Close active panels, halt streams, or clear an empty prompt
Since 1.2.10 pressing Esc while the suggestions dropdown is open closes the dropdown first; a second Esc is needed to interrupt the agent's turn.
Since 1.2.17 dismissible announcement cards for model launches and service notices appear above the prompt, one at a time, newest first; pressing Esc on an empty prompt dismisses the current card permanently, and sending a message hides it for the rest of the session.
1.4.6 fixed the Esc disabled during file edits warning naming the wrong reject key when the prompt shows an extra choice, such as for edits outside the workspace in accept-edits mode.
Named in the 1.3.5 release notes alongside the new --mode auto value and the /config Agent Mode option; absent from the cli/reference global-controls and prompt-focus keybinding tables.
1.4.4 extended auto to Windows 11 version 24H2 and later for personal Google Accounts; previously Linux/macOS only.
Promoted to verified on the 2026-11-12 sync: cli/reference's keybinding table now lists this binding as cli.cycle_mode, cycling default > accept-edits > auto > plan.
Paste media files or clipboard blocks into the prompt
1.4.4 fixed media attachments with parameterized MIME types, such as audio/webm; codecs=opus, being rejected as unsupported.
1.4.8 fixed pasting images on Windows when they were copied from a browser, Snipping Tool, or another app that puts PNG or V5 bitmap data on the clipboard, and made pasting retry when another app briefly holds the clipboard.
1.4.9 added image pasting through the terminal's own paste shortcut, such as Cmd+V or Ctrl+Shift+V with only an image copied, and by dragging image files into the prompt; this now also works over SSH with X11 or Wayland forwarding, and the paste key names the clipboard tool to install, such as wl-clipboard, when one is missing.
1.4.5 added a blank line between a collapsed thought block and the response that follows it, and changed the collapsed block to show only its Thought for … header instead of a snippet of the thinking; expand it to see the full thinking.
1.5.4 fixed a thought that follows a tool's result line, such as Read 22 lines, being indented as if it were part of that tool call in this expanded view.
1.5.8 added click-to-expand in full screen (altscreen) rendering mode: clicking a tool group, a thought, a prompt's attached images, or multi-line command output now expands or collapses just that item, as an alternative to this shortcut.
1.5.9 fixed tool groups jumping around when expanded: the collapse hint now stays on the same line as the expand hint, and in low verbosity the token summary line no longer disappears from the expanded view.
1.2.8 fixed this hanging with a Vim: Warning: Output is not to a terminal error for full-screen editors like vim; the editor now runs against a real terminal stdout.
1.5.8 fixed opening your external editor on Windows when EDITOR or the editor setting is a normal Windows path such as C:\Tools\ed.cmd; backslashes were dropped, so this shortcut, /open, /keybindings, and artifact review failed with an editor ... not found in PATH error.
1.4.4 fixed this mishandling pasted text, merging typing with recalled history entries, recording no-op Backspace or Delete steps, and truncating characters near emoji.
1.5.0 fixed undo in the main prompt bringing back a message you had already sent, and undo in ! shell or : search mode bringing back prompt text that Enter would then run as a command or search.
1.5.6 fixed undo and redo, including u in Vim mode, moving the cursor to the end of the input; the cursor now returns to where the undone or redone edit was made.
Named in the 1.2.6 release notes alongside Ctrl/Alt+arrow keys, which stop at punctuation as well as whitespace; absent from the reference keybinding table.
Named in the 1.2.6 release notes; since that release deletion stops at punctuation, so one press removes a single path or URL segment. Absent from the reference keybinding table.
Send using the opposite of your Queued Messages setting
Added in 1.4.0 (Option+m on macOS when Option acts as Alt): send immediately while in Queue mode, or hold a message in Send Immediately mode. Named only in the release notes; absent from the reference keybinding table.
1.5.2 changed the macOS default so Option-key shortcuts such as this one work out of the box without setting Option to act as Alt in your terminal, where previously pressing them did nothing or typed a character such as µ; set AGY_CLI_MAC_OPTION_AS_ALT=0 to go back to the old behavior.
Promoted to verified on the 2026-11-12 sync: docs/settings?tab=cli now documents this shortcut directly alongside the queuedMessages setting, though it is still absent from the keybinding table itself.
1.5.4 fixed paging skipping text: a page that ended exactly on a message boundary could jump past the whole next message, and in full screen (altscreen) rendering mode the lines hidden under the pinned prompt header were scrolled past without ever being shown.
1.5.4 fixed paging skipping text: a page that ended exactly on a message boundary could jump past the whole next message, and in full screen (altscreen) rendering mode the lines hidden under the pinned prompt header were scrolled past without ever being shown.
1.4.7 trimmed the shortcuts tab by leaving out basic navigation keys such as the arrow keys, so the more useful shortcuts are easier to find.
1.5.8 fixed side questions asked here, like /btw, failing with stream timed out after 5m0s while the question was still working; they now run until they finish or you press Esc.
Summarize the conversation on demand to free up context
Added in 1.3.5: on its own it summarizes and ends the turn; with a message it summarizes then answers. Shows a cancelable Compacting conversation... progress label and a step/size-reduction summary when done. A command by this name was removed as unverified from the first draft of this placemat (v1.2.2) for not existing yet; still absent from the cli/reference core slash command table.
1.4.4 improved this and automatic compaction to keep tool declarations on the summarization request, so the model API can reuse its cached prompt prefix.
1.4.6 improved compaction to keep images, PDFs, and other attached files in the conversation summary, so the agent can still see them afterward.
1.4.7 fixed conversations compacting over and over when a single message, tool output, or attached document was too large to fit in the context.
Roll back conversation history to a previous message
Reverts workspace files to the matching snapshot as well as the transcript; 1.2.6 fixed the lookup sweeping back to an older snapshot and silently overwriting kept work, and skips the elided diff of any edit over 1 MiB.
Since 1.2.9 each step shows a relative timestamp, and the panel focuses the most recent step when it opens.
1.3.1 fixed very long conversations listing older, space-saving-compacted messages as (empty message) and erroring when one was picked; such messages now show their original text, dimmed and labeled cleared to save space, can't rewind here, and can no longer be selected.
1.5.8 fixed this failing with a step_index out of bounds error when you pressed Esc while a turn was still starting and then rewound to remove that message.
Open the conversation picker overlay to load a previous thread
1.2.14 sped up resuming conversations with thousands of steps by no longer scanning every step up front before loading.
1.2.14 fixed resuming a conversation whose history had a missing step, for example after a crash or an interrupted write, hiding the most recent steps and letting new messages overwrite them.
1.2.15 fixed the picker's search box wrapping long queries so their beginning scrolled out of view; it now uses the full terminal width.
1.3.2 fixed the picker crashing when you pressed Enter to import a conversation from the Other tab on a narrow terminal, and fixed conversation titles with CJK characters being cut off mid-character.
1.3.4 fixed @conversation: mention suggestions getting stuck on Fetching matches for ... when search is slow; after 2 seconds the list now falls back to your local conversations.
1.4.8 improved resuming a conversation to use about 10% less CPU by not rendering already-finished tool calls a second time.
1.5.9 sped up opening the picker in long sessions by no longer loading every open conversation's full history first, and let /resume <id> take any ID prefix that matches exactly one conversation, such as the short ID shown in the status line, reporting conversation <id> not found for an unknown one instead of opening the picker.
Since 1.2.16 highlighting a setting and pressing ←/→ switches it to the previous or next value and saves it right away, without opening the dropdown.
Since 1.3.5 an Agent Mode entry here picks between accept-edits, plan, and the new auto execution mode; see --mode on card-cli.
1.5.1 renamed the Local and Remote tabs here during --remote-agent sessions to TUI and Remote Agent, so it is clear which tab configures the terminal UI and which the remote agent; the settings on each tab are unchanged.
Choose your preferred reasoning model for the session
1.2.4 added an inline argument autocomplete dropdown when typing /model and a fuzzy search bar inside the picker that filters by name or ID.
Since 1.2.6 a /model <name> submitted while a turn is running waits for the queued prompt instead of switching the in-flight turn.
Since 1.3.4 this picker hides third-party models your Google Cloud project can't access for Business sign-in on pay-as-you-go, instead of failing once you send a message.
Open the interactive tool permissions manager panel
1.3.4 fixed "Always allow" rules approved from a Remote Control browser session still applying after you removed them in the browser's rules editor or this panel's shared tab; such rules now save only to the shared config.json.
1.4.9 fixed "Always allow" rules for multi-line commands, such as a python -c or node -e script spanning several lines, never taking effect; existing multi-line rules now apply.
1.4.9 changed execute_url permissions so approving a Google site such as www.google.com or cloud.google.com no longer also allows sensitive Google apps such as mail.google.com, docs.google.com, drive.google.com, accounts.google.com, or console.cloud.google.com; each now needs its own approval.
1.5.6 fixed editing project rules here clearing the marker that records the project's permission rules as already migrated, which made the Antigravity app migrate them again and could drop command(...) allow rules when the sandbox is on.
1.5.7 fixed saved command allow and deny rules being ignored, so every command asked for approval, when an organization's admin policy requires commands to run in the terminal sandbox but the machine has no OS sandbox; your saved rules now apply and unlisted commands still ask before running.
