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
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.
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.
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.
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
1---2name: antigravity-cli3description: >-4 Run Google Antigravity CLI (agy) for web-grounded research, cross-AI review,5 or validation when those capabilities are explicitly needed or another6 skill delegates to agy. One-shot only; do not select it for an ordinary7 review without an explicit provider choice.8allowed-tools:9 - Bash10 - Read11metadata:12 source: https://skillmd.com/skills/alexandrbasis/antigravity-cli13---1415# Antigravity CLI1617> **Announcement**: Begin with: "I'm using the **antigravity-cli** skill for cross-AI validation with Antigravity (agy)."1819Use this wrapper for an explicit Antigravity second opinion or for research that20needs its provider-specific grounding. Read the shared lifecycle in21`../codex-cli/references/cross-ai-run.md` first.2223## Provider adapter24251. Check `command -v agy`, then `agy --version`. Read `agy --help` before using26 print mode, output, timeout, model, or effort flags. If the binary or a27 required capability is absent, report the prerequisite and stop.282. For an explicit Gemini run, inspect `agy models`, choose an exact currently29 listed `gemini-*` slug, and pass it as `--model`. Record the selected30 provider and model. If no model was requested, omit `--model` and retain the31 provider's configured default; an unreported default model is unverified and32 does not establish model-level independence.333. For review or research, resolve the loaded skill directory, run from the34 target repository root, and use the runner with a fresh output directory:3536 ```bash37 ANTIGRAVITY_SKILL_DIR="/absolute/path/to/loaded/antigravity-cli"38 python3 "$ANTIGRAVITY_SKILL_DIR/scripts/review.py" \39 --prompt-file /tmp/agy-prompt.md \40 --file path/to/file.py \41 --file path/to/requirements.md \42 --model '<exact slug from agy models>' \43 --effort high \44 --timeout 120 \45 --output-dir /tmp/agy-review-unique46 ```4748 `ANTIGRAVITY_SKILL_DIR` must be the actual directory containing this loaded49 skill; do not resolve it from the target repository's current directory.50 `--file` is repeatable. `--effort` is optional and accepts `low`, `medium`,51 or `high`. The runner verifies and pre-reads those files, then52 inlines their contents so a denied provider file read cannot look like a53 successful review. `--model` is optional and has no runner default; omit the54 line when the configured provider default is intended. Use a new output55 directory for every attempt and do not retry implicitly.564. Read `receipt.json`, `stdout.json`, `version.stderr.log`, `response.txt` on57 success, and `stderr.log` from the output directory. A valid completed58 review requires process exit 0, payload59 `status: SUCCESS`, a non-empty `response`, and no `denied_actions`. A60 `SUCCESS` result with an61 empty or visibly partial answer, or with any denied action, is `PARTIAL` and62 must never be reported as a passed review. Inspect and retain stderr even on63 exit 0. Return the shared receipt; the result never authorizes edits, shell64 work, or publication.6566The runner invokes the supported headless shape: `agy -p --output-format json67--print-timeout <duration>`, with an explicit `--model <slug>` only when chosen68by the caller. The old `--approval-mode` and `-o` forms are incompatible with69the tested Antigravity CLI. Do not use permission-bypass flags, settings/auth70edits, or automatic model fallback.7172For upstream behavior, use the official [headless CLI docs](https://antigravity.google/docs/cli/headless/),73[migration guide](https://antigravity.google/docs/cli/gcli-migration/), and74[installation guide](https://antigravity.google/docs/cli/install/). Verify local75behavior first because flags, models, and output fields can change.7677## Prompt requirements7879Include the repository root, task/requirements path when task-attached, files or80diff scope, and concrete questions. For web research, ask for source URLs and81separate sourced claims from repository observations. Resolve task evidence82through `../setup/references/task-context.md`; standalone calls do not create a83task. Do not put unverified `@path` references in review prompts; pass verified84files through the runner's repeatable `--file` option.8586## When not to use8788- The caller did not explicitly choose Antigravity or delegate to it89- A normal review or research tool already satisfies the request90- The request needs interactive refinement or file-writing work9192## References9394- `../codex-cli/references/cross-ai-run.md` — shared lifecycle and receipt95- `reference.md` — local-help-first flags, models, output, and migration notes96- `templates.md` — prompt shapes and runner examples