3.0

⚠️ Breaking changes

• Route changed — /single/multipart-form and /multiple/multipart-form are gone. Use POST /classify instead.
• Input format changed — Multipart form data replaced with raw application/octet-stream (send image bytes directly).
• Multiple images per request no longer supported — The batch /multiple/multipart-form endpoint has been removed. Classify each image individually with separate API calls.

What's new

• Simplified API — Switched from /single/multipart-form and /multiple/multipart-form (multipart) to a single POST on /classify endpoint accepting raw application/octet-stream. No more form parsing overhead.
• Structured logging — Added @logtape/logtape with ANSI color output. Server logs classification duration and request size at info/debug levels.
• AGENTS.md — Agent instructions for AI coding assistants (Bun/Hono/Biome stack, Docker commands, conventions).

What's better

• Faster inference — Image processing pipeline now resizes directly to 224×224 raw RGB tensor without an intermediate JPEG re-encode. Sharp cache disabled for lower memory usage.
• Memory safety — Tensor disposal is guaranteed via try/finally, preventing GPU/CPU memory leaks even when model.classify throws.
• Raw binary transport — Accepts application/octet-stream instead of multipart form data, avoiding multipart parsing and base64 overhead.
• Dependencies bumped — Hono 4.11.9 → 4.12.22, NSFWJS 4.2.1 → 4.3.0, TypeScript 5 → 6, Biome 2.4.2 → 2.4.15, @types/bun pinned.
• Updated tsconfig — Added "types": ["bun"] for proper Bun type resolution.
• README rewrite — New table of contents, performance deep-dive (SIMD, jemalloc, pipeline details), usage examples in JS/Python/httpie/curl, local dev instructions, license section. Badges added.
• editorconfig / .gitignore — Added .zed to gitignore.