Collaborating With Antigravity

Collaborating with Antigravity (agy)

Use the Antigravity CLI (agy) as an independent collaborator while the calling agent stays responsible for verification, synthesis, and final user-facing decisions. Antigravity CLI is Google's replacement for the retired Gemini CLI (hosted auth shut off 2026-06-18).

The bridge script (scripts/agy_bridge.py) wraps agy -p (headless print mode), returns structured JSON, and manages session continuity via SESSION_ID. Always go through the bridge — raw agy -p surfaces no conversation ID and no structured output (see below).

Why the bridge is mandatory

• agy won't reveal the conversation ID (no --session-id, never printed). The bridge recovers it from agy's cache, returns it as SESSION_ID, and resumes via --conversation. It also serializes calls with a file lock — don't run bridge calls concurrently.
• agy has no structured output. The bridge shapes plain text into JSON, streams live progress to stderr, and enforces its own wall-clock kill as the real timeout.

Mechanics, upstream issue numbers, and the verified flag surface: references/agy-cli.md.

Commands below write <skill_dir> for the absolute path of the directory containing this SKILL.md. Your harness usually reports that path when it loads the skill. If it does not, use this SKILL.md's own location. Substitute it before running, for example ~/.claude/skills/collaborating-with-antigravity.

Safety

agy can read and write files and run tools in the workspace. To constrain it:

• --sandbox is on by default in the bridge — it applies agy's terminal restrictions. Recommended for review/consultation. Use --no-sandbox only when you deliberately want edits/shell.
• Include OUTPUT: Unified Diff Patch ONLY. Do not modify any files. in the prompt when requesting code changes. This is a convention, not enforced.
• --skip-permissions maps to agy's --dangerously-skip-permissions (auto-approves every tool). Only use it with explicit user consent, ideally in an isolated worktree.
• Never hand agy secrets, private keys, or production data.

Host-side approval (the bridge call itself)

Everything above governs the child agy. The host agent's own permission layer gates the python3 … agy_bridge.py Bash call first — and under classifier-gated auto-approval (Claude Code auto/dontAsk, Codex non-interactive runs), a long-running script that spawns another agent over the codebase pattern-matches "high-risk" and can be denied silently: the delegation never starts. A host permission error instead of bridge JSON means the host blocked the bridge, not that agy failed.

• Pre-authorize the bridge instead of relying on the classifier. Claude Code host: add "Bash(python3 *collaborating-with-antigravity*bridge.py*)" to permissions.allow in your settings.json. The wildcard form keeps matching wherever the skill is installed. Sandboxed hosts also need "python3 *collaborating-with-antigravity*bridge.py*" in sandbox.excludedCommands, because sandbox network policy blocks the child CLI's API traffic even after the command is allowed. Install and approval runbooks: docs/setup/ in the source repo.
• Codex host: the sandbox is the second gate. The child agy CLI needs network for Google auth and inference, which the host sandbox blocks in read-only and workspace-write. Run the bridge call through an approved escalation, or knowingly grant network for that call.
• Never degrade silently. If the host denies the bridge call, report it and propose the allowlist fix — don't substitute your own answer for the independent second opinion that was requested.

When to use

• Second opinion on design tradeoffs, edge cases, or test gaps.
• Web search / research (Antigravity has built-in grounding).
• Reviewing a diff or proposing changes as a unified diff.
• Screenshot / image analysis.

When not to use

• Trivial one-shot tasks — do them directly.
• Tasks requiring file edits — the calling agent edits, agy advises.
• Anything involving secrets, private keys, or prod data.

Quick start

⚠️ Backticks / $VARS in prompts trigger shell expansion — use a single-quoted heredoc, or --prompt-file for large/generated prompts. See references/shell-quoting.md.

bashCopy code
PROMPT="$(cat <<'EOF'
Review src/auth.py around login() and propose fixes.
OUTPUT: Unified Diff Patch ONLY.
EOF
)"
python3 <skill_dir>/scripts/agy_bridge.py \
  --cd "." --model "Gemini 3.5 Flash (Low)" --PROMPT "$PROMPT"

Returns (stdout JSON): { "success": true, "SESSION_ID": "...", "agent_messages": "...", "model": "...", "warnings": [...] }. Live progress streams to stderr; the bridge exits non-zero on failure. Run non-trivial calls in the host's background-command mode and watch the stderr progress.

Multi-turn sessions

Capture SESSION_ID from the first call and pass it back (use the same --cd):

bashCopy code
# Turn 1
python3 <skill_dir>/scripts/agy_bridge.py \
  --cd "." --PROMPT "Analyze the bug in foo()."

# Turn 2 — resume by ID
python3 <skill_dir>/scripts/agy_bridge.py \
  --cd "." --SESSION_ID "<id>" --PROMPT "Propose a fix as a unified diff."

# Or continue the most recent conversation
python3 <skill_dir>/scripts/agy_bridge.py \
  --cd "." --continue --PROMPT "What about edge cases?"

--SESSION_ID and --continue are mutually exclusive; --continue resumes agy's most-recent conversation globally (not per---cd), so prefer --SESSION_ID for deterministic resume. Sessions persist as SQLite under ~/.gemini/antigravity-cli/conversations/<id>.db. Save turn 1's SESSION_ID — agy can't pre-assign one and never reprints it, so an unsaved conversation can't be resumed.

Bridge flags

Set the host's timeout_ms to 600000 (10 min) when invoking via a command runner.

Models

Probe the live list — don't hardcode it:

bashCopy code
python3 <skill_dir>/scripts/agy_bridge.py --list-models
# -> { "success": true, "models": ["Gemini 3.5 Flash (Low)", "Claude Sonnet 4.6 (Thinking)", ...] }

Pass the exact string to --model. agy itself does not validate --model — a misspelled or unknown name silently runs the default — so the bridge validates against agy models and errors on an unknown name (--no-validate-model to skip). Rough guide: Flash (Low) for quick checks, Gemini Pro or Claude for hard tasks. Full snapshot in references/agy-cli.md.

Prompting

Use assets/prompt-template.md for structured starters. Key principles:

• Point, don't paste — give file paths and line numbers, not code blocks.
• One objective per prompt — competing goals produce noisy output.
• Enforce output format — append OUTPUT: Unified Diff Patch ONLY. for code changes.
• Leverage grounding — agy can search the web; ask for research when needed.

Screenshots

agy reads files inside the workspace. Copy screenshots in first, then reference the path (example for a Codex host — adapt the clipboard source path to your host):

bashCopy code
mkdir -p .codex_uploads && cp "${TMPDIR:-/tmp}"/codex-clipboard-*.png .codex_uploads/

Don't add the upload dir to .gitignore (agy, like Gemini, may refuse to read ignored paths). Delete screenshots when done.

Verification

• Smoke: python3 <skill_dir>/scripts/agy_bridge.py --help
• Syntax: python3 -m py_compile <skill_dir>/scripts/agy_bridge.py
• Auth: run agy once and sign in with Google (token at ~/.gemini/antigravity-cli/).
• Session test: run one prompt; confirm JSON has success: true and a SESSION_ID, then resume with --SESSION_ID and confirm continuity.

Collaboration State Capsule

Keep this block updated while collaborating:

Copy code
[Antigravity Capsule] Goal: | SID: | Model: | Sandbox: | Files: | Last: | Next:

References

• references/agy-cli.md — verified agy flags, models, file layout, quirks
• assets/prompt-template.md — structured prompt patterns
• references/shell-quoting.md — heredoc quoting for backticks