1.5.8 fixed "Always allow" command rules on Windows being bypassable when the shell is cmd, cmd.exe, powershell.exe, pwsh.exe, or a full path to one of them; commands run in PowerShell or Command Prompt now always need an exact match, since a command with hidden extra statements could otherwise be approved by a rule meant only for a safe one.
Set or inspect the active model's reasoning effort level
Added in 1.2.11 per the release notes, alongside the same gauge inside /model and the --effort flag; promoted to verified now that cli/reference's core slash command table lists it, with levels ranging from low through max depending on the model.
Enable fast mode, bypassing reasoning plans for quick actions
Marked unverified on the 2026-11-04 sync: absent from the current cli/reference core slash command table and from every changelog entry checked so far.
1.3.3 fixed custom alt keybindings such as alt+s not working with the macOS Option key in terminals that don't treat Option as Meta, such as Terminal.app and VS Code-based terminals.
1.3.3 fixed this silently doing nothing when the configured editor can't be found; it now shows an error.
1.4.6 changed the in-progress status line to show what the agent is waiting on, such as initializing MCP servers or Model unavailable (attempt 2/10), retrying in 12s with a live countdown, instead of a generic working label.
1.5.3 added a model field to each subagents entry in the status line JSON, naming the model that subagent is using.
1.5.8 fixed status messages such as "Copied 3 lines (42 characters)" after selecting text, settings save errors, and permission reload notices never appearing when you use a custom status line; they now briefly replace it, which comes back once the message clears, without shifting the layout.
Open the Agent Manager Panel for background subagents
Since 1.2.9 typing @<subagent> <message> sends a message directly to a subagent's conversation, with autocomplete listing running and completed subagents.
1.2.11 fixed project custom agents in .agents/agents/ not being found or selectable here, in workspaces created by the Desktop App or already trusted, under execution with --agent, or in headless (-p / --prompt) runs.
Since 1.2.16 the agent hands image-generation requests to a built-in image-generator subagent, which writes the prompt, checks results with up to three attempts, and saves images to the conversation's artifacts; generating an image now appears here as a subagent run.
1.3.1 fixed Markdown custom agents ignoring skills:, plugins:, rules:, agents:, and hooks: entries written as paths relative to the agent's own file when the agent ran as a subagent or as the main agent; such entries now resolve against the agent's own directory.
1.3.1 fixed subagents that stopped on an error, such as running out of quota or model capacity, showing as Done here; they now show Error: with the reason, which clears once the subagent makes further progress.
1.3.3 fixed subagents started with a branch or share workspace in a multi-repository workspace silently working in your real checkout, and fixed a stopped subagent staying listed here with a pending approval instead of clearing it right away.
1.3.6 fixed questions asked by subagents and workflow agents being lost, which left them waiting forever: an <agent> has a question alert now appears, and Ctrl+K or a on it opens a picker that sends your answer; headless -p runs skip these questions so the subagent can continue.
1.4.3 fixed the agent losing track of its most recently started subagents after a conversation was compacted; the summary now keeps only each subagent's type, role, and result instead of its full initial prompt.
1.4.4 fixed subagents losing plugins inherited from the parent session, and fixed plugin-bundled subagents with plugins.inherit_user: false and restricted MCP tools failing to start with mcp spec not found for provider.
1.4.8 fixed subagents failing to run when their custom agent's name contains a colon, and fixed deleting a conversation leaving its running subagents, their saved history, and their shared or branched workspaces behind; they are now stopped and cleaned up with it.
1.4.10 fixed subagents losing the task their parent agent gave them, and the parent's ID, after their context was compacted, including with /compact; a subagent's summary now keeps the parent agent's 10 most recent messages.
1.5.1 improved blocked subagent rows to name the tool they are waiting on, such as Blocked · AskQuestion, instead of a generic Blocked · Processing....
1.5.3 added each subagent's model next to its name in the subagent rows above the prompt and in /agents (hidden on narrow terminals), with the full model and effort shown in a subagent's detail view; /agents now shows a single Tip: Press enter on a subagent to view details. line instead of repeating [view details] on every row.
1.5.4 fixed crashes and freezes: the CLI no longer crashes when a subagent's question or approval prompt is drawn in a very narrow terminal or before the terminal size is known, a subagent's question no longer closes by itself while you are answering it, and opening a subagent's view no longer occasionally freezes the CLI.
1.5.5 fixed subagents and swarm workers that get their own copy of a Git repository starting from the last commit and losing uncommitted work; their worktree now includes staged, unstaged, and untracked changes, and a subagent or worker outside version control can get an isolated copy of the workspace folder, for folders up to 100,000 entries and 4 GiB.
1.5.8 fixed custom agent and rule files saved with a UTF-8 byte order mark (BOM), as Windows PowerShell, Visual Studio, and Notepad's "UTF-8 with BOM" option do, going missing from here and the agent picker or losing their frontmatter trigger.
Since 1.2.7 LaTeX math ($$...$$, fenced math/latex/tex blocks) and Mermaid diagrams render inline via Kitty graphics, with a Unicode ASCII fallback and Left/Right arrow panning for wide diagrams; press m to cycle Image, ASCII, and Raw views.
1.2.9 added a footer hint when a wide diagram overflows the window, pointing at the Left/Right pan keys.
1.2.10 pre-renders diagrams in the background on Kitty-compatible terminals so they appear as soon as the viewer opens, fixed Ctrl+= / Ctrl+- zoom flickering between sizes, and combined the footer's separate top/bottom hints into one g/G top/bottom entry labeling m with the view it switches to.
1.2.11 restored the ASCII fallback for Mermaid diagrams and LaTeX in WezTerm and the VS Code integrated terminal, which had been rendering as garbled placeholder characters, and fixed copying text when Copy on Select is off in /config — the viewer no longer captures the mouse, so the terminal's own selection and copy shortcut work, and Ctrl+C goes back to interrupting or exiting.
1.2.12 coalesces bursts of scroll events before redrawing the full-screen view, cutting CPU use and scroll lag.
1.2.13 clarified the viewer's m hint to name the content and the mode it switches to, such as diagram ASCII, diagram source, or math image, instead of generic labels like toggle ASCII or toggle raw.
1.2.15 fixed large WebP images opened by the agent not being scaled down to fit image size limits the way PNG and JPEG images are.
1.2.16 improved mouse text selection in the altscreen conversation view and this viewer: dragging past the top or bottom now auto-scrolls, the selection stays on its text while content streams or you scroll with the wheel, and copying includes lines that scrolled off screen.
1.2.17 fixed .tiff images being reported with the wrong file type and silently converted to PNG when the agent viewed them.
1.3.2 lowered the image size images are scaled down to, to at most 2576 pixels per side (previously 4096), so conversations with many images stay within Claude models' limits.
1.3.4 fixed a conversation failing on every later request after the agent read a truncated or header-only WebM file; such files are now rejected when read instead of being attached.
1.3.6 fixed Mermaid diagrams with YAML frontmatter, a %%{init: ...}%% directive, %% comments, or Mermaid v11 id@{ shape: ..., label: ... } node syntax showing as raw source instead of rendering, and fixed / search in this viewer highlighting the wrong characters or crashing on lines with letters whose lowercase form has a different length, such as İ.
1.4.3 improved n/N search navigation here to step through matches one at a time, including several on the same line, instead of jumping from line to line.
1.4.9 fixed conversations failing every later model request with INVALID_MEDIA after the agent viewed a corrupt image or one with text in front of its header, such as a screenshot captured with adb exec-out screencap -p that also captured a warning; such images are now rejected with an error that says where the real image header starts.
1.4.10 capped attached images at 2000 pixels per side and sent large PNGs as JPEGs when that is much smaller, to prevent requests from being rejected for exceeding the model's size limit, and fixed an image saved as JPEG but labeled PNG, such as a pasted image or browser screenshot, being rejected for a mismatched type.
1.5.8 fixed / search here missing text split across different styles, such as a sentence containing a number or inline code, and leaving the rest of a styled line unstyled after a highlighted match; the current match is also now easier to pick out in every color scheme.
1.2.15 fixed signing in to MCP servers whose OAuth client registration responds with HTTP 200 instead of 201, which previously failed with "registration failed with status 200".
1.3.1 fixed restarting a server from here sometimes leaving it stopped, with an error such as "Failed to stop existing instances for reload" when the old process took too long to exit; the server now always starts again.
1.3.4 fixed remote MCP servers that use OAuth signing you out after a temporary network failure, timeout, or server error; saved sign-ins are now cleared only when the server or its authorization server actually rejects them.
1.4.1 fixed a server continuing to show as Ready after its process exited unexpectedly, which left later tool calls failing with client is closing: EOF; it's now marked Error and its closed session is removed.
1.4.4 fixed MCP client sessions and background connection routines leaking when tools/list failed during server startup, or when the CLI shut down while a server was still connecting.
1.4.7 increased the inline result size for on-demand MCP tools to 10 KB; previously anything over 4 KB was saved to a file the agent had to read in an extra step.
1.4.7 fixed plugin and MCP server sign-ins not being saved on macOS and Windows when the token was too large for the system keychain, storing such tokens in the CLI's private secrets file instead, and fixed a plugin MCP server that signs in with OAuth being able to use a token saved by another user of the same computer account, or keeping the previous user's connection after switching accounts; tokens now belong to the signed-in user.
1.5.0 improved the error for a stdio server whose command isn't installed, which now says, for example, "uvx" is not installed or not on PATH, install it or set "command" to its full path instead of a raw executable file not found in $PATH error.
