Image input¶

All four Kimi Code models accept images. When a request carries image content, the adapter resolves each image through the harness attachment service (ctx.attachments) and inlines it as a base64 image_url data URL in the OpenAI-compatible chat-completions body