flush prevention.
Tool state export/import via GetToolState RPC and initial_tool_state_json, enabling session warm-start and cloning.
GROK_MEMORY=0 force-disables memory regardless of config.toml or remote settings; CLI flag still overrides.
Per-tool MCP timeout overrides via tool_timeouts in config.toml and toolTimeoutsMs in ACP session meta.

Bug Fixes

Slash command completion now respects optional arguments, fixing /load, /model, /compact, and /theme execution.
Per-session model, sampling, and YOLO tracking in leader mode, preventing cross-client state contamination.
Complete session replay on reconnect by switching to unbounded client channels, preventing silently dropped notifications.
Zombie process prevention on leader respawn with child reaper, config.toml auto_update respect, and stale-binary resolution.
OOM prevention in codebase indexer via 5MB file-size cap, binary detection, and hidden-directory filtering.
Per-client YOLO mode isolation in leader mode — toggling no longer contaminates other clients' sessions.
A/B merge-back recovers files in new directories created via terminal, fixing the inotify recursive-watch race.
Worktree cwd preserves subdirectory offset instead of resetting to the repo root.
8 MiB stack for session threads prevents stack overflow on macOS where the default is 512 KB.
Grep tool description clarifies raw regex syntax to prevent the model from quoting patterns.
Tool kind params merge across all tools of the same ToolKind, fixing template render failures for multi-tool configs.
User query placed before attached file contents in the prompt for better model attention.
Read file errors now include the file path and underlying cause across all tool implementations.

Performance

Lightweight index status queries via GetFileCount/GetStats, eliminating a full ScopeGraphIndex clone on every status check.
Hunk tracker skips unnecessary file reads in AgentOnly mode and eliminates a redundant content clone.

0.1.154-alpha.4

Features

Background command lifecycle reworked with graceful SIGTERM→SIGKILL escalation, non-blocking I/O, and proper agent tracking for user-backgrounded commands.

Bug Fixes

Auto A/B comparisons deferred until worktree pool is ready, preventing 30-60s stall on large repos.
Headless mode (grok -p) exit panic fixed by flushing telemetry before runtime teardown.

0.1.154-alpha.3

Features

MCP server management via mcp/list, mcp/call, and mcp/servers_updated extension methods with multi-scope connector resolution.

Bug Fixes

PTY notifications reach the correct client via _meta routing metadata, with shared helpers and session-aware cwd fallback.
Concurrent tool calls no longer crash the tools server; semaphore serializes access to the thread-local toolset.
Cross-session notification leaks eliminated in leader mode for relay, dead-client, and ext/notification routing paths.
Leader auto-update pre-downloads the binary before shutdown and resolves the ~/.grok/bin symlink at spawn time.
First-compaction memory flush fires correctly by pre-incrementing the compaction counter before the flush guard.
Interval memory flush resumes after compaction by resetting the conversation length counter when history is compacted.
Session content restored on TUI reconnect by clearing stale scrollback, resetting turn state, and gating live updates during replay.
Memory flush entries accumulate across cycles by appending to daily log files instead of overwriting each flush.
A/B experiment tool output shows original project paths instead of internal worktree directories in streaming updates.
File write truncation eliminated by switching to tokio::fs and disabling ACP filesystem routing from the TUI.
Token refresh works for JWTs with aud claims by disabling audience validation during expiration parsing.
Session resume and fork no longer crash with "Is a directory" error caused by a regression in the Tool trait migration.

0.1.154-alpha.2

Features

Structured headless prompts via --prompt-json and --prompt-file flags, with --verbatim to skip query wrapping.
Bring-your-own-key for A/B comparisons by reading [ab] openrouter_api_key from config.toml.

0.1.154-alpha.1

Features

Mid-session MCP server toggling via x.ai/session/update_mcp_servers extension method with optimistic rollback.
Sandbox profile configurable via GROK_SANDBOX env variable, defaulting to workspace profile on devboxes.
Agent version exposed in InitializeResponse metadata for relay and client version discovery.
grok worktree subcommand for listing, inspecting, removing, and garbage-collecting session worktrees.
Character-budget BFS directory listing replaces depth-based summarization, expanding small deep directories when budget allows.

