# Antigravity Python Prompt

> Send prompts directly to Google Antigravity (Gemini + Claude) via Python using OAuth refresh tokens, without OpenCode; includes correct headers, model ID normalization, endpoint selection, SSE streaming, and 429 backoff.

- Skill: `tankygranny05/antigravity-python-prompt` (Agent Skill, multi-file: 3 files)
- Install (CLI): `npx skillmds@latest add tankygranny05/antigravity-python-prompt`
- Raw SKILL.md: https://api.skillmd.com/api/skills/tankygranny05/antigravity-python-prompt/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: tankygranny05 (https://skillmd.com/u/tankygranny05)
- Updated: 2026-09-10
- Page: https://skillmd.com/skills/tankygranny05/antigravity-python-prompt

---


# Antigravity Python Prompt
*[Created by Codex: 019ba92d-c995-77c0-b090-e9109492f53b]*

## Overview

Send prompts directly to Antigravity using Python (no OpenCode), with correct request envelope, headers, model naming, and streaming SSE parsing.

## Quick Start

- Run the bundled script: `scripts/antigravity_prompt.py`.
- Prefer `--quota gemini-cli` for Gemini if Antigravity is 429, but **Claude only works on Antigravity**.

Copy/paste:

```bash
python ~/.codex/skills/antigravity-python-prompt/scripts/antigravity_prompt.py \
  "Write a short story about programmers (~500 words)" \
  --model gemini-3-pro --thinking high --quota gemini-cli
```

## Workflow

### 1) Load credentials

- Read refresh tokens from `~/.config/opencode/antigravity-accounts.json` (default), or pass `--accounts-file`.
- Select account with `--account-index` (default `0`).

### 2) Normalize model IDs

- Strip prefixes (`google/…`, `antigravity-…`).
- Map common versioned Claude names (e.g. `claude-opus-4-20250514`) to Antigravity model IDs.

### 3) Choose quota bucket + endpoints

- For Gemini:
  - `--quota antigravity` uses Antigravity header style and sandbox endpoints (may 429).
  - `--quota gemini-cli` uses Gemini CLI header style and production endpoint (often a separate quota pool).
  - `--quota auto` tries Antigravity first, then Gemini CLI.
- For Claude/GPT: only `--quota antigravity` is valid.

### 4) Send request

- Build Antigravity request envelope:
  - Top-level: `{ project, model, request, requestType, userAgent, requestId }`
  - `request.contents` uses Gemini-style `{role: "user"|"model", parts:[{text:"…"}]}`.
- Inject `request.systemInstruction` with Antigravity system instruction when using Antigravity header style (this matches OpenCode plugin behavior and avoids flaky Claude throttling).

### 5) Stream SSE

- Parse `data: {json}` lines and print text deltas from `candidates[].content.parts[].text`.

### 6) Handle throttling

- On HTTP `429`, apply exponential backoff (script supports `--retry-on-429` and `--max-backoff-seconds`).
- No reliable “remaining quota” header exists; treat 429 as the signal.

## Examples

### Gemini 3 Pro (streaming, use Gemini CLI quota)

```bash
python ~/.codex/skills/antigravity-python-prompt/scripts/antigravity_prompt.py \
  "Summarize this repo's purpose in 3 bullets." \
  --model gemini-3-pro --thinking high --quota gemini-cli
```

### Claude Opus (streaming, Antigravity only)

```bash
python ~/.codex/skills/antigravity-python-prompt/scripts/antigravity_prompt.py \
  "Explain the difference between rate limit and capacity." \
  --model claude-opus-4-20250514 --retry-on-429 12
```

## Notes

- Avoid asking the model to run tools: this script is a raw LLM call, not an agent runtime.
- Prefer `--quota gemini-cli` for Gemini workloads when Antigravity sandbox returns 429.