HiAPI GPT Image 2.5

This released 0.3.3 skill wraps HiAPI's unified async image API for six GPT Image 2.5 model IDs in two routes:

• Mode route (default). When --model is omitted, the CLI selects gpt-image-2.5-<family>/text-to-image without reference images and gpt-image-2.5-<family>/image-to-image with them. --family is flare by default; use sunburst only when the caller asks for Sunburst. Priced per image by resolution (1K/2K/4K).
• Pro route (--route pro). gpt-image-2.5-flare@pro or gpt-image-2.5-sunburst@pro, used only when the caller asks for pro, quality tiers, or a PNG/JPEG/WebP output choice. Priced per image by quality.

Never send the bare IDs gpt-image-2.5-flare or gpt-image-2.5-sunburst; the CLI rejects them.

Each accepted task produces one image. Do not infer a quality, speed, or visual advantage between Flare and Sunburst or between the routes.

Required sequence

1. Decide whether the request is text-to-image (no --image-url) or image-to-image/editing (1–16 --image-url values). The mode route accepts public HTTP(S) URLs or data:image/png|jpeg|webp;base64, URIs; the pro route accepts public HTTP(S) URLs only. Local file paths are not uploaded.
2. Validate the prompt and options locally. All six model IDs require a non-empty prompt of at most 20,000 UTF-16 code units after trimming surrounding whitespace (JavaScript string.length; astral emoji count as two). The API returns INVALID_REQUEST with a maxLength message above that. Text-to-image additionally has an internal o200k token cap that returns PROMPT_TOO_LONG; image-to-image has no token cap. Run --dry-run --estimate before a paid create. The estimate reads the current /api/pricing snapshot and creates no task. The CLI rejects fields that belong to the other route (--quality/--output-format on the mode route, --resolution on the pro route) and --background with resolution 2K/4K.
3. Submit once with a stable --idempotency-key. The CLI prints the key and task ID to stderr immediately after acceptance. If acceptance is ambiguous, reuse the same key; never blind-retry with a new key.
4. Use the default wait, or --no-wait to return after submission. Use --resume-task-id to poll/download an existing task without creating another task. Use --no-save when you only need the output URL.
5. Keep the downloaded image through technical and creative checks. Treat temporary output URLs as expiring.