Bug Fixes

Tool-server finalize subcommand aligned with the Python client via named --tools/--truncation args and --json output.
Completion requirement tracking restored in recovery loop, fixing duplicate tool calls introduced by the DynTool migration.
Foreground command cancellation no longer deadlocks — terminal backend access moved outside the registry mutex.
A/B session filesystem isolation fixed by replacing overlayfs with bind-mounts, eliminating split brain between bash and file tools.
ACP file operations were silently bypassing the client filesystem backend, breaking ask-mode write rejection in GrokCode.
Zero-argument MCP tool calls no longer fail with JSON parse error when the model emits an empty arguments string.
Model selection in leader mode now persists across sessions — changing models mid-session no longer reverts on /clear.
Git index.lock contention on devboxes resolved by adding --no-optional-locks to all background git operations.
System reminders now fully respect the disabled flag, and template parameter defaults auto-populate from input schemas.
Memory tools now render in TUI and web UIs with search/read cards, query text, and file path display.

Performance

Session thread isolation via per-session tokio runtimes, preventing tool execution in one session from blocking another.

0.1.154

Features

/flush slash command in the TUI for on-demand memory flush to disk.

Bug Fixes

A/B session file preservation via host-side overlayfs, fixing terminal-created files silently lost during merge-back.
Memory flush no longer infers preferences from unchallenged assistant actions, requiring explicit user statements.

0.1.152-alpha.1

Features

Slash command autocomplete with ACP-advertised built-in and skill commands, keyboard navigation, and fuzzy filtering.
Compact and rewind controls exposed to the web frontend via typed SessionClient methods with backwards-compatible camelCase aliases.
OpenCode agent mode with 8 dedicated tools (bash, read, edit, write, grep, glob, todowrite, skill) using opencode conventions.

Bug Fixes

Codex agent tool accuracy restored by switching to dedicated codex-specific read, list-dir, and grep implementations.
File editing on macOS fixed by canonicalizing search_replace paths to match read_file's case-normalized tracker keys.
Workspace disambiguation in /load prevents session list overwrites when multiple workspaces share the same directory basename.
ACP agent profile support restored for JSON object payloads, fixing permission mode and prompt loss for web and remote clients.

0.1.151

Features

Interactive web terminals let clients run full PTY shells over ACP WebSocket with resize, reconnect replay, and unified terminal lifecycle APIs.
Configurable Codex system prompts let builds toggle custom base templates per agent definition while preserving correct agent-type resolution from the active default model.

Bug Fixes

Graceful tool cancellation handling prevents agent-loop aborts when gRPC calls are cancelled, with structured CANCELLED errors and end-to-end cancellation coverage.
Accurate edit preview line numbers now appear before approval by sending diff metadata through standard tool_call_update events instead of permission payload plumbing.

0.1.149

Features

Offline relay session sync keeps shared sessions resilient with reconnect and persisted cursors, adding connection-state notifications for reliable TUI status updates.
Foreground tool progress visibility shows active tool execution in the TUI status bar, generalizing streaming state beyond bash commands.
Config-driven skill loading lets users add and ignore custom skill paths from config.toml, extending discovery beyond default local and user directories.
Persistent leader availability after disconnects by spawning leader with --no-exit-on-disconnect, allowing new clients to reconnect without restarting the subprocess.

Bug Fixes

Reliable CLI updates now refresh grok-latest during installs, ensuring symlink-based setups launch the newly installed binary in fresh shells.
Parallel tool-call stability prevents Anthropic ordering errors by deferring follow-up user messages until all tool_result blocks in a batch are emitted.
Channel-aware update checks correctly install alpha builds during channel switches by combining semver comparison with explicit stable-versus-alpha mismatch logic.
More reliable tool execution in forks validates concatenated JSON against the named tool schema and preserves FileReadTracker state by copying tool_state.json.
Client disconnect cancellation now stops running tools and triggers cleanup for foreground commands in unary and streaming execution paths.
Stable A/B fork path behavior keeps model-visible project paths consistent while resolving stale absolute paths safely to worktrees and preserving edit-read tracking state.
Reliable session resume across providers by repairing dangling tool calls throughout conversation history, preventing persistent 400 errors on interrupted legacy sessions.
Accurate replayed terminal state by skipping streaming bash update handling during replay, preventing ghost in-progress timers from out-of-order persisted chunks.
Cleaner TUI tool-call logs by sending no placeholder raw_input until tagged ToolInput arrives, eliminating repeated 'missing field variant' parse errors.

