[jellyjelly814]

Summary

Adds MiniMax-H3-Image to the CLI and SDK: text-to-image (POST /v1/images/generations) and image editing (POST /v1/images/edits). The default model stays image-01; H3 is opt-in with --model MiniMax-H3-Image.

# text-to-image
mmx image "An orange cat on a windowsill" --model MiniMax-H3-Image --aspect-ratio 16:9 --area 2K --n 2
mmx image "Poster" --model MiniMax-H3-Image --size 2720x1536 --output-format png --no-prompt-enhance

# image editing: --image switches to /v1/images/edits (1-16 images; local file, URL, data URI or mm_file://)
mmx image generate --model MiniMax-H3-Image --prompt "Replace the cat in Picture 1 with the dog in Picture 2" --image cat.jpg --image https://example.com/dog.png

await sdk.image.generate({ model: 'MiniMax-H3-Image', prompt: 'cat', aspect_ratio: '16:9', area: '2K' });
await sdk.image.generate({ model: 'MiniMax-H3-Image', prompt: 'swap', images: [{ image_url: 'https://...' }] });

Design notes

• Editing is image generate --image rather than a new image edit subcommand, so the mmx image "<prompt>" shortcut (single-subcommand forwarding) keeps working. This mirrors --image on H3 video.
• Validation and request building live in src/image/h3.ts (pure functions shared by CLI and SDK, same pattern as src/video/v2.ts).
• Requests are checked locally against the documented limits: size WxH area 655,360–20,000,000 px and ratio 1:3–3:1; size WxH cannot be combined with aspect_ratio/area; aspect_ratio X:Y with integers 1–300; n 1–10; 1–16 reference images; 64 MiB file/body; b64_json total area (20 MP png / 80 MP jpeg·webp, size auto counted by area); --no-prompt-enhance is text-to-image only and needs a concrete size or ratio.
• The endpoints are synchronous (server waits up to 600 s) and not idempotent, so the request timeout is at least 600 s and nothing is retried.
• Partial success (HTTP 200 with some data[].error): successful images are saved, files keep their original index (image_001, image_003, ...), failures go to stderr (also with --quiet) and to failures / failed_count in JSON output.
• All-failed responses (422/402/500): the trailing (NNNN) code in error.message is parsed for /v1/images/* only (1026/1027 → content filter, 2068 → plan message), and per-image reasons from data[] are added to the hint.
• H3-only flags with image-01, and --seed / --prompt-optimizer / --subject-ref with H3, fail with a usage error. --width/--height are shorthand for --size on H3. --response-format base64 is accepted as an alias of b64_json.
• Out of scope: streaming output.

API docs: CN, Global

Docs

README.md, README_CN.md, SDK.md, skill/SKILL.md.

Tests

• bun run typecheck, bun run lint (0 errors; existing warning in test/sdk/speech.test.ts), bun run build.
• bun test: 674 pass, 1 fail. The failure (Zed configuration > ... XDG_CONFIG_HOME) also fails on main locally and is unrelated.
• New: test/image/h3.test.ts, test/errors/api-images.test.ts; extended test/commands/image/generate.test.ts (mock server: partial failure, edits routing) and test/sdk/image.test.ts.
• --dry-run checked for text-to-image, editing, --no-prompt-enhance, b64_json limits and the legacy image-01 path.
• Not yet run against the live API.

---

^{Open in [code]smith · Need help on this PR? Tag @codesmith-bot with what you need.}

[SaladDay]

LGTM, THANKS!!