1.5.1 fixed stdio MCP servers on Windows leaving their child processes running after the server stops, which could lock plugin folders and make plugin upgrades fail because the files were in use.
1.5.2 fixed signing in to an MCP server that uses OAuth failing with Invalid PKCE code_verifier when the server reconnected during sign-in, such as when this manager reloaded servers or a new turn started, so the sign-in link you already opened keeps working.
1.5.8 fixed permission rules for MCP tools, such as an mcp(server/tool) deny or ask rule, being ignored for MCP tools loaded up front (marked eager in mcp_config.json) and for the /browser subagent's Chrome DevTools tools; these now ask for approval or are blocked like other MCP tools.
1.5.8 fixed remote MCP servers that sign in with OAuth without a client secret signing you out as soon as the access token expired, typically after an hour; token refreshes now send the client ID the way the server requires, including for sign-ins saved by earlier versions.
1.5.8 changed the agent telling you to click a Sign in button or use an Accounts tab, which the CLI doesn't have, when a plugin MCP server needs a sign-in; this panel now shows the same /plugin sign-in guidance as /plugin itself for a server waiting on one.
1.5.9 fixed a race where reloading MCP servers while several subagents were starting could launch duplicate server processes, leaving some subagents with broken MCP tools or failed sign-ins.
1.5.9 fixed MCP server and plugin OAuth sign-in on machines without an IPv4 loopback address, such as IPv6-only VMs, by falling back to [::1] when 127.0.0.1 is unavailable instead of failing silently.
/skills reload (1.2.4) re-discovers skills and slash commands asynchronously, without restarting the session or blocking input.
1.2.16 fixed the / menu listing built-in skills that the current agent does not enable, which had inserted instructions for tools the agent could not use when one was selected.
1.5.2 added skills from the cloud plugins you connected with /plugin to --remote-agent sessions: they appear in the / menu, and this command lists them grouped by plugin under Remote agent skills; /skills reload lists them again.
Open the interactive Plugins Manager, or run install/enable/disable/list inline
Since 1.2.11 a plugin placed directly in ~/.gemini/config/plugins whose MCP server needs configuration variables starts disabled until you enable it, matching plugins installed this way.
1.3.1 fixed installing, reinstalling, and uninstalling a plugin on Windows while one of its MCP servers was running; the CLI now stops the plugin's MCP servers before replacing or deleting its folder and restarts them afterward, and tells you to close any other Antigravity process still holding the plugin open. Plugin manifests, mcp_config.json, and hooks.json files saved with a UTF-8 byte-order mark now load, and installing from a marketplace no longer overwrites an existing plugin folder that came from a different source.
1.3.2 opens this manager onto the official Antigravity plugin marketplace: browse and install plugins bundling skills, MCP servers, subagents, rules, and hooks, install one from your shell with agy plugin install <plugin-name>@antigravity-plugins-official, or pick + Install from local directory...; the plugin list now opens instantly, an interrupted install never leaves a half-installed plugin, and Windows installs/removals are more reliable.
1.3.3 improved search in the Installed tab: type spaces in a query, press Enter to finish searching, then use Space to toggle the selected plugin and Enter to expand its details; typing again resumes the search.
1.3.4 fixed agy plugin validate, the summary agy plugin install prints, and marketplace previews in this manager's Discover tab leaving out subagents a plugin declares in its root agents.json.
1.3.6 fixed /plugin subcommands such as /plugin List failing with Unknown subcommand when not typed in lowercase; they now work in any case.
1.4.3 fixed updating or reinstalling a marketplace plugin on Windows leaving it half-deleted when another agy session, the Antigravity desktop app, or the Remote Control background service had it open; the installed plugin now stays intact until the replacement is ready.
1.4.5 fixed plugins failing to load when plugin.json gives skills as a single path string instead of an array, as some plugins written for other tools do, and fixed marketplace installs and automatic updates failing with a checksum mismatch or network error when a download was briefly stale or interrupted by retrying the download once.
1.4.6 added a Configure MCP servers... row for installed plugins, so you can set or change a plugin's MCP server variables at any time without disabling and re-enabling it; plugins with missing values show ⚠ needs setup, and saving restarts an enabled plugin's servers so the new values take effect.
1.4.9 grouped the Discover tab's marketplace plugins into the same sections as the Antigravity plugin marketplace instead of one flat list, listing a plugin under every section it belongs to.
1.4.11 improved the Discover tab for long shelves: each shelf now shows its first 6 plugins followed by a ⋯ N more — enter to show row, Ctrl+E expands or collapses the shelf under the cursor, and [ and ] jump to the previous or next shelf.
1.5.0 added sign-in for a partner plugin's MCP servers when you install it from here, with a Connect <plugin> screen for each server that opens its sign-in page (Esc skips), a ⚠ needs sign-in marker and a Sign-in: row for the plugin in the Installed tab, and a third-party developer tools notice for partner plugins in the Discover list, in a plugin's details, and after install.
1.5.3 improved the plugin sign-in screen here, which now indents the sign-in link like the surrounding text and sets it apart with a blank line, no longer underlines trailing blank space in wrapped links, and uses the plugin's display name from plugin.json, such as Canva instead of canva.
1.5.3 fixed the agent's requests to sign in to a plugin being dropped from the conversation, so when a plugin needs a sign-in the CLI now shows a line such as ⚠ The <plugin> plugin needs you to sign in. Open /plugin to sign in.
1.5.8 fixed this, agy plugin, and headless -p runs ignoring your organization's Gemini Enterprise admin controls for plugins: when an administrator disables plugins, installing, importing, or enabling one now fails with plugins are disabled by your organization, and when the marketplace is disabled, marketplace catalogs are not fetched and marketplace installs fail; listing, disabling, and uninstalling still work.
1.5.8 changed the agent telling you to click a Sign in button or use an Accounts tab, which the CLI doesn't have, when a plugin needs a sign-in; it now tells you to sign in here and then send a message to continue.
1.5.9 fixed signing in to a plugin from here sometimes reporting that the sign-in didn't finish when it had just succeeded.
Delegate web browsing, UI checks, and in-browser debugging to a subagent
1.4.9 fixed this subagent not receiving the current page's snapshot at the start of each turn with Chrome DevTools MCP 1.11 and later, because the snapshot request left out the now-required page ID and failed silently.
1.4.10 fixed this subagent not receiving a page's snapshot when the selected page belongs to an isolated browser context, such as a page opened with isolatedContext or attached from an Incognito window.
1.5.1 fixed the built-in browser agent being listed among the agents you can start a conversation with; it still runs only as a subagent.
1.5.9 switched this subagent's session recordings from .webm to .mp4 files, made with Chrome's built-in screen recording.
Capture session corrections as a reviewable rule or skill proposal
1.5.8 fixed this repeatedly appending to rule files such as GEMINI.md or AGENTS.md until they grew past the rule size limit and their endings were silently cut off; it now checks the documented size limits before updating a rule and, when close to or over them, proposes condensing existing entries, splitting them into a scoped rule, or moving step-by-step guidance into a skill.
Set a one-shot timer or recurring schedule that wakes the agent
1.4.5 made recurring schedules fire at a consistent per-schedule offset instead of exactly on the cron boundary, to avoid traffic spikes; schedules repeating within a day fire somewhere within their period, fixed-time schedules still fire within their scheduled minute, and @every schedules are unchanged.
1.5.9 fixed scheduled automations on Windows receiving only the first line of a multi-line prompt; each scheduled or manually triggered run now gets the full prompt.
Open the Task Manager Panel for background shell logs
Since 1.2.5 a task sent to the background with an explicit name keeps it in completion notifications and the task list.
1.2.16 fixed the agent stalling when a background command it was waiting on crashed or was killed by the system, for example from an out-of-memory condition; it now sees the command finish with its exit code and keeps going, and headless -p runs no longer exit before the agent can respond to a background command that finishes after the first turn.
1.3.1 extended this panel, the active task list, and the status line's task count beyond background shell commands to also show timers, recurring jobs, and other background tools, reported as completed or failed without a made-up exit code.
1.4.0 fixed file-change, task, and session-cost tracking reverting to an earlier point whenever a background command finished while the agent was still working on something else.
1.4.0 fixed a background command's exit output sometimes being dropped from its task log instead of always being captured on completion or cancellation.
1.4.3 changed timer labels here and in the active task list to show durations in days, hours, minutes, and seconds, such as 1d 2h timer or 5m timer, instead of raw seconds such as 93784s timer.
1.5.1 sped up how quickly the agent moves on after a terminal command exits while a background process it started still holds the command's output open, to at most about 1 second instead of 5.
1.5.7 fixed the agent not noticing that a command it moved to the background was waiting for input, such as a password or y/n prompt printed before the command went to the background; the agent is now notified once that output settles instead of waiting silently.
1.5.7 fixed background task details from manage_task losing their formatting when the task's recent output, or input sent to the task, contained Markdown code blocks or runs of backticks.
1.5.9 fixed quitting the CLI or killing a task here leaving the programs it started running and the task marked as still running; these now end along with the command.
Launch collaborative multi-agent teams for long-horizon work
/teamwork-preview marked unverified on the 2026-11-04 sync: the current cli/reference core slash command table lists only /teamwork, with no alias.
