| name | antigravity_image |
| description | Generate raster images with Antigravity (the official `agy` CLI, Google's Nano Banana 2 model) on Nick's Google AI Pro quota. USE ONLY WHEN NICK EXPLICITLY ASKS for Antigravity / agy / Gemini / Nano Banana image generation, or invokes /skill antigravity_image. For ordinary image requests use the built-in imagegen skill instead — this is an explicit, opt-in alternative that routes generation through agy rather than the built-in OpenAI image tool. Do not pick it implicitly. |
| metadata | {"category":"media-generation","distribution_scope":"internal","invocation_strategy":"explicit","runtimes":["codex","openclaw"],"source_of_truth":"github:nicshik/gemini-claw"} |
Antigravity Image (agy / Nano Banana 2)
Generate one or more images through the official Antigravity CLI (agy), which drives
Google's Nano Banana 2 image model on Nick's Google AI Pro quota. This is a
deliberate, explicit alternative to the built-in imagegen skill (which uses the OpenAI
image tool). It exists so Nick can say "сделай N вариантов через antigravity_image" and
get images produced by agy, not by OpenAI.
When to use (explicit only)
Use this skill ONLY when Nick explicitly asks for it, e.g.:
- "через antigravity_image ...", "через agy", "через Gemini / Nano Banana";
/skill antigravity_image <описание>;- he directly asks for image generation specifically via Antigravity / agy.
For any other image request, prefer the default imagegen skill. Do NOT silently pick
this skill over imagegen.
How to run
~/.openclaw/workspace/bin/antigravity-image [-n COUNT] [--aspect RATIO] [--model NAME] "<image description>"
-n, --count COUNT — number of variants, 1..10. Each variant is a separate agy run.--aspect RATIO — one of 1:1 2:3 3:2 3:4 4:3 9:16 16:9. Default: the pult's current
aspect (~/.gemini/antigravity-skill.json -> image_aspect).--model NAME — agy reasoning model (default: the pult's default_model). The image
itself is always Nano Banana 2 (fixed by agy); this only picks the reasoning model.--reference PATH — reference image for image-to-image / edits: the result is based
on that file (object shape, colours, label text carry over). Repeatable. Use it when
Nick attaches a photo and asks to edit it or build around it (поменять фон, добавить
элемент, обложка для маркетплейса с этим товаром). The path must be a real local file
from the current task (an inbound attachment or a task-dir file) — never invent one.--dry-run — verify agy/env without spending quota.
Example — Nick's "5 variants" request:
~/.openclaw/workspace/bin/antigravity-image -n 5 --aspect 16:9 "фотореализм, зарянка и пума идут гулять в парк, естественный дневной свет, без текста и водяных знаков"
Example — edit an attached product photo:
~/.openclaw/workspace/bin/antigravity-image --aspect 3:4 --reference /path/to/inbound-photo.jpg "обложка карточки маркетплейса: этот же товар, чистый белый фон, мягкая тень"
Delivering the result to the user
The wrapper prints one line per produced image:
IMAGE: /var/lib/openclaw/.openclaw/workspace/outputs/antigravity-skill-images/<file>.jpg
For EACH IMAGE: line, send that file to the user as a photo/attachment. The files live
under the workspace outputs/ root (an allowed outbound-media root), so they can be
attached directly. Also tell the user how many of the requested variants succeeded.
A result is ONLY a fresh IMAGE: line from the current run. Send exactly the files
this invocation printed as IMAGE: lines, and nothing else:
- If the run printed no
IMAGE: line (or exited non-zero), the generation did not
happen. Say so plainly ("не выполнено, нового изображения нет"), give the reason the
wrapper reported (quota 429 / overload 503 / the model never called the generator), and
attach nothing. Do not soften a failure into a success.
- Never re-send an image from an earlier request, from chat history, or from an inbound
attachment as if it were a new result or a reference — the wrapper already guarantees
every
IMAGE: path did not exist before this run, so if there is no such line there is
no new image to send.
- Match the count and aspect the user asked for this time. Don't carry over a previous
request's parameters (e.g. answering "5 variants" when the user now asked for 2).
- Before delivering, confirm the file actually came from this run — the wrapper's fresh-path
attribution does this for you, so trust the
IMAGE: lines and do not substitute anything.
Timing & reliability
- ~60-90 seconds per image;
-n 5 runs serially ≈ 5-8 minutes. Tell the user it's
generating; don't silently block.
- agy's reasoning model is occasionally flaky and returns no image (~1/3 of runs); the
wrapper auto-retries each variant up to 3 times. If a variant still fails, the wrapper
reports fewer images than requested — offer to re-run the missing ones.
- If Google is overloaded (503) or the image quota is exhausted (429), the wrapper
produces no image and exits non-zero; relay that it's a Google-side condition, not a
bug. Image quota is separate from text models.
Quota & limits — do not invent numbers
If the user asks how many images are left, how many were used, or what the exact daily
limit is, do NOT fabricate a figure or present a mechanism as fact. State only what is
actually known:
- Image quota is separate from and metered independently of the text models.
- Google does not publish the exact bucket size or reset window for this route, and there
is no live counter / "fuel gauge" to read.
- The only authoritative signal is agy's own
429 RESOURCE_EXHAUSTED, which sometimes
carries a reset window ("resets in ~N"). Relay exactly that and nothing more.
Do NOT claim a specific number (e.g. "100/day"), and do NOT explain it as a "developer API
vs web app" split or backend-specific throttling — auth is Nick's Google AI Pro
subscription over OAuth, not a separate API key. If asked how many were generated today,
there is no quota gauge to answer that; past attempts live only in the local brain
transcripts, which is not a limit readout. See the repo README "Image quota" section for
the sourced details.
Notes
- No secrets: agy authenticates via Google OAuth (run the host's
scripts/login.sh once).
- This skill does NOT touch the
/antigravity control-panel plugin; it's an additive,
agent-facing route to the same agy binary.
