Google Antigravity runs the official CLI
1 feature · 20 fixes · 6 performance changes
Features
Google Antigravity runs the official CLI
The agy built-in now launches the official Antigravity CLI as agy acp instead of requiring a separately installed standalone ACP server. Install the CLI with Google's instructions and confirm agy acp --help succeeds; an interactive agy sign-in is reused, so the previous server archive, its login, and its separate model catalog are no longer involved. The provider ID, display name, authentication mode, permission mode, and session records are unchanged.
Model and effort now come from the CLI's live catalog. On the verified CLI 1.5.9, seven logical models are advertised and each carries its own thought_level option: Hydrogen offers low, medium, high, and xhigh, while Argon offers three levels; max is unsupported. CompozyOS selects the model first, then applies the advertised effort and confirms the provider's acknowledgement before the first prompt. Effort is never silently lowered to a value the model does not offer. Availability depends on your Google account, CLI version, and client identity, so the observed list is not a fixed built-in catalog.
Under Full access, Antigravity now selects a mode its CLI actually advertises (auto on the verified version) instead of the previous injected yolo. Host permission decisions are unchanged, and no permission-bypass flag is added.
A live rewalk showed that the CLI runs its own write_file without asking the host even in its plan mode: an approve-reads session acknowledged mode=plan, gemini-4-hydrogen, and xhigh, then created a file with no permission request. Because CompozyOS cannot mediate that boundary, a native agy acp start is refused before any prompt under every policy below Full access — deny-all, approve-reads, approve-edits, and auto. The refusal never widens the policy and is never worked around by an agent raising its own level: choose Full access for a native Antigravity session, or keep the narrower policy on a runtime or authored route whose tool calls return through CompozyOS. An authored standalone or custom command keeps its route and its own approval behavior. For the same reason, a native agy acp start under Full access is refused when permission_rules contains a deny entry; allow entries are accepted because they never narrow what runs. The check recognizes the CLI under both executable names Google ships, agy and antigravity.
Antigravity sends its questions as permission requests whose options are the possible answers. CompozyOS now routes those questions to the operator as a choice prompt under every policy, including Full access, and cancels the request when no operator can answer instead of approving the first answer. Every Antigravity session also disables the CLI's run_workflow tool, which otherwise writes skills/system/run-workflow/SKILL.md into the workspace.
Migration notes:
- Install and login: install Antigravity CLI, then run
compozy provider auth login agyandcompozy provider auth status agy --remote. A standalone-only installation reportsmissing_cli; a standalone-server login alone reportsneeds_login. CompozyOS never copies or deletes either Google token store, and the remote probe never starts consent. For headless or SSH setup, Google'sagy acp --loginis the operator alternative and is never run as a background probe. - Existing overrides: an authored provider/agent/fallback command and any explicit provider-private ACP options keep their exact bytes and precedence. An option the CLI no longer offers fails with the existing actionable validation error; CompozyOS does not silently rename or convert it. Remove an obsolete standalone command override yourself when you want the new built-in route.
- Upgrades and continuity: a session whose executable route changed continues the same logical session from its retained transcript through the existing replay path. This is not native-history import: Google's own history is untouched and CompozyOS does not claim provider-private context was transferred. A session without verifiable accepted-route metadata also replays rather than assuming an old native ID belongs to the CLI store. Sessions recorded by earlier releases can carry an effective provider mode without a separately recorded selection intent — for example, the old injected
yolo— so CompozyOS cannot distinguish an automatic mode from an explicit one. It preserves that ambiguous mode instead of deleting or remapping it; if the new CLI does not offer it, the start fails with the existing unsupported-option repair guidance and keeps the session and transcript, rather than silently dropping it. To establish current intent, select the session's provider, model, and effort again withcompozy session runtime set <id> --provider agy --model <advertised-model> --reasoning-effort <offered-effort>; that writes a complete next-prompt snapshot without the obsolete raw mode, and the same logical session then replays on a fresh native ACP session ID.compozy session runtime clear <id>returns to the effective/agent default and is not that confirmation. A session whose accepted route still matches an authored standalone command keeps that route; only a mismatched route replays. After the new native bind, later restarts restore that session natively when the provider supports it. - MCP: the default hosted-tool transport stays stdio. If you author an HTTP MCP configuration for Antigravity, Google requires its
--process-per-sessionoption; CompozyOS does not enable HTTP or a different worker mode implicitly.
Fixes
The daemon boots with a missing project, stops silent turns on time, and installs updates right after a release
- Boot with a renamed or removed project folder. If a registered project's folder was renamed, removed or became unreadable, the daemon could fail to start after a restart or an update, which locked you out. It now starts, logs a warning naming the affected project and profile, and leaves that project's saved layout as it was. Other projects load normally.