1.5.5 fixed two runs in the same workspace sharing .agents/teamwork/ and overwriting each other's run state; each run now keeps its state in its own folder under the CLI's app data directory.
Start a swarm of worker agents that collaborate on a shared message board
Added in 1.5.5 as a preview, labeled (Preview & Quota Heavy) in the / menu because each worker is a separate agent that uses your quota. The main agent becomes a coordinator that adds worker agents, each in its own conversation, working on the task in parallel.
Workers can be forked from an existing member's conversation, run on any model in the /model picker, and pause instead of retrying when their model runs out of quota; a monitor worker alerts the coordinator when the whole swarm stalls.
A ● Swarm(<title>) row appears in the active-items tray below the prompt while a swarm runs, showing the latest post and post count and then Done when it finishes; press Enter on the row to open /swarm board, or Alt+X to dismiss it.
1.5.6 fixed the coordinator sometimes ending its turn with a one-line filler reply instead of its real answer after being sent back to re-read board posts it had already read.
1.5.6 fixed the coordinator offering a sidecar:// board link the CLI cannot open instead of /swarm board on later turns, and fixed agy -p runs that start a swarm pointing you to /swarm board, which headless runs cannot open.
Added in 1.5.5, also opened with Alt+S (rebindable as swarm.board). Shows a Roster of members with each one's state and model, a Posts feed, and a Message pane that renders the selected post as Markdown.
Tab / Shift+Tab move between panes, z zooms the focused pane, Enter shows or hides the message, [ / ] step through posts, p filters to one member's posts and mentions, o opens a member's subagent details, f follows new posts, and Esc backs out.
1.5.6 added bee visuals: a queen bee in the board header and a 🐝 beside each live member in the Roster, with an animated colored bee card per member on Kitty-graphics terminals such as Kitty and Ghostty, and the queen shown as pixel art elsewhere when the window is at least 60×35.
1.5.6 fixed pressing Esc in a swarm member's conversation opened from here with o dropping back to the prompt; it now returns to the board with the same selection, filter, focused pane, and zoom.
Coordinate many subagents with a Python script run in parallel or staged phases
Added in 1.5.9 as Dynamic Workflows: the agent writes and runs a short Python script that starts agents with parallel() or pipeline(), grouped into named phases you can follow live; each started agent works in its own conversation and returns a structured result, keeping file reads, searches, and test logs out of your main conversation's context. Available on Google AI Pro and Google AI Ultra plans.
Runs in a separate Python process bundled with the CLI, so no local Python install is needed; ask the agent to save a working script to your project to reuse it later.
List, create, rename, and switch between local projects
Added in 1.5.0: a project groups a set of workspace folders, agent settings, and conversation history; create one from a name and folders given as absolute, ~, or workspace-relative paths, change a project's name or folders, or switch to a project, which adds its folders to the workspace and starts a new conversation. Named only in the release notes as /projects; absent from the current cli/reference core slash command table and from docs/projects, which documents only the --project and --new-project flags and the /fork <project_id> move.
1.5.2 fixed adding PDFs, other binary files, or files larger than about 8 MB to a remote project from here in --remote-agent sessions, which failed and now upload successfully.
Ask a side question without interrupting the main thread
1.3.3 fixed this running your user and workspace hooks, which could make a stop hook loop or add extra steps to a quick question.
1.5.0 let side questions call tools, such as reading files or searching, showing each call live in the /btw card and a Called N tools summary once the answer is done; a tool call that would ask for permission is denied automatically, which also fixed questions failing with an empty-response error when the model wanted to use a tool.
1.5.8 fixed this and /help side questions failing with stream timed out after 5m0s while the question was still working; they now run until they finish or you press Esc, and a side question that needs an approval the card can't show stops right away and tells you to ask in the main conversation instead.
1.4.10 improved the after-compaction size shown here and in the compaction marker to also count what the summary carries into the next prompt, such as earlier requests and artifact, subagent, and running-task snapshots, so it is closer to the real size; the (N summary) label still shows the summary alone.
1.5.3 fixed paths that share a prefix with your home directory, such as /home/alice when your home is /home/al, showing with a mangled ~ prefix in the file list here and in the keybindings reset hint.
/copy btw (1.2.3) copies the full text of the active /btw side-question response even when its card is collapsed; typing /copy shows argument ghost hints.
cli/reference now also documents /copy <n>, copying the n-th most recent agent response instead of only the latest.
1.5.1 fixed copy-on-select and this command inside GNU screen, so copied text now reaches your clipboard and long selections no longer leave stray characters in the terminal.
1.5.9 fixed this and other clipboard actions on Windows over SSH: the copy now reaches your local clipboard, and the copied text is no longer written to the CLI log file.
Since 1.3.0 j/k in the file view move the line cursor like the up/down arrow keys, matching the file list and Vim, instead of jumping to the next or previous file; use ←/→ to switch files.
1.3.1 improved file navigation: ←/→ now open the next or previous file at its first change with context in view, each file remembers its own cursor and scroll position when you switch away and back, and n/N continue into the next or previous file's changes instead of stopping at the current file's last change.
1.3.1 added a right-aligned change A/B · file X/Y position indicator to the file view header, counting the block under the cursor and the file's place in the file list.
1.3.2 fixed Tab, →, and ← in the Commit Tree view moving the cursor instead of switching views as the footer hint says, and added Shift+Tab to switch to the previous view in every tab.
1.3.3 fixed this showing No uncommitted changes in Git, Mercurial, and Jujutsu repositories configured to always color output, such as with color.ui = always, and fixed the error shown when the fallback git diff also fails to surface its real cause instead of the original git diff HEAD error.
1.3.5 added syntax highlighting to the file view, with a background tint on added/removed lines and a Terminal color scheme that matches your terminal's light or dark background; added folding, with Ctrl+O expanding or collapsing the fold nearest the cursor and f switching between only the changes and the whole file; and improved performance on very large files by rendering only the visible rows, though expanding folds and showing the whole file turn off above 100,000 lines or 10 MB.
1.3.6 colorized the Commit Tree view: commit hashes and change IDs are highlighted, descriptions are dimmed, the Working Copy (Uncommitted Changes) row stands out, and @ is bold.
1.4.0 reworked the Commit Tree view: the header count now reflects commits rather than lines, the cursor is confined to commit rows with Enter opening the selected one, the view expands to fill the terminal, and up to 100 recent commits scroll into view instead of just 10.
1.4.3 added / search across the file list and file view: matches highlight as you type, the search prompt shows how many matches the diff and each file contain, Enter jumps to the first match, and n/N step through matches across every file, with a match A/B indicator in the file view header.
1.4.4 extended the whole-file toggle (f) and hunk folding (Ctrl+O) to the Commit Tree view, consolidated Per Turn diffs so multiple edits to one file in a turn appear as a single combined diff, and fixed long rows wrapping instead of truncating at the terminal width, and Per Turn headers truncating prompts at 50 characters.
1.4.5 counted matches inside folded unchanged lines during / search, such as match 2/5 · +3 hidden, marked the selected commit in Commit Tree the same way as All Changes and Per Turn, and made Per Turn start with the cursor on the first file instead of its turn header.
1.4.8 added [ and ] to jump to the first file of the previous or next turn in Per Turn, sped up opening in large repositories and long conversations by showing tracked changes immediately and loading untracked files and older turns in the background, and listed large or binary untracked files by size instead of reading them into memory.
1.4.9 changed the position indicator under long lists to say what it counts, such as [1-44 of 195 files] in All Changes and [1-44 of 195 rows] in Per Turn, where turn headers count as rows.
1.5.0 fixed this drawing an empty ↕ ··· fold divider at the top of new files, and of edits starting at line 1, where nothing was folded away.
1.5.8 fixed this showing a made-up change at the end of a turn's diff for files larger than 500 KB when an earlier edit changed the file's length; the two versions are now cut off at the same line, so the preview shows only real edits.
Toggle a session-scoped Remote Control link for the active terminal session
Added in 1.2.6; follow and control the session from another device. /remote-control off or closing the session tears down the tunnel and unregisters the device.
1.3.3 fixed starting this while signed in with Application Default Credentials; ADC isn't supported yet, so this command is now hidden and --remote-control exits with a clear error in that case.
1.4.1 fixed this rejecting on/off typed with uppercase letters, such as /remote-control Off.
1.5.2 made Remote Control connections retry a brief network or server error while setting up a new connection instead of dropping it.
1.5.7 fixed Remote Control showing a link that could not be opened at the end of some agent replies, such as the swarm coordinator's link to the swarm UI; these links are now left out.
1.5.9 fixed the model selector in the Remote Control browser view on phones, where the list could not be scrolled by touch past the first screen, so models near the end of the list could not be selected.
Since 1.2.15 -p "/usage" prints quota reset times in your local time zone when run interactively; piped and --json output keep UTC timestamps, so existing scripts are unaffected.
Since 1.3.3 reaching your individual quota is reported right away instead of being retried for a while first.
Since 1.2.8 recordings automatically stop and finalize at 3 minutes 30 seconds, so dictation past 4 minutes no longer fails with a deadline error.
1.5.2 fixed voice dictation in Chinese and Japanese showing a space between every character in the live draft before the final transcript arrived, and fullwidth Latin words losing the spaces between them.
Launch an interactive Antigravity CLI coding session
1.4.0 fixed every message failing with an Eligibility check failed error, requiring a restart, after a brief service outage or rate limit at launch; such errors now retry instead of locking the session.
