Skip to content

antigravity-cli.md

Latest commit

 

History

History
97 lines (63 loc) · 6.58 KB

antigravity-cli.md

File metadata and controls

97 lines (63 loc) · 6.58 KB
Raw

Antigravity CLI Provider

Overview

The Antigravity CLI provider enables CAO to work with Antigravity CLI (agy), Google's terminal-native AI coding agent and the successor to the Gemini CLI after Google retired the free Gemini Code Assist "Login with Google" path for the gemini binary (2026-06-18). agy runs as an interactive full-screen TUI that keeps scrollback history in tmux.

Prerequisites

  • Antigravity CLI: Install via curl -fsSL https://antigravity.google/cli/install.sh | bash
  • Authentication: Run agy once and complete the Google sign-in + one-time onboarding (color scheme + Terms/Data-Use). The provider expects a fully onboarded CLI so no setup screens appear inside CAO-spawned sessions.
  • tmux 3.2+

Verify installation:

agy --version
agy models   # confirms auth + lists model strings

Backend Support

The provider currently requires the tmux backend (cao-server --terminal tmux, the default). It opts into screen-based status detection (supports_screen_detection = True), which is driven by the FIFO / pyte pipeline that tmux provides.

The herdr backend is not yet supported: herdr uses an event inbox and skips the FIFO pipeline for providers it has no native status integration for, so agy's state is never observed and terminals time out. Generic herdr FIFO support for non-native providers is tracked as a follow-up; until then, run antigravity_cli on the tmux backend.

Quick Start

# Launch with CAO
cao launch --agents reviewer_gemini --provider antigravity_cli

Set a model in the agent profile (model: field). Model names are the human-readable strings agy models prints, e.g. "Gemini 3.1 Pro (High)", "Gemini 3.5 Flash (High)", "Claude Sonnet 4.6 (Thinking)".

Launch Command

agy --dangerously-skip-permissions [--model "<model>"] [-i "<system prompt>"]
  • --dangerously-skip-permissions auto-approves tool calls so orchestrated (handoff / assign) flows do not block on per-tool approval prompts.
  • --model selects the model (profile model: wins over a constructor override).
  • -i (--prompt-interactive) injects the agent profile's system prompt (+ skill catalog, + security prompt when tool-restricted) as the first message, with an explicit "acknowledge your role and wait" guard so the agent adopts its role without exploring on launch.

Status Detection

The provider classifies agy states from the tmux output buffer (footer-anchored, render-stable):

Status Pattern Description
PROCESSING Footer esc to cancel (or a braille spinner line, e.g. ⣽ Working...) Response streaming or tool executing
IDLE Footer ? for shortcuts, no turn delivered yet Ready for first input
COMPLETED Footer ? for shortcuts, ≥1 turn delivered Turn finished
WAITING_USER_ANSWER Approval / picker prompt ([y/n], ↑/↓ navigate, …) Blocked on user input
ERROR Error:, panic:, Traceback patterns Error detected

The TUI footer is identical for IDLE and COMPLETED, so the two are split on an internal turn counter (mark_input_received), exactly as the Cursor CLI provider does. This preserves the "wait for IDLE before delivering the task" contract right after initialization.

TUI Structure

─────────────────────────────   (full-width U+2500 rule)
> <user question>
  <assistant response>            (2-space indented)
─────────────────────────────   (input box top rule)
>                                 (idle input prompt)
─────────────────────────────   (input box bottom rule)
? for shortcuts            <model>

Response extraction returns the text between the last echoed > <query> line and the next full-width separator, with TUI chrome filtered out: banner, separators, footer, tips, spinner, tool-call lines (● …), and the collapsed-thought header (▸ Thought for …) together with the single auto-generated title line that follows it.

MCP Servers

agy reads MCP servers from ~/.gemini/config/mcp_config.json (top-level mcpServers key). There is no per-invocation MCP config flag, a custom agent's mcpServers is not a working alternative, and a separate HOME per worker would need its own sign-in, so every CAO worker shares that one file — and every agy process starts every server listed in it.

The provider therefore writes entries that carry no terminal identity:

  • cao-mcp-server gets CAO_TERMINAL_ID from the agy process itself. The backend sets it in the terminal's pane environment, and agy passes its environment to its stdio MCP servers (an entry's env is merged on top). Each worker's agy runs its own copy of the server, and each copy reports its own terminal. A CAO_TERMINAL_ID in a profile's server env is dropped with a warning.
  • Workers whose profiles declare the same server config share one entry, keyed <server>-<8-hex digest of its config>. A different config for the same server name gets its own key. The key stays the same across terminals and restarts, so the tool names agy shows the model do too.
  • Each entry lists the terminals using it under caoTerminalIds. cleanup() removes its own terminal, and the entry is removed with its last user. Registration also releases terminals that are no longer live. Fields you add to an entry (disabled, disabledTools, ...) are kept, and your own entries are never touched.

Before #940 the provider wrote one entry per terminal (<server>-<terminal_id>) with CAO_TERMINAL_ID pinned in its env. A worker launched while another was alive also started that worker's entry, so its send_message, handoff and assign calls could reach the supervisor under the other terminal's id. Registration removes entries in that old shape. An agy that is already running keeps the servers it started, because agy reads the file when it launches.

Agent-plugin MCP working directory

Its MCP config format documents a working-directory key, so CAO writes an agent plugin's declared cwd there directly. See Agent Plugins for the full per-provider table.

Tool Restrictions

agy is in SOFT_ENFORCEMENT_PROVIDERS: tool restrictions are advisory. When a profile is not allowed every tool (e.g. the read-only reviewer), the security prompt is appended to the injected system prompt. There is no native hard-block flag.

Exit

agy exits on the /quit slash command (also Ctrl-D pressed twice).