Image GenerationSAFE

Open-source super AI assistant & Agent Harness. Plans tasks, runs tools and skills, self-evolves with memory and knowledge. Multi-agent, multi-model, multi-channel. Lightweight, extensible, one-line install.

The instruction file, verbatim from the audited commit — this is the text the model reads, and the surface the audit's instruction layer examines. Quoted here so you can judge it without cloning anything.

---
name: image-generation
description: Generate or edit images from text prompts. Use when the user asks to create, draw, design, or edit an image, illustration, photo, icon, poster, or any visual content.
metadata:
  cowagent:
    requires:
      anyEnv:
        - OPENAI_API_KEY
        - GEMINI_API_KEY
        - ARK_API_KEY
        - DASHSCOPE_API_KEY
        - MINIMAX_API_KEY
        - LINKAI_API_KEY
---

# Image Generation

Generate and edit images using AI models. The script automatically picks a backend based on which API keys are configured — **you don't need to specify a model unless the user explicitly names one**. Never guess or invent a `model`: when the user doesn't name one, omit `model` entirely so the configured default/provider is used.

Supported models (passed via `model` only when the user asks for a specific one):

- **OpenAI** — `gpt-image-2.5-flare` (recommended default: best quality/latency trade-off), `gpt-image-2.5-sunburst` (highest editing precision), `gpt-image-2`, `gpt-image-1`
- **Gemini Nano Banana** — `nano-banana-2`, `nano-banana-pro`, `nano-banana`
- **Seedream (Volcengine Ark)** — `seedream-5.0-lite`, `seedream-4.5`
- **Qwen (DashScope)** — `qwen-image-2.0`, `qwen-image-2.0-pro`
- **MiniMax** — `image-01`

## Usage

Run `scripts/generate.py` with a JSON argument. Use the absolute `<base_dir>` path — **do NOT `cd` into the skill directory**, since it is a builtin skill and gets reset on restart (anything written there is lost).

```bash
python <base_dir>/scripts/generate.py '<json_args>'
```

Images are saved to the workspace (or the open project dir) under `images/`, so they persist and stay reachable to the client.

**Set bash timeout to at least 600 seconds**, as image generation can take 30–200s per provider, and the script may try multiple providers sequentially. Call it synchronously (do not run it in the background) and wait for its output in the same turn.

### Parameters

| Parameter | Type | Required | Default | Description |
|-----------|------|----------|---------|-------------|
| `prompt` | string | yes | — | Image description |
| `image_url` | string / list | no | null | Input image(s) for editing: local file path or URL. Multi-image fusion is supported (pass a list) |
| `quality` | string | no | auto | `low` / `medium` / `high` (only some backends honour this) |
| `size` | string | no | auto | `512` / `1K` / `2K` / `3K` / `4K`, or pixel value (`1024x1024`) |
| `aspect_ratio` | string | no | null | `1:1` / `3:2` / `2:3` / `16:9` / `9:16` / `21:9` (some backends also support extreme ratios like `1:4` / `8:1`) |

**Higher `quality` and larger `size` cost more and run slower.** In normal cases, when the user does not explicitly specify, `low` or `medium` is sufficient. Only use `high` when the user asks for it.

### Example — generate

```bash
python <base_dir>/scripts/generate.py '{"prompt": "A corgi astronaut floating in space"}'
```

With aspect ratio:

```bash
python <base_dir>/scripts/generate.py '{"prompt": "Isometric miniature city of Shanghai at sunset", "size": "2K", "aspect_ratio": "16:9"}'
```

### Important: Editing vs Generating

When the user asks to **edit, modify, or improve an existing image**, pass the original image via `image_url`. Prefer **local file paths** directly — the script handles file reading internally. Without `image_url`, the script generates a brand-new image instead of editing.

### Example — edit (image-to-image)

```bash
python <base_dir>/scripts/generate.py '{"prompt": "Add a Santa hat to the dog", "image_url": "/path/to/dog.png"}'
```

Multi-image fusion — pass a list:

```bash
python <base_dir>/scripts/generate.py '{"prompt": "Combine these characters into a group photo", "image_url": ["/path/a.png", "/path/b.png"]}'
```

### Output

Prints JSON to stdout:

```json
{
  "model": "doubao-seedream-5-0-260128",
  "images": [
    {"url": "/path/to/output.png"}
  ]
}
```

After success, display the image to the user. You can either embed it in markdown (`![description](/path/to/output.png)`) or use the `send` tool.

On error:

```json
{
  "error": "error message"
}
```

### Setup

The script needs **at least one** of these API keys (set via `env_config` or `config.json`):

`OPENAI_API_KEY` / `GEMINI_API_KEY` / `ARK_API_KEY` / `DASHSCOPE_API_KEY` / `MINIMAX_API_KEY` / `LINKAI_API_KEY`

Each also has an optional `*_API_BASE` for custom endpoints. The script automatically picks the first configured backend and falls back to the next if it fails — no need to specify a model.

### Error Handling

If the script returns an error after trying all configured backends, **do NOT retry with the same parameters** — the failure is almost always a configuration issue (wrong API key, unsupported API base). Tell the user to fix it via `env_config`, then retry.

### Notes

- HTTP timeout is 300s — high-resolution generation can take over 200s.
- Omit `quality` / `size` to let the model pick automatically (`auto`).
- Input images for editing are auto-compressed to ≤ 4MB / longest edge ≤ 4096px.

SAFEgrade B · trust 89/100 Nothing in the source contradicts what it says it does. Grade A is reserved for packages that have also passed the behavioural sandbox.

What the source does

Filesystem

none-observed

Network

declared (5 observation(s))

Shell

none-observed

Dependencies

pinned

Secrets in source

none-found

Findings (4)

return base64.b64decode(value)

raw = base64.b64decode(inline["data"])

paths.append(_save_image(base64.b64decode(b64), output_dir))

paths.append(_save_image(base64.b64decode(b64), output_dir))

Gates applied: no_behavioural_pass.