1.5.0 fixed certain models failing a turn on rate-limit (429) and server (5xx) errors instead of retrying them, now honoring the server's Retry-After header.
1.5.3 fixed agent turns failing with http2: timeout awaiting response headers when the model took more than 60 seconds to start responding, which are now retried.
1.5.3 fixed every model request failing with a 404 error when an enterprise model gateway URL ended in /v1beta, even though the model list loaded; a trailing /v1beta in the gateway URL is now ignored.
1.5.5 changed enterprise model gateway URLs to require https:// unless the gateway runs on the same machine (localhost, 127.0.0.0/8, or ::1); a remote gateway configured with a plain http:// URL is now rejected instead of receiving your prompts unencrypted.
1.5.5 fixed enterprise model gateways that validate the request format rejecting every request, because the CLI sent a non-standard sessionId field in Gemini-format requests.
1.5.6 fixed some models served through an enterprise model gateway being sent requests without native tool declarations, which kept the agent from calling tools reliably; every model in the gateway's catalog, including models listed in AGY_LLM_GATEWAY_MODELS, now gets native tool calls.
1.5.6 fixed custom Gemini models served through an enterprise model gateway or an admin model list, such as tuned models, cutting off long edits and file writes at 8,192 output tokens; they now default to the same 65,535-token output limit as built-in Gemini models, and an explicit admin maxOutputTokens value still takes precedence.
1.5.7 fixed Gemini 4 Carbon requests through an enterprise model gateway failing with an HTTP 400 Thinking budget is not supported for this model error; the CLI now sends Gemini 4 Carbon a thinking level and no longer leaves the Low, Medium, or High suffix in the model ID it sends to the gateway.
1.5.8 added support for Gemini 4 Hydrogen through an enterprise model gateway.
1.5.8 added two admin settings, gateway.routeWebFetch and featureControls.remoteControlEnabled (see card-settings); with an enterprise model gateway, web pages are now fetched directly from your machine, through proxy.address when set, instead of through Google's fetch service, unless an admin sets gateway.routeWebFetch to disable URL reading.
1.5.8 fixed follow-up turns in a conversation resumed this way ending on a plain-text reply with no structured result if it was started with --json-schema and an earlier turn had produced structured output.
Run prompt non-interactively and print response to stdout
Since 1.2.6 an agent or model API failure prints a structured AGY_ERROR: {...} JSON line on stderr (status, HTTP/gRPC code, retryability, error ID) and exits with code 3 instead of 1.
1.2.9 fixed these runs leaving daemon background processes running past exit, which could hang scripts reading the CLI's output until end-of-file; daemon processes now terminate when the run ends.
1.2.10 fixed runs that streamed part of a response and then ended on a model or agent error exiting with code 0; they now exit with code 3 and print the AGY_ERROR line, and JSON error output includes the partial response, while multi-turn stream-json sessions still warn and continue.
1.3.4 fixed combining this with --prompt-interactive exiting with status 0 after reporting the conflict; it now exits with status 2 like other print-mode usage errors.
1.5.2 stopped these runs from making a background model request to generate a conversation title that print mode never shows.
1.5.5 fixed these runs sometimes never exiting after printing their result when a background task, such as conversation title generation, got stuck; the CLI now finishes shutting down within 30 seconds.
1.5.8 fixed these runs, along with /plugin and agy plugin, ignoring your organization's Gemini Enterprise admin controls for plugins.
Execute initial prompt and continue interactive session
1.5.9 fixed this together with --remote-agent and --conversation or --continue: the prompt is now sent to the resumed conversation after its history loads instead of starting a new one, and is handled like typed input, showing right away, waiting behind a running turn, and reporting a failed send.
Set agent execution mode (accept-edits, plan, auto)
auto added in 1.3.5: approves file edits and safe commands without asking, still prompts before irreversible actions, and has a separate agent review commands that would leave the terminal sandbox; also chosen from the /config Agent Mode entry or by pressing Shift+Tab. Linux and macOS only, with Windows support coming later. 1.3.6 made that sandbox-escape review faster, let it allow the workspace's own format/lint/fix commands instead of asking, and made it block commands that would upload sensitive data. Named only in the release notes; not present in any agy --help this card was verified against.
1.4.3 fixed subagents and custom agents in Auto mode asking you to approve every command that would leave the terminal sandbox; their commands are now reviewed by the same separate agent as the main agent's, so safe commands run without asking.
1.4.4 extended auto to Windows 11 version 24H2 and later for personal Google Accounts, and made the sandbox-escape review consider the assistant's preceding question on short replies like yes, check every subcommand in a pipeline, and prompt before destructive Git or file commands.
1.4.7 improved Auto mode to ask for approval less often on routine commands, such as cleaning up temporary files, stopping a local process, or running a code formatter.
1.4.8 gave the sandbox-escape review more time to decide and stopped any single slow check after 10 seconds, so it times out less often and asks you to approve safe commands less often.
1.4.9 made the sandbox-escape review aware of whether the working directory is inside a Git or other repository, so it no longer spends its time budget running git status in folders that aren't repositories.
1.5.0 improved this review to time out and ask less often on more safe commands, such as uploads that don't request a review, stopping a job the agent started, installing or removing packages from a registry, and removing files that can be restored from version control.
1.5.4 sped up this review by using a lighter-thinking Gemini Flash model for it, so it decides faster and times out less often.
1.5.6 made this review ask less often on safe reads: GitHub API reads, fetches into an existing clone, and single-file downloads such as raw.githubusercontent.com URLs no longer get the repository popularity, age, and release checks that cloning a repository still gets, host names that start with a digit are no longer mistaken for IP addresses, and reading ~/.ssh/config, known_hosts, authorized_keys, or public keys (*.pub) is no longer treated as reading secrets, while private keys are still protected.
1.5.9 sped up this review further by letting it search the conversation, list directories, and search files directly instead of starting a sandboxed command for each check, so it decides faster and times out less often.
Run the session's conversations on Remote Agents instead of this machine
Added in 1.5.0, default false; with --project, conversations open in the named remote project, otherwise in your default remote project. In these sessions @-mention completion searches the remote agent's files, artifacts that can't render in the terminal get an Open in browser link, the header shows remote workspace instead of your local directory, and the / menu no longer offers skills from this machine.
1.5.3 fixed files larger than the remote read limit being silently cut off in these sessions, which now load in full up to 20 MiB and report an error above that instead of showing partial content.
1.5.4 fixed the Sign-in update needed warning, which asks you to run /logout and sign in again when your sign-in is missing a newly required permission, not appearing in these sessions, where only a cloud conversations are not enabled error was shown.
1.5.8 improved the first message of a conversation here, which no longer waits for its remote environment to start or resume: the CLI now prepares a new conversation's environment as soon as the session knows its remote project, and starts a resumed conversation's idle environment in the background as soon as it opens.
1.2.11 fixed this flag not finding project custom agents defined in .agents/agents/ across workspaces; see /agents.
1.3.4 fixed a name matching no agent silently running the default agent instead: headless -p runs now exit with code 1 and list the available agents, and interactive sessions show an Agent Not Found warning before falling back to the default agent.
Run inside sandbox with terminal execution restrictions
1.4.6 restricted sandboxed commands from reading or modifying ~/.docker, ~/.kube, or ~/_netrc by default; previously ~/.docker and ~/.kube were writable inside the sandbox, so commands that use Docker registry or Kubernetes cluster credentials now need an explicit grant.
1.5.1 hardened how plugin OAuth credentials are protected here: a plugin's real token now replaces its stand-in only in a request's Authorization header, so a sandboxed command can no longer get a website to echo the real token back by sending the stand-in in another header or in the URL.
Auto-approve tool permission requests without prompting
1.3.2 stopped this flag from also auto-approving the agent's implementation plan in planning mode; it now only skips tool permission prompts, so plans wait for your review again. Set artifactReviewPolicy in /config if you want plans and other artifacts to proceed automatically.
1.5.2 changed this flag so commands on the built-in list of protected commands are no longer approved automatically and go through the usual checks, while every other command still runs without asking; this flag and plugin agents set to auto-allow also no longer approve terminal commands or file access outside the workspace when your Gemini Enterprise administrator requires those to be reviewed, and headless runs now deny such requests instead.
1.5.2 fixed subagents stalling for about 60 seconds on file permissions under this flag, so reads and writes outside the workspace now go through without a prompt, and headless -p runs no longer wait on any remaining interaction, while your explicit deny rules still apply.
Output format for print mode (text, json, stream-json)
1.4.11 added a finished_reason field to json and stream-json results that says why a turn ended early — print_timeout, interrupted, or error — and that is left out when the turn finishes normally.
Enforce structured output using JSON schema string or file
Since 1.2.14 this flag rejects plain text, bare type names such as string, and missing schema files, and requires the schema's root to be "type": "object"; invalid input now fails at startup with an error and exit code 1 instead of silently being treated as a string schema.
1.5.8 fixed follow-up turns in a conversation using this flag, once resumed with --conversation, ending on a plain-text reply with no structured result after an earlier turn had produced one; each new message now gets its own reminders to return the structured output.
Timeout for a print-mode run; unset, the run continues until the response completes
Before 1.2.6 headless runs defaulted to 5m0s; pass this flag explicitly to impose a limit.
1.2.9 fixed background tasks being canceled about 5 seconds after the agent went idle; runs now wait for background tasks until this deadline, capped at 30 minutes.
