Antigravity CLI

Antigravity CLI

Announcement: Begin with: "I'm using the antigravity-cli skill for cross-AI validation with Antigravity (agy)."

Use this wrapper for an explicit Antigravity second opinion or for research that needs its provider-specific grounding. Read the shared lifecycle in ../codex-cli/references/cross-ai-run.md first.

Provider adapter

1. Check command -v agy, then agy --version. Read agy --help before using print mode, output, timeout, model, or effort flags. If the binary or a required capability is absent, report the prerequisite and stop.

2. For an explicit Gemini run, inspect agy models, choose an exact currently listed gemini-* slug, and pass it as --model. Record the selected provider and model. If no model was requested, omit --model and retain the provider's configured default; an unreported default model is unverified and does not establish model-level independence.

3. For review or research, resolve the loaded skill directory, run from the target repository root, and use the runner with a fresh output directory:

   ANTIGRAVITY_SKILL_DIR="/absolute/path/to/loaded/antigravity-cli"
   python3 "$ANTIGRAVITY_SKILL_DIR/scripts/review.py" \
     --prompt-file /tmp/agy-prompt.md \
     --file path/to/file.py \
     --file path/to/requirements.md \
     --model '<exact slug from agy models>' \
     --effort high \
     --timeout 120 \
     --output-dir /tmp/agy-review-unique
   

   ANTIGRAVITY_SKILL_DIR must be the actual directory containing this loaded skill; do not resolve it from the target repository's current directory. --file is repeatable. --effort is optional and accepts low, medium, or high. The runner verifies and pre-reads those files, then inlines their contents so a denied provider file read cannot look like a successful review. --model is optional and has no runner default; omit the line when the configured provider default is intended. Use a new output directory for every attempt and do not retry implicitly.

4. Read receipt.json, stdout.json, version.stderr.log, response.txt on success, and stderr.log from the output directory. A valid completed review requires process exit 0, payload status: SUCCESS, a non-empty response, and no denied_actions. A SUCCESS result with an empty or visibly partial answer, or with any denied action, is PARTIAL and must never be reported as a passed review. Inspect and retain stderr even on exit 0. Return the shared receipt; the result never authorizes edits, shell work, or publication.

The runner invokes the supported headless shape: agy -p --output-format json --print-timeout <duration>, with an explicit --model <slug> only when chosen by the caller. The old --approval-mode and -o forms are incompatible with the tested Antigravity CLI. Do not use permission-bypass flags, settings/auth edits, or automatic model fallback.

For upstream behavior, use the official headless CLI docs, migration guide, and installation guide. Verify local behavior first because flags, models, and output fields can change.

Prompt requirements

Include the repository root, task/requirements path when task-attached, files or diff scope, and concrete questions. For web research, ask for source URLs and separate sourced claims from repository observations. Resolve task evidence through ../setup/references/task-context.md; standalone calls do not create a task. Do not put unverified @path references in review prompts; pass verified files through the runner's repeatable --file option.

When not to use

• The caller did not explicitly choose Antigravity or delegate to it
• A normal review or research tool already satisfies the request
• The request needs interactive refinement or file-writing work

References

• ../codex-cli/references/cross-ai-run.md — shared lifecycle and receipt
• reference.md — local-help-first flags, models, output, and migration notes
• templates.md — prompt shapes and runner examples