Cloudflare Workers AI changed the API for flux-2-dev, flux-2-klein-4b,
and flux-2-klein-9b to require multipart/form-data (instead of JSON) and
now returns {"image":"<base64>"} instead of raw PNG bytes.
- Add requiresMultipart() helper for the three FLUX.2 models
- callImageAPI builds multipart body for those models, JSON for others
- Parse {"image":"<base64>"} JSON response; fall back to raw bytes for legacy models
- Use "steps" field name (not "num_steps") in multipart forms per CF docs
- Book page: capture and display actual backend error message instead of blank 'Error'
CF Workers AI has a ~4MB JSON body limit. Large cover images base64-encoded
can easily exceed this, causing Cloudflare to return a 502 Bad Gateway.
Added resizeRefImage() which scales the reference down so its longest side
≤ 768px (nearest-neighbour, re-encoded as JPEG) before passing it to the
GenerateImageFromReference call. Images already within the limit pass through
unchanged.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Llama 4 Scout returns `result.response` as an array of objects
[{"generated_text":"..."}] instead of a plain string. Decode into
json.RawMessage and try both shapes; fall back to generated_text[0].
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
The aura-2-en model enforces a hard 2 000-character limit per request.
Chapters routinely exceed this, producing 413 errors.
GenerateAudio now splits the stripped text into ≤1 800-char chunks at
paragraph → sentence → space → hard-cut boundaries, calls the API once
per chunk, and concatenates the MP3 frames. Callers (runner, streaming
handler) are unchanged. StreamAudioMP3/WAV inherit the fix automatically
since they delegate to GenerateAudio.
Adds backend handlers and SvelteKit UI for an admin text generation tool.
The tool lets admins propose and apply AI-generated chapter titles and book
descriptions using Cloudflare Workers AI (12 LLM models, model selector shared
across both tabs).