1.5.0 changed headless runs without this flag to wait for background tasks to finish, where they previously stopped them after a hidden 30-minute limit and exited with success; when this flag now cuts a turn short, the run exits with code 3, prints an AGY_ERROR line naming the terminated tasks, and writes a [agy] print timeout after ... line to stderr.
1.3.3 fixed validate and plugin listings missing skills declared through the skills field in plugin.json or ignoring skills.json exclusions, and fixed a race that could drop results when importing plugins.
1.3.6 fixed a manifest named Plugin.json not being found by install, validate, enable, and disable, since plugin names and manifests were case-sensitive where the agent was not.
1.4.0 rewrote validate to apply the CLI's own plugin-loading rules instead of a simplified check, surfacing each error and warning by file and line and flagging common plugin.json mistakes; --strict now fails on warnings and --json gives a machine-readable report.
1.4.1 expanded validate to also check a plugin's skills, subagents, and rules.
1.4.3 expanded validate further to check a plugin's hooks, sidecars, and component manifests, and made install and import follow NO_COLOR, CLICOLOR, and CLICOLOR_FORCE the same way validate already did, instead of always coloring piped output.
1.4.4 made validate stricter across plugin.json, MCP server configs, hooks.json, and sidecar manifests, catching invalid JSON syntax, unknown fields, bad variable placeholders, unrecognized hook tool matchers, and missing hook scripts.
1.4.6 expanded validate to check a plugin's context providers (missing label or command, invalid or duplicate triggers, unknown fields, missing command files) and to warn about skills entries in plugin.json that don't exist or contain no SKILL.md.
1.4.8 expanded install output to name each installed skill, agent, rule, command, MCP server, and hook and to print the install folder, and expanded validate to check a plugin's author, homepage, and repository fields in plugin.json.
1.4.9 made validate warn when a plugin's name isn't lowercase kebab-case, such as Neo4j or foo_bar, which the marketplace can't publish, and added checks for a plugin's logo and context provider iconUri paths.
1.5.0 fixed auto-upgrades of a plugin installed under a different folder name, such as after a rename, writing to a new folder and retrying on every startup instead of updating the existing folder in place.
1.5.1 fixed these commands being unavailable, with no plugin marketplaces listed, when you sign in with Application Default Credentials (AGY_ADC_AUTH) or through an enterprise model gateway.
1.5.2 hardened import, install, uninstall, enable, and disable against malicious plugins: they now reject plugin names that are not safe file names with invalid plugin name, and refuse to write outside the plugins folder or through symbolic links, so an imported plugin can no longer overwrite other configuration files such as mcp_config.json.
1.5.3 fixed uninstall, enable, disable, and uninstalling from /plugin, failing for plugins installed as symbolic links, such as plugins installed from Antigravity, or for plugins whose names contain spaces or symbols such as c++-tools, and fixed list omitting symlinked plugins.
1.5.5 fixed import turning Claude Code plugin commands into skills that did not load or were all named SKILL; each command now becomes a skill named after its folder, including commands with no frontmatter or with Claude Code-style values such as argument-hint: [base branch] [focus].
1.5.5 fixed import copying certain plugins' hooks/hooks.json unchanged, which caused the plugin's hooks to be skipped; PreToolUse, PostToolUse, Stop, and startup SessionStart hooks are now converted, with tool names in matchers mapped to built-in tools and a default 60-second timeout.
1.5.6 improved validate to warn when a plugin hook command uses ${PLUGIN_ROOT} or $PLUGIN_ROOT, which older CLI versions do not set, and to suggest a relative path instead, since hooks already run in the plugin's folder; the warning fails validation only with --strict.
1.5.6 fixed install and import leaving a half-copied plugin that still loaded as enabled when an install failed partway; installs are now swapped in all at once and the previous version is restored on failure, plugins containing symlinks that point outside the plugin, FIFOs, or device files are refused, and a command whose file name cannot be used as a skill folder name is skipped with a warning instead of failing the import.
1.5.6 fixed install ignoring the skill folders that a Claude Code plugin lists in the "skills" field of its manifest; those skills are now installed and loaded, and the install summary warns about listed paths that are missing, contain no skill, or point outside the plugin.
1.5.8 fixed /plugin, this command, and headless -p runs ignoring your organization's Gemini Enterprise admin controls for plugins; an administrator disabling plugins or the marketplace is now enforced instead of silently ignored.
1.5.9 fixed validate running all of its checks on a hooks.json saved with a UTF-8 byte order mark (BOM) and reporting the real syntax error and line instead of blaming the BOM.
1.2.10 fixed machines set up with the old Remote Control installer script crash-looping after an upgrade: when launched by that deprecated background service, the CLI now unregisters it and points to remote-control start instead of restarting repeatedly.
1.2.14 improved remote-control start on Linux machines without a systemd user service manager, such as most containers: instead of failing, it starts the daemon as a background process that won't restart after a crash or at boot; remote-control status shows its PID.
1.3.4 improved Remote Control so the browser's Settings modal shows the settings the CLI actually uses (project and global permissions, queued messages, AI credits, telemetry, and artifact review) instead of saying settings can't be changed there, and agy remote-control serve sessions now read and save changes back to your saved settings.json.
1.4.2 fixed conversation history not showing up here when running in daemon mode.
1.4.3 fixed Remote Control on iPhone and iPad asking you to accept the Security Notice & Data Use terms and showing the Enable Notifications prompt again every time you reopened it; it now remembers those answers, along with your theme, layout, and dismissed tips.
1.4.4 added a multi-pane layout with touch-sized controls for tablets and foldables instead of collapsing to the phone view, and kept popovers open while dragging scrollbars; also fixed clipped headers/dialogs and iOS Safari input zoom on phones, Enable Telemetry/Marketing Emails reverting while Settings loads, dragged URLs losing query or # parameters, broken nested code blocks, $VAR shell variables parsed as LaTeX, broken media previews, blurry or blank diagrams in Copy Image/Export as PDF, and background-tab timer drift.
1.5.4 fixed this background service staying disconnected for good when it started before the network was up, such as right after a reboot, and stopped it from printing a sign-in link it could not complete when checking the saved sign-in only hit a network error.
Run as an Agent Client Protocol server over stdio for ACP-capable editors
Added in 1.5.0: your editor launches this as a subprocess and talks JSON-RPC over stdin/stdout, so editors such as Zed, JetBrains IDEs, Android Studio, and Xcode can run the agent in their own chat panel; the agent, tools, sign-in, and settings all come from Antigravity CLI. Sign in with an interactive agy session, or run agy acp --login to reuse the cached credentials.
1.5.0 gave this prompt steering by sending another prompt mid-turn, default/accept-edits/plan session modes, session/resume restoring a session without replaying its history, each session keeping its own model and thinking level, images and embedded file context in prompts, and Google sign-in from a terminal without a browser.
1.5.1 widened model access here to every model available to your account for every ACP client, where previously only a fixed list of editors were offered models other than Gemini, and fixed session bugs: closing, reloading, or resuming a session while a prompt runs now ends that prompt as cancelled instead of leaving it hanging, cancelling a prompt no longer also stops the next prompt sent right after it, session/update notifications no longer arrive after the matching session/prompt reply, and permission requests and question forms now close when answered elsewhere or their prompt ends.
1.5.2 fixed the / menu listing the commands and skills of a default agent instead of those of the agent the session actually runs, so in these sessions it now shows that session's agent, including a custom agent's commands.
1.5.3 fixed these sessions exiting before the agent backend finished shutting down when the client closed the connection, which could interrupt saving conversation data.
1.5.4 added Auto mode to these sessions, so editors can switch a session to auto wherever the CLI itself offers Auto mode; it approves edits inside the workspace and runs commands in the terminal sandbox, and still asks before commands that would leave the sandbox unless Auto mode's review clears them.
1.5.4 fixed these sessions dropping back to the default mode when an editor reloaded or resumed an open session, so a session in accept-edits no longer starts asking for permission on file writes again.
1.5.4 fixed turns hanging when a tool call needed a second permission approval, such as a command that asks again after its first approval; the editor now receives a new permission request instead of the turn waiting forever.
1.5.8 added context window usage updates to these sessions, so editors such as JetBrains AI Chat can show how much of the model's context window the conversation is using; the value is sent when a session starts or is reloaded, after each model call, and after a model or thinking level change.
1.5.9 fixed several issues here: a session no longer fails with Authentication required after agy acp --login when settings.json names an auth method with no credential, such as gemini-api-key without GEMINI_API_KEY, because it now falls back to your saved login; prompts no longer fail when the ACP client passes an MCP server whose name has spaces or other special characters; a failed turn now reports its specific error instead of a generic execution-terminated message; the diffs sent for completed edits no longer carry an extra trailing newline; and a settings.json saved with a UTF-8 byte order mark (BOM) no longer makes this command fall back to default settings and sign-in options.
Also accepts light/dark/solarized/colorblind-friendly/tokyo night variants.
1.3.5 fixed colors being reduced to 256 in iTerm2 over SSH, where COLORTERM isn't forwarded; full 24-bit color is now used whenever LC_TERMINAL=iTerm2 is set, unless COLORTERM or NO_COLOR says otherwise.
1.4.1 brought this scheme to the inline diff preview shown when the agent proposes to edit or create a file: syntax highlighting, background tints on added/removed lines (picked from your terminal's light or dark background under the Terminal scheme), and an untinted line-number gutter.
