[rekram1-node]

Problem

Image edits through OpenRouter never worked. OpenRouter image models use the OpenAI images protocol, and when a request has input images that protocol posts multipart to /images/edits. OpenRouter has no such endpoint, so every edit came back as HTTP 404 for every model.

Just changing the path isn't enough either. OpenRouter's image API (POST /api/v1/images, alias /images/generations) is JSON only and reads input images from input_references: [{ type: "image_url", image_url: { url } }]. If you send OpenAI's images field instead, it silently ignores it and you get an unrelated text-to-image redraw.

Change

I added an imageInput option to OpenAIImages.model, next to the sizing hook from #59927. It defaults to "edits", so OpenAI, Azure, Vercel and the other hosts behave exactly as before. OpenRouter passes "references":

 OpenRouter image request
-  images? -> multipart POST /images/edits        (404)
+  always  -> JSON POST /images/generations
+  images  -> input_references: [{ type: "image_url", image_url: { url } }]
+             bytes as data URLs, http(s) URLs as is
+  mask    -> UnsupportedOperation (OpenRouter has no mask parameter)

A request without images sends the same body as before. Response parsing didn't need changes, since OpenRouter returns OpenAI-style data[].b64_json.

This is stacked on #59927 (image-sizing-fixes) because both touch the same lines in openai-images.ts and openrouter.ts. Once that merges this retargets to dev with only its own commit.

Out of scope

• Image.stream through OpenRouter. Its stream events carry media_type but no output_format, and the protocol currently requires output_format on partial and completed events, so streaming there fails to decode. That's a separate fix.
• Per-model reference limits. GET /api/v1/images/models lists input_references {min, max} per model (some take 0). I'm leaving the count check to OpenRouter.
• Vercel AI Gateway edits already work through /v1/images/edits, so I left them alone.

Verification

• Live: POST /api/v1/images/edits returns 404. /api/v1/images/models lists input_references for nearly every model and no mask parameter anywhere.
• Recorded openrouter-images/edits-an-image-through-input-references with openai/gpt-image-1-mini (about $0.002). The black circle fixture came back as the same circle in green, so it's a real edit.
• New unit tests in test/image.test.ts: no images keeps the plain body, bytes and URLs become input_references on /images/generations as JSON, and a mask fails with UnsupportedOperation before any request. These and the recorded test fail without the fix.
• packages/ai: bun typecheck, bun run typecheck:tsc, bun test (1979 pass). Root: bun run check.

Visual

<svg xmlns="http://www.w3.org/2000/svg" width="240" height="72" viewBox="0 0 240 72" fill="none" stroke="currentColor" stroke-width="2">
  <style>
    .dot { animation: go 2.4s ease-in-out infinite; }
    .out { animation: fill 2.4s ease-in-out infinite; }
    @keyframes go { 0% { transform: translateX(0); opacity: 0; } 15% { opacity: 1; } 55% { transform: translateX(72px); opacity: 1; } 70%, 100% { transform: translateX(72px); opacity: 0; } }
    @keyframes fill { 0%, 50% { fill-opacity: 0.15; } 70%, 90% { fill-opacity: 0.9; } 100% { fill-opacity: 0.15; } }
    @media (prefers-reduced-motion: reduce) { .dot, .out { animation: none; } .dot { opacity: 0; } .out { fill-opacity: 0.9; } }
  </style>
  <rect x="24" y="12" width="48" height="48" rx="6" opacity="0.6"/>
  <circle cx="48" cy="36" r="12" fill="currentColor" fill-opacity="0.35" stroke="none"/>
  <line x1="84" y1="36" x2="156" y2="36" stroke-dasharray="3 5" opacity="0.4"/>
  <path d="M150 30 L157 36 L150 42" opacity="0.4"/>
  <circle class="dot" cx="84" cy="36" r="4" fill="currentColor" stroke="none"/>
  <rect x="168" y="12" width="48" height="48" rx="6"/>
  <circle class="out" cx="192" cy="36" r="12" fill="currentColor" fill-opacity="0.9"/>
</svg>