0.1.146

Features

A/B filesystem isolation prevents cross-session contamination by running forked sessions in overlayfs and syncing winner side-effects safely.
A/B safety guardrails now block unsafe side-effecting commands during comparisons and abort contaminated experiments before execution.

Bug Fixes

Responsive A/B cancellation now interrupts worktree preparation and mid-copy operations, avoiding stuck comparisons and stale vote blocking.
Background task visibility now shows immediate execution feedback and caps blocking waits, preventing apparent freezes during long-running commands.

0.1.145

Bug Fixes

Reliable hunk acceptance and replication by refreshing baselines on git state changes and preserving tracked paths after accept/reject workflows.
Cleaner client tool event stream by suppressing updates for skipped tool calls while still persisting cancellation context in conversation history.

0.1.144

Features

Offline-safe feedback retention writes feedback and A/B votes to local session JSONL before network submission, including previously dropped vote rationale fields.
Agent environment metadata becomes queryable by merging GROK_AGENT_METADATA into initialize responses and persisting the full blob in backend agent records.
Dynamic leader-mode control resolves from CLI, local config, then remote settings, and startup also adopts remote upload limits.

Bug Fixes

Complete prompt trace capture preserves full pre-truncation text in GCS and records truncation metadata for reliable debugging of long prompts.
Backward-compatible metadata semantics distinguish absent legacy prompt flags from explicit false values by serializing truncation/image indicators as optional booleans.
Rewind preview no longer crashes on multibyte text by truncating at UTF-8 character boundaries instead of raw byte offsets.
Large file attachments are safely bounded by emitting metadata stubs above a token threshold, preventing context blowups while preserving file discoverability.
Non-interactive edit workflows no longer fail read-before-edit checks because tools server defaults skip_read_before_edit to true at startup.
A/B forks now keep subdirectory context so comparison sessions run in the same repo subpath and replicate changes correctly.
Server-side A/B gating is now enforced by honoring classification results before forking, reducing unnecessary comparisons for low-value prompts.
Forked A/B sessions now inherit current YOLO mode so tool permissions stay consistent with user state and comparisons avoid approval stalls.

Performance

Faster A/B fork creation removes backend registration latency from fork setup by spawning session sync as background telemetry work.
Lower A/B fork overhead skips low-value telemetry calls for ephemeral comparison forks, avoiding extra session-load work during experiments.
Quicker dual-worktree setup reuses one precomputed dirty-state scan across both A/B syncs, eliminating redundant git status calls on large repos.
Shorter A/B comparison startup overlaps both fork-and-spawn flows with concurrent joins, reducing sequential network wait in dual-fork initialization.
Pool restart latency drops by adopting valid orphan worktrees instead of recreating them, with tested cleanup and adoption flow hardening.

0.1.143

Features

Prompt introspection from the CLI is available via grok prompt, with JSON or section output and persisted session prompt_context snapshots.
Prompt image tracing now captures user-supplied images as decoded per-turn files in GCS, improving multimodal debugging and auditability.
A/B comparisons now support ask/plan read-only mode by forking sessions without git worktrees, enabling comparisons outside repositories with safer non-mutating defaults.
Per-server MCP timeouts can be set via _meta.mcpConfig, with relay passthrough and precedence over config defaults during server startup and tool calls.
Ask/plan A/B can force worktrees through [cli].ab_force_worktrees, overriding read-only defaults when operators need full filesystem-isolated comparisons.
Installation diagnostics help troubleshoot conflicting grok binaries by listing canonical and invoked paths, versions, update status, and optional JSON output.

Bug Fixes

Reliable todo status updates now accept merge-only status patches by defaulting missing content to todo IDs instead of rejecting updates.
Accurate prompt re-rendering now respects active tool overrides and disabled tools by centralizing all prompt assembly through a shared PromptContext.
A/B cancellation no longer double-finishes comparisons by suppressing completion notifications and after-uploads when cancel handlers already removed active comparison state.
Running outside git repos no longer panics by skipping gitignore construction without a repo root and guarding absolute-path ignore checks.
npm installs now use a canonical Grok binary path via postinstall copy to ~/.grok/bin/grok, preventing installer conflicts and version confusion.