1.4.7 brought box-drawing table borders to the default Terminal color scheme in Markdown tables, matching the other color schemes instead of ASCII | characters.
1.2.16 improved resize performance in no-flicker (altscreen) mode: resizing the terminal no longer freezes or lags the CLI, and the full-screen view redraws on-screen messages first instead of re-rendering the whole conversation at the old width.
1.3.3 added a pinned prompt header in this mode while you scroll back through a long conversation: the prompt that produced the response you're reading stays at the top, and clicking it jumps back to that prompt.
1.5.0 renamed the always option's label in /config's Rendering Mode entry from no flickering (altscreen) to full screen (altscreen); the saved setting is unchanged. Named only in the release notes; docs/settings?tab=cli still showed the old label as of the 2026-11-06 sync.
Prompt editing model: default (flat) or vim (modal)
Since 1.2.9 Vim mode supports numeric count multipliers on operators, motions, and actions in Normal and Visual modes, including 3dw, 2d3w, 3dd, and counted text objects such as 2di(.
1.2.12 fixed Vim mode ignoring non-ASCII characters after r/f/F/t/T, fv/fV leaving Visual mode instead of extending the selection, and D/C/Visual ~/u/U recording an empty undo step and clearing redo history when nothing changed.
1.2.15 fixed the Vim mode cursor stopping inside emoji such as 👩💻 or characters with combining marks such as Devanagari; h, l, and other motions now move over whole characters, and x/r act on the whole character instead of a hidden part of it.
1.2.16 fixed k in Normal mode on the top line recalling the last command-history entry instead of your queued messages, which could send a queued message twice; k now pulls queued messages back into the editor like the Up arrow does.
1.4.7 fixed Vim mode's r replace, in Normal and Visual mode, leaving part of a collapsed [Pasted text #N] placeholder behind; the whole placeholder is now replaced.
1.4.9 fixed Vim mode's r, f, F, t, and T canceling when the target character is a space, and fixed a half-typed operator such as d carrying over after you submit a prompt.
Visual verbosity level for tool-call and thought output
Historically high (full output) or low (minimal); the 1.2.10 release notes describe a new medium level between the two that groups related tool calls and thoughts into concise summaries while keeping commands and responses visible.
1.2.12 fixed resuming a long conversation at medium or low verbosity taking many seconds and lagging while history replayed; earlier tool groups now appear already finished instead of re-animating.
1.3.0 changed the default from high to medium for anyone who never explicitly picked a verbosity, including anyone who had selected high while it was still the default.
1.4.4 fixed the prompt box staying mid-screen after closing panels, dropdowns, or inline prompts, and blank gaps appearing above the prompt in Medium verbosity mode. docs/settings?tab=cli now documents all three levels (high/medium/low) directly, confirmed on the 2026-11-04 sync.
1.5.7 fixed agent replies stalling for 20 seconds or more while a group of tool calls with long output was still animating its token counts; streamed steps now keep appearing while the group re-renders.
Follow-up message handling during an active turn, default "queue"
Added in 1.2.14 and exposed in /config; set "send-immediately" to interrupt the agent with queued follow-ups instead of holding them until the turn ends — this key was previously ignored when set directly in settings.json.
Promoted to verified on the 2026-11-04 sync: docs/settings?tab=cli now documents it as the Queued Messages entry, press Alt+M to send one message the other way.
1.2.16 fixed "send-immediately" messages sent while earlier messages were still queued joining the queue instead of going to the agent together with them, in order; a message that fails to send mid-turn now returns to the queue instead of being lost.
1.3.1 fixed "send-immediately" messages being silently queued until the end of the turn whenever the agent had already written or updated an artifact, such as a task list or plan, during that turn; they now reach the agent right away.
1.5.2 fixed the agent re-running a failed turn in a rapid loop when messages were queued at the moment it hit an error, so queued messages now wait for your next message.
1.5.3 fixed a message sent during a turn with the Send Immediately queued-messages mode briefly disappearing from the conversation and reappearing when it arrived just as the turn ended.
1.5.8 fixed a single empty or corrupted queued message file, for example one left behind by an unclean shutdown, blocking every later message to the agent; the unreadable file is now set aside and the remaining messages are delivered.
Session execution mode: auto, default, accept-edits, or plan
Confirmed by name in the docs/settings?tab=cli CLI settings reference on the 2026-11-04 sync, which had previously only been named in release notes and the /config Agent Mode entry (see --mode and Shift+Tab on the other cards); auto is the default where available and approves edits and safe sandboxed commands automatically, default prompts for diff review and follows toolPermission for commands, accept-edits auto-approves edits only, and plan prepends /plan before any edits.
1.5.7 changed sessions with no saved mode, including headless -p runs, to start in auto wherever it's available and default elsewhere; choosing Default for Agent Mode in /config keeps the previous behavior by saving an empty string here.
Let file read/write tools navigate outside the workspace root
1.5.2 changed the prompt shown when the agent asks for access to a file outside the workspace, which now reads Allow access to this file? and offers to always allow or deny that one path for the conversation or in settings.json, instead of turning this setting on for everything outside the workspace.
Restrict local execution commands to OS containment rings
Since 1.2.10 the agent asks to bypass the sandbox less often: it knows sandboxed commands can read and write its own artifact and scratch directories, and it retries the sandbox first even after an earlier command needed a bypass.
1.2.17 improved the Windows command sandbox so sandboxed commands no longer need administrator rights, and common tools such as Python, Git, and npm work inside it out of the box.
1.3.3 changed the macOS sandbox to block sandboxed commands from reading ~/Library/Application Support by default, since desktop apps keep credentials and session tokens there; it's still readable if you allow it explicitly in your sandbox rules.
1.3.3 fixed sandboxed commands failing with Access is denied. on Windows when PowerShell 7 was installed from the Microsoft Store; the agent now falls back to Windows PowerShell in that case.
1.4.1 fixed agent commands failing to start on Windows versions that don't support the terminal sandbox (Windows 10 and Windows 11 before 24H2) when a sandboxed permission mode was selected: the CLI now turns this off and falls back to asking before running commands, and the Remote Control browser Settings modal hides sandboxed permission options on those systems.
1.4.3 fixed the CLI offering options that need the terminal sandbox on systems where it can't run at all, such as Linux systems that block unprivileged user namespaces and Windows before 11 24H2.
1.5.3 fixed a security issue where a command run in the terminal sandbox could change environment state later loaded by persistent terminal commands run outside the sandbox, and hardened persistent terminal IDs against path traversal.
1.5.6 fixed duplicate approval prompts and sandbox errors for organizations that enforce sandbox mode on machines without an OS sandbox, such as some Windows hosts or Linux with user namespaces disabled; commands now ask for review once and then run.
1.5.8 fixed a security issue in this sandbox's network proxy where a command could get around path-based network rules, such as those limiting where credentials are sent or which Git and API paths are allowed, by using percent-encoded slashes, backslashes, or .. segments in a URL path; every way of reading the path is now checked against the rules.
Admin setting disabling URL reading unless a gateway proxy address is set
Added in 1.5.8, so an enterprise model gateway's web page reads can't go around it, unless proxy.address is configured. Named only in the release notes; absent from docs/settings?tab=cli as of this sync.
Admin setting that stops all plugins, including built-in ones, from loading
Added in 1.5.9 for organizations that don't use Gemini Enterprise; also blocks installing or enabling plugins. Gemini Enterprise accounts keep following their organization's Cloud Console policy instead. Named only in the release notes; absent from docs/settings?tab=cli as of this sync.
Use personal AI credits once plan quotas are exhausted
1.2.15 fixed requests made once both the plan quota and this AI credits balance are exhausted: instead of retrying for about two and a half minutes and then showing a generic execution-terminated error, the CLI now immediately says the AI credits balance is too low to continue.
Allow usage-data collection to help improve Antigravity
1.4.1 shows this as off (disabled) with an explanation for Business accounts, such as Google Cloud project and Application Default Credentials sign-in via AGY_ADC_AUTH, where prompts, content, and model responses are never collected.
Marked unverified as a settings.json key on the 2026-11-06 sync: docs/settings?tab=cli now describes Enable Telemetry as saved with your account instead of in settings.json, unlike useG1Credits, showTips, and showFeedbackSurvey on the same page, which still show their settings.json key names.
Authenticate with a Gemini API key instead of Google sign-in
Also set modelProvider: "gemini" in settings.json; suits headless and CI runs. The CLI reads this credential only from the environment, not from .env files.
Since 1.2.6 headless runs authenticated this way keep daemon background commands, such as dev servers, running after the turn finishes; they still terminate once the run itself exits (see -p).
Since 1.2.12 a session using this key stops immediately when the Gemini API reports an exhausted daily quota, a spend cap, or depleted prepaid credits, instead of retrying for several minutes; short per-minute rate limits still retry.
1.3.1 fixed sessions using this key failing a turn as if it had been cancelled when the connection to the Gemini API was dropped internally; the CLI now retries the request instead.
1.4.3 fixed Gemini 4 Argon not being shown in /model and --model when connected with this key.
1.5.0 fixed conversations using this key re-sending a prompt forever when the API's safety filter blocks it, which now ends the turn with a content-filter message, and fixed the pro and flash subagents on this connection running models two generations older instead of Gemini 4 Argon and Gemini 3.8 Flash.
