laurentvv/agents-kit/agy

agy - headless Antigravity CLI

agy (Google Antigravity) runs an agent non-interactively: one prompt in (-p), the response on stdout, diagnostics on stderr, process exits. It reuses the logged-in account's cached credentials - authenticate once with an interactive agy session before any headless use.

Everything below was verified live on agy 1.2.16 (Windows 11, Git Bash), 2026-10-04. Official reference: https://antigravity.google/docs/cli/headless/

Generate an image (the core recipe)

Image generation is an agent TOOL (generate_image, delegated to an image-generator subagent), not a model slug: agy models lists text models only, any of them can drive the generation. The image bytes are produced server-side; the agent then copies the file into the working directory with a shell command - and that copy is the whole permission problem.

1. Work from a disposable directory (scratch/..., a temp dir): the run

needs --dangerously-skip-permissions (see below), so give it nothing valuable to touch.

1. One image, one command:

mkdir -p scratch/agy-images && cd scratch/agy-images
agy -p "Generate an image: <subject, color/material, background, lighting, framing>. Save it as <name>.png in the current directory." \
    --dangerously-skip-permissions --output-format json

1. Verify THREE things before claiming success - a denied run still exits 0

with status SUCCESS:

• the envelope has .status == "SUCCESS" AND denied_actions absent or

empty,

• the file exists on disk with the right magic bytes (PNG / JPEG),
• the image was actually LOOKED at (open it, screenshot it, vision pass):

verified means seen.

uv run --no-project python -c "import pathlib; b=pathlib.Path('<name>.png').read_bytes()[:8]; print('PNG' if b.startswith(b'\x89PNG\r\n\x1a\n') else b)"

Why --dangerously-skip-permissions

• The final file copy runs through run_command; in headless mode an

unapproved command is SOFT-DENIED: the run "succeeds" (exit 0, status SUCCESS, empty or partial response) and the envelope carries denied_actions - the image never lands in the working directory.

• Telling the agent to "use cp" does NOT reliably hit an existing

permissions.allow rule - verified twice, denied both times.

• write_to_file cannot carry binary image data (observed: junk test

files, then a denial).

So keep the flag and contain the blast radius: disposable cwd, a prompt you wrote, nothing sensitive in reach. Scoped permissions.allow rules are the alternative when a run must work without the flag.

Salvage a denied run

The generation itself usually completed before the denied copy. The raw image waits under the CLI's brain directory, named after the prompt with a timestamp - and the extension can drift (a .png request landed as .jpg):

~/.gemini/antigravity-cli/brain/<conversation_id>/<prompt_name>_<timestamp>.jpg

Take conversation_id from the envelope, copy the newest file out, rename it.

Iterate on an image

Conversations continue headless: -c continues the most recent one, --conversation <id> a specific one. Reference-based edits (recolor, restyle, variation) work, reading the earlier image from the conversation context; expect roughly double the cost and duration of a fresh generation.

agy -c -p "Same subject, <change only>. Save it as <new-name>.png in the current directory." \
    --dangerously-skip-permissions --output-format json

Machine-readable output

• --output-format text (default): the response text only.
• --output-format json: one envelope at the end - conversation_id,

status, response, error, duration_seconds, num_turns, usage (token counts) and denied_actions (undocumented in the official page, present in 1.2.16).

• --output-format stream-json: NDJSON events (init, then step_update

with tool calls, then one result) - the way to see WHICH tools ran.

Do not assume jq is installed - parse with Python:

agy -p "..." --output-format json | uv run --no-project python -c "import json,sys; e=json.load(sys.stdin); print(e['status'], e['response'])"

--json-schema constrains the answer to a schema; --model <slug> pins a model and fails loudly (exit 1) on an unknown slug; --effort low|medium|high trades depth for speed (1.2.16 also accepts xhigh|max).

Expectations and pitfalls (measured 2026-10-04, agy 1.2.16)

• One image: 45-100 s and 45k-80k tokens; one reference edit: ~200 s and up

to 166k cumulative tokens. Budget before batching.

• The local binary can drift from the docs: 1.2.16 waits indefinitely by

default (--print-timeout 0s; the docs describe a 5 m ceiling) - set the flag explicitly in scripts, and re-read agy --help on the target machine before trusting memorized flags.

• The enclosing repository leaks into the run: agy reads the workspace

AGENTS.md and .agents/skills/ of the cwd's repo (observed: replies in the repo's chat language, workspace skills auto-loaded). Run from a neutral directory unless that context is wanted.

• CI or any terminal-less environment without cached credentials exits with

authentication required instead of hanging.

• --disable-slash-commands turns off slash-command and skill expansion in

print mode.

Safety

--dangerously-skip-permissions approves every tool call of the run, file writes and shell commands included. Disposable directory, self-written prompt, nothing sensitive within reach - and never the user's working tree.