0.1.142

Features

Hidden directories now appear in file listings by defaulting fs list requests to include hidden paths and removing unused session fs/git handlers.
Faster A/B comparisons in large macOS repos via reusable prefilled worktree pools, atomic claiming, and preparation signaling.
CLI self-update downloads now use authenticated, time-limited URLs with environment-scoped auth for updates.
Prompt context now includes editor focus state by rendering focused files, open files, and regular resource links in structured system reminders.
Richer editor context in prompts now includes focused and open files as resource links with cursor metadata for better code-grounded responses.
More reliable self-updates in restricted environments by falling back to GitHub Releases via gh when npm or internal installers fail.

Bug Fixes

Web search works again by passing sampling credentials through toolset overrides so the web_search tool is registered correctly.
Unique agent identities across containers by incorporating Linux HOSTNAME into ID derivation and normalizing stored agent IDs to UUIDv5.
Embedded file attachments parse more URI formats by accepting optional file:// prefixes and both #Lstart-end and #Lstart-Lend fragments.
Single-@ file references now work in TUI prompts by broadening reference parsing to support optional @ and L-prefixed line ranges.
Compacted conversations preserve query structure by wrapping summaries in <user_query> tags so resumed sessions keep consistent context formatting.
Duplicate session messages are prevented by making saveSessionData append only unseen message tails based on persisted message counts.
A/B winner changes now land in your main repo by replicating worktree edits to the original source directory explicitly.
Debug extension methods now accept camelCase params while keeping snake_case aliases, improving client compatibility without breaking existing callers.
Fewer stale HTTP client failures by disabling idle connection pooling for sampling requests and relying on HTTP/2 keep-alive checks.

Performance

Faster worktree pool operations reduce acquire and cleanup overhead with skip-clean paths and streamlined removal logic.

0.1.141

Breaking Changes

Per-model concise behavior replaces global [toolset].use_concise; migrate by setting model..use_concise=true because the old global key is ignored.

Features

Faster completion acceptance lets you confirm prompt, file, and history suggestions with Right Arrow alongside Enter and Tab.
Inline code review comments can now be created and deleted via new extension methods, with append-only GCS event records for pipelines.
Machine-readable update checks add --check and --json support to update and version commands for automation-friendly version status reporting.
Automatic project-instruction discovery surfaces newly encountered AGENTS.md and Claude.md paths during tool access, including post-compaction reminders outside the initial directory chain.

Bug Fixes

Cleaner pager command output relies on shell no-color mode, removing local ANSI stripping and preserving streamed text exactly.
Accurate restored session metrics persist full SessionSignals snapshots and reload them on resume, avoiding turn-count drift after compaction.
More reliable long streaming requests use per-sampling HTTP clients with keep-alive tuning instead of a shared global client.
Startup stability for external instruction files avoids gitignore panics by skipping ignore checks for paths outside the repository root.
Tool-call protocol compliance now emits cancellation and rejection outputs for unexecuted tools, preventing model errors when expected tool results are missing.

224 KiB Raw Blame History Unescape Escape

Changelog

0.2.101 — 2026-07-13

Features

Bug Fixes

0.2.100 — 2026-07-13

Features

Bug Fixes

0.2.99 — 2026-07-12

Features

0.2.98 — 2026-07-12

Breaking Changes

Features

Bug Fixes

0.2.97 — 2026-07-11

Features

Bug Fixes

0.2.96 — 2026-07-10

Features

Bug Fixes

0.2.95 — 2026-07-09

Features

Bug Fixes

Performance

0.2.94 — 2026-07-09

Features

Bug Fixes

0.2.93 — 2026-07-08

Breaking Changes

Features

Bug Fixes

0.2.92 — 2026-07-08

Features

Bug Fixes

Performance

0.2.91 — 2026-07-07

Bug Fixes

0.2.90 — 2026-07-07

Features

Bug Fixes

0.2.89 — 2026-07-07

Features

Bug Fixes

Performance

0.2.88 — 2026-07-06

Features

Bug Fixes

0.2.87 — 2026-07-05

Features

Bug Fixes

0.2.86 — 2026-07-04

Features

Bug Fixes

0.2.85 — 2026-07-03

Features

Bug Fixes

0.2.84 — 2026-07-03

Features

Bug Fixes

Performance

0.2.83 — 2026-07-02

Features

0.2.82 — 2026-07-02

Features

Bug Fixes

0.2.81 — 2026-07-01

Features

Bug Fixes

Performance

0.2.80 — 2026-07-01

Features

Bug Fixes

Performance

0.2.79 — 2026-06-30

Features

Bug Fixes

Performance

0.2.78 — 2026-06-30

Features

Bug Fixes

224 KiB

Raw Blame History