1.5.8 added Gemini 4 Hydrogen (Low, Medium, High, and XHigh) and Gemini 4 Carbon (Low, Medium, and High) to /model and --model when connected with this key.
Authenticate using Google Application Default Credentials
Named in two independent release notes: 1.3.3 (this sign-in method doesn't yet support --remote-control, so that flag and /remote-control are disabled under it) and 1.3.6 (Claude Opus 5.5 and Sonnet 5.5 become available this way, gated by what your Google Cloud project can access). Absent from every docs page in sources.json; unverified pending confirmation from an auth/install reference.
1.4.1 changed first-run setup under this sign-in method to show the Google Cloud Platform Terms of Service without the consumer telemetry opt-in checkbox.
1.5.5 added Gemini 4 Carbon (Low, Medium, and High) to /model and --model when signed in this way.
1.5.6 changed pay-as-you-go Google Cloud accounts that sign in this way to use the same permission settings as other pay-as-you-go accounts, making Auto mode available to them; Gemini Enterprise accounts keep following their organization's admin-controlled permission settings and are not offered Auto mode.
1.5.8 added Gemini 4 Hydrogen (Low, Medium, High, and XHigh) to /model and --model when signed in this way.
Force Kitty graphics image mode for the artifact viewer
Named in the 1.2.10 release notes for terminals with image passthrough configured inside tmux or GNU screen, which otherwise fall back to ASCII diagrams; the install/auth docs that source this card don't mention it.
1.2.12 fixed the CLI sending its image-support probe through screen, tmux (including iTerm2 tmux -CC tabs), or Zellij, which corrupted the terminal or tab title; the same release notes name CLI_GRAPHICS=kitty again as still forcing image mode afterward, so the row is promoted to verified on two independent changelog mentions.
1.2.15 changed the SSH image-support probe to ask the terminal to identify itself first; over SSH only Kitty and Ghostty are now treated as image-capable, fixing WezTerm and Konsole showing garbled image placeholders and macOS Terminal.app leaving stray escape-code text behind after exit.
Added in 1.4.0: on a supporting terminal, the tab or window title reflects whether the CLI is working, waiting on input, done, or failed, after a capability check; set off to disable, or on to force it when a terminal supports this but fails that check. Named only in the release notes; absent from the install/auth docs sourced for this card.
Restore the pre-1.5.2 macOS behavior where Option-key shortcuts need terminal setup
Named only in the 1.5.2 release notes alongside the new macOS default for Option-key shortcuts such as Alt+M; absent from the install/auth docs sourced for this card.
Stop the CLI from extracting and running its bundled ripgrep
Added in 1.5.5, for example when endpoint security software blocks the bundled binary; searches then use a ripgrep on your PATH, or the built-in search if none is found.
1.2.15 fixed the agent being blocked from reading global rules and customization files in ~/.gemini/config, including the rules/ folder, AGENTS.md, GEMINI.md, skills.json, rules.json, plugins.json, and agents.json; it can now read them and asks before editing them, while other files in that folder stay off-limits. The same release fixed global rules being added to context more than once when ~/.gemini/GEMINI.md, ~/.gemini/AGENTS.md, or their ~/.gemini/config/ counterparts are symlinks to the same file.
1.4.4 fixed the system prompt's wording so GEMINI.md and AGENTS.md rules explicitly override earlier built-in instructions too, instead of only ones that follow them.
Workspace-level rule files, 12,000 characters each
Backward-compatible with the older .agent/rules path.
Since 1.2.7 user and workspace rules get a dedicated 20,000-token budget, cutting oversized rules on newline boundaries, so large rule sets no longer evict skills, workflows, subagents, or MCP tools.
1.2.16 fixed skills.json, rules.json, and other customization manifests in a parent .agents/ directory being ignored when a session started in a subdirectory; manifests now load from every .agents/ directory between the working directory and the project root.
Workspace-level skill definition with YAML frontmatter
Workspace customizations in .agents/ override global and built-in skills.
1.3.6 fixed skills, rules, workflows, and Markdown agents failing to load with mapping values are not allowed in this context when an unquoted description: line contains a colon, such as description: Runs checks: formatters and linters.
1.4.4 fixed rules and SKILL.md skills loading twice when discovered from both a standalone customization folder and an enabled plugin.
1.5.2 restricted plugins and skills listed in a workspace's .agents/plugins.json or .agents/skills.json to loading only from inside the workspace or its repository root; entries that point elsewhere through absolute paths, ../, or symbolic links are now skipped.
Config file in .agents/ or ~/.gemini/config/ mapping hooks to events
1.2.4 fixed these configurations being silently dropped while customization token-budget truncation was active.
1.4.7 added lastUserInput to the JSON that hook commands receive on stdin, so a hook can inspect the latest message you sent, for example to screen prompts before the agent acts on them.
1.5.3 added the PLUGIN_ROOT and PLUGIN_DATA environment variables to hook commands that come from a plugin, pointing to the plugin's folder and its persistent data folder, so plugin hooks can use files bundled with the plugin and keep state between runs; agy plugin validate no longer warns that these variables are not set for hooks.
1.5.3 fixed hooks receiving a placeholder or empty model name when the session uses a custom model, so hooks now get the name of the model actually in use.
1.5.4 added hookEventName, cwd, and toolCall.id to the JSON input hooks receive, so a single hook command can handle several events, knows the agent's working directory, and can match a PostToolUse call to its PreToolUse call.
1.5.8 fixed hook commands on Windows that contain double quotes, such as type "C:\my dir\deny.json", failing and being silently skipped, which let a PreInvocation or PreToolUse hook that should block the agent go unenforced; hook commands are now passed to cmd.exe exactly as written.
1.5.8 fixed a hooks.json saved with a UTF-8 byte order mark (BOM), as Windows PowerShell, Visual Studio, and Notepad's "UTF-8 with BOM" option do, leaving /hooks empty with none of its hooks running; such files are no longer rejected as malformed.
1.5.7 fixed a hook returning auto_approve, or a wildcard command(*) allow rule, bypassing an organization's admin-enforced review: commands that would leave the terminal sandbox, commands the admin policy holds for review, and file reads or writes outside the workspace under an "always ask" file-access policy now always ask for your approval.
1.4.3 fixed a hook that fails to run here — for example a missing command, a non-zero exit, a timeout, or invalid JSON output — marking the tool call it follows as failed even though the tool succeeded; the failure is now logged and the tool keeps its result.
1.4.11 added terminationBehavior: a handler returning "terminate" stops the turn before the model is called, and can set an errorMessage whose userMessage reaches you without being sent to the model, which also makes a -p run exit with a non-zero code. Absent from docs/hooks, which still documents only injectSteps for this event's output.
1.5.6 added a continuation_active field to this hook's JSON input, true while the agent is already working on a continuation an earlier Stop hook requested, so a hook can tell it already asked and avoid looping.
Manifest bundling skills, rules, and MCP configurations
Named in the 1.3.2 release notes: the manifest can also live in a .antigravity-plugin/ or .gemini-plugin/ folder, list skills directly in plugin.json, store data in a per-plugin folder exposed as PLUGIN_DATA, and ship command-line tools in a bin/ folder the agent can run by name; absent from docs/plugins.
Named in the 1.3.3 release notes: a root-level agents.json next to skills.json and rules.json lets a plugin declare its own subagents; absent from docs/plugins.
1.3.4 added a --json flag for machine-readable output, and fixed this showing only imported plugins; it now lists every installed plugin, including ones installed from a marketplace, with its version and whether it is enabled.
Import compatible plugin bundle from Claude ecosystem
1.4.11 fixed this reporting No claude extensions found. for plugins installed from a Claude Code marketplace: it now reads Claude Code's installed-plugin list, skips ones disabled there or installed for a single project only, keeps imported plugins' remote MCP servers ("type": "http" with url and headers) and servers from an .mcp.json file without an mcpServers wrapper, expands ${CLAUDE_PLUGIN_ROOT} to the installed plugin's directory, and lists a server it can't convert, such as one using the legacy SSE transport, with a warning instead of reporting it as installed.
1.4.0 fixed plugin MCP servers sending a literal ${TOKEN}-style placeholder instead of its configured value in a request header, matching how the server's command, arguments, and environment variables already resolved.
1.4.4 added _meta["antigravity.google/root_conversation_id"] to MCP tools/call requests so servers receive the top-level conversation ID even when a tool is called from a nested subagent.
1.5.2 fixed MCP tools that declare their own explanation, toolSummary, toolAction, or waitForPreviousTools parameter receiving calls without that argument, so the value the model sends now reaches the tool.
1.5.9 fixed a security issue where custom headers configured here for an HTTP MCP server, such as API keys, could be forwarded to a different host when the server responded with a redirect; they are now sent only to the configured server's own origin.
Server-side flag that marks an MCP tool as always loaded
Named in the 1.3.3 release notes: an MCP server sets this on a tool in its tools/list response so the agent gets that tool directly instead of discovering it on demand; absent from docs/mcp.
Per-plugin store of API key values for its bundled MCP servers
Named in the 1.4.12 release notes: a key entry can be marked "optional": true for a service with a free tier, so when you haven't set the key, the plugin's commands see it as unset and make unauthenticated requests at the free rate limit instead of failing with 401 errors or prompting you to sign in; absent from docs/plugins and docs/mcp.