canghe-image-gen
AI image generation with OpenAI, Google, DashScope and Canghe APIs. Supports text-to-image, reference images, aspect ratios. Sequential by default; parallel generation available on request. Use when user asks to generate, create, or draw images.
1How It Works & Mechanism
The execution entity of this skill is a Bun/TypeScript CLI (scripts/main.ts), launched by the agent using npx -y bun according to SKILL.md; argument parsing, configuration merging, provider selection, and result saving are all completed at this entry point.
npx -y bun ${SKILL_DIR}/scripts/main.ts --prompt "A cat" --image cat.png
Automatic provider selection: Explicit --provider takes priority; when --ref is provided, the order is Google -> OpenAI -> Canghe; without --ref, it picks the first available key found in the environment (Google takes precedence if multiple keys exist).
const hasGoogle = !!(process.env.GOOGLE_API_KEY || process.env.GEMINI_API_KEY);
The four backends are modularized individually and unify under the ProviderModule interface (getDefaultModel / generateImage), each with different protocols: OpenAI uses images/generations and edits, Google uses Gemini generateContent or Imagen predict, DashScope uses the Tongyi Wanxiang endpoint, and Canghe uses OpenAI-compatible chat/completions.
const res = await fetch(`${getBaseUrl()}/chat/completions`, {
Reference image support varies by backend: Google multimodal reads reference images as base64 into inlineData, OpenAI uses the /images/edits form, and Canghe transmits via image_url (local files are converted to data URIs, while http URLs are passed directly); DashScope and Imagen explicitly throw errors as unsupported.
const { data, mimeType } = await readImageAsBase64(refPath);
The configuration hierarchy priority is fixed as CLI args > EXTEND.md > environment variables > cwd/.canghe-skills/.env > ~/.canghe-skills/.env; main.ts parses the YAML front matter of EXTEND.md and .env key-value pairs on its own without relying on dotenv.
**Load Priority**: CLI args > EXTEND.md > env vars > `<cwd>/.canghe-skills/.env` > `~/.canghe-skills/.env`
Generation results are written in binary to the user-specified --image path (automatically appending .png if no extension is present), with directories created recursively prior to writing; failed generations are automatically retried only once, and errors such as 'reference image unsupported/missing key' are not retried.
await writeFile(outputPath, imageData);
Batch and parallel operations are two distinct modes: images are generated sequentially one by one by default; parallel generation is only enabled when explicitly requested by the user, running via the host's multiple background sub-agents (recommended 4, max 8); the script itself contains no parallel logic.
**Parallel Generation**: Only use when user explicitly requests parallel/concurrent generation.
2Core Capabilities
3External Dependencies
| Kind | Dependency / Endpoint |
|---|---|
| cli | bun（经 npx -y bun 运行 TS 入口，会联网获取包） |
| api | OpenAI Images API |
| api | Google Generative Language API（Gemini / Imagen） |
| api | DashScope（阿里通义万象） |
| api | Canghe 第三方网关 |
| package | Node 内建模块 node:fs/promises、node:path、node:os（经 Bun 运行） |
4Recommended Governance & Evidence
- scripts/: Local script or command execution: npx -y bun ${SKILL_DIR}/scripts/main.ts (the sole CLI entry point executed by the agent according to SKILL.md)
- 网络端点: Declared external network access: POST https://api.openai.com/v1/images/generations (openai.ts)
- 本地工作目录: File read/write and artifact output: Image files at user-specified --image path (auto-appends .png if missing; main.ts runs mkdir -p before writeFile)
5Second-Pass Verification
- [ok] 执行入口是否唯一且真实存在 — SKILL.md 的调用示例与 main.ts 的 printUsage 一致（npx -y bun scripts/main.ts --prompt ... --image ...），main.ts 末尾 main().catch 会 process.exit(1)，证明它是 CLI；目录仅 8 个档（SKILL.md + preferences-schema.md + main.ts + 4 个 provider + types.ts），无其它可执行文件。
- [ok] 四家 provider 端点与凭证读取点 — openai.ts→api.openai.com/v1 + OPENAI_API_KEY；google.ts→generativelanguage.googleapis.com + GOOGLE_API_KEY/GEMINI_API_KEY；dashscope.ts→dashscope.aliyuncs.com + DASHSCOPE_API_KEY；canghe.ts→api.canghe.ai/v1 + CANGHE_API_KEY。均可用对应 *_BASE_URL 覆盖。
- [ok] 是否漏掉网络外发 — 全目录除 providers/*.ts 的 fetch 外无网络能力：main.ts 只做 IO 与编排，types.ts / preferences-schema.md 无网络代码；未见 socket/child_process/exec/urllib。
- [ok] TLS 是否被降级 — 四家后端都用 Node/Bun 原生 fetch（默认校验证书），未见 ssl.CERT_NONE / rejectUnauthorized=false / --no-verify，因此不触发红档。
- [discrepancy] 「并行生成」宣称是否有代码支撑 — SKILL.md 的 Parallel Generation 段描述用宿主 Task 子代理 run_in_background=true 并发并给并发建议，但 scripts/main.ts 只有单次 --n 的串行生成，未实现任何并发；该能力完全依赖 agent 编排层，脚本不提供。
- [ok] 默认 provider 与质量口径 — SKILL.md 写 default provider google、quality 2k；detectProvider 多 key 时返回 available[0]（google 排第一），main.ts 对空 quality 置 '2k'，与文档一致。
- [ok] 参考图 provider 限制 — SKILL.md 称 --ref 受 Google 多模态 / OpenAI edits / Canghe 支持；google.ts 对 Imagen、openai.ts 对 dall-e、main.ts 对 dashscope+ref 均抛错，与文档相符。
6Verdict & Scenarios
6d47e36b6c02f891…