- Silent agent turns are stopped on schedule. An ACP agent can report a failed tool call and then
never answer the prompt. The quiet watchdog now starts measuring silence right away, so it warns at
session.supervision.quiet_after(30 minutes by default) and stops the turn aftersession.supervision.stop_grace(10 minutes by default). Before, it waited extra watchdog intervals while the turn still looked active. Any new update from the agent clears the warning, and after the stop you can use Resume to retry. - Hosted MCP calls no longer hang forever. A tool call made without its own deadline could wait
indefinitely if the daemon stopped answering. Such a call now has a time budget built from the
longest tool wait (24 hours), the longest approval wait (10 minutes) and
client.request_timeout(30 seconds by default). When it runs out, the call ends with a tool error. Clarifications without a timeout and long extension calls keep running as long as the daemon reports progress. Calls with their own deadline behave as before. - App updates right after a release. For a few minutes after a release was published, every Update or Retry update click could fail with "The app update could not be installed." because the update feed still named the previous version. The desktop app now installs the version the daemon planned, straight from its GitHub release, and still checks its digest against the release checksums.
Forked and continued sessions show the conversation they came from
A forked or continued session used to show only a divider and "Nothing said here yet", so it looked as if its messages were lost. It now shows a collapsed, read-only Earlier conversation row above the divider, with the number of messages it carried. Nothing loads until you expand it. Expanding it shows the newest 50 messages from the source up to the fork or continue point, and Show earlier messages loads older ones. The row is hidden when nothing was carried over.
These messages belong to the source. You can't edit them or fork from them, and they are not part of the session's search, outline, message counts or find results. The expanded panel says what the agent received, that is, how many of the carried messages it gets, or that the agent continues from its own copy of the conversation. If the source was deleted, archived, moved, or rewound or cleared past that point, the panel says the source is no longer available and that the agent's carried context is unaffected.
Nothing is copied into the new session, so forking and continuing cost the same as before.
compozy session history and compozy__session_history still return only the session's own
messages.
Session lists load again after a parent is deleted, and the rail keeps its place
- Workspace session lists after a deleted parent. On v0.7.7, deleting or cleaning up a session
whose children were still around could make every workspace-scoped session read fail with
session_not_found. That brokecompozy session list --workspace,compozy worktree list,compozy worktree createandcompozy session new --workspace, and such children silently dropped out of the sidebar. These reads work again, and those children show in the sidebar. - Children of deleted side chats stay hidden. When you delete a side chat while a child it
started, such as a subagent, keeps running, that child no longer shows up as a top-level session
in the session list, the catalog or
compozy__session_list. You can still read it by ID. - The rail keeps its scroll when you switch sessions. Clicking a row below the fold in a Session window's rail no longer sends the list back to the top. The rail also returns to where you were when its list finishes loading after the switch, until you scroll it yourself. While the picked session loads, the real rail stays beside the loading conversation instead of placeholder rows. A reload still starts at the top.
- Notifications reuse your Session window. Clicking a toast, an OS notification, or a session or notice row in the attention bell no longer opens a new Session tab for every session. It lands in your Session window by the same rule that new sessions use, and a session that already has a window is focused there. The rail scrolls the session into view.
- Folders no longer push other sessions off the first page. A folder with many newer sessions
could take the whole first page of the list and hide the sessions outside folders. The sessions
outside folders and each folder now page on their own, each with its own Previous and Next
controls and its own total. Archived and Later keep their Manual order. Over HTTP and UDS, the
sessions list accepts an optional
folder=<id>orfolder=@rootselector, which needs one project and one profile; leaving it out returns the whole list as before. - Global always-on agents appear once. When the session list or the Kanban board is grouped by project, Global always-on agents now show only in the Global row, not again in the User group.
- Group totals leave out always-on main sessions. A quiet Global always-on agent could make the
User group read one session with no rows, and project totals counted agent main sessions too.
Requests that send
exclude_instance_main=true, as the List and Kanban views do, now leave those sessions out of the counts as well. Requests without the flag count as before.
Current subagent usage, terminal cap recovery and more fixes
- Usage for resumed Claude subagents. When Claude runs a finished background subagent again, its token and tool counts now show the current run instead of the highest figure from any earlier run. This applies to the subagent drawer, the CLI, HTTP and UDS, and the native subagent status tool. Existing subagents keep their recorded usage on upgrade.
- Terminal cap dialog after a refused Close. A Terminal window waiting behind the terminal limit dialog used to stay on "Opening a terminal" when the daemon refused the Close you chose there. The limit dialog now opens again, and dismissing it leaves the window on its empty state.
- Empty peer lists explain themselves. When an always-on agent's peer list is empty because
every peer was hidden by authorization,
GET /api/agent-instances?peers=true, the CLI andcompozy__agent_instance_listnow return apeers_notesaying so, instead of looking like there are no peers. The note names no agents, IDs or projects. - Pull request monitors reject branches with spaces.
compozy__monitor_createwithkind=pull_requestnow refuses a branch selector that contains whitespace anywhere, including leading or trailing spaces, instead of waiting for a pull request that can't exist. - Extension tools with line separator characters. A TypeScript extension tool whose schema contained U+2028 or U+2029 in a property name or string value could fail to build or become unavailable. Those tools now work, and published schema digests and Loop pins are unchanged.
- Local sockets for stdio MCP servers. Stdio MCP servers run with their own isolated home folder, so a server that finds a local service's socket through your home folder can't see it. The MCP documentation now shows how to pass the absolute socket path through the server's own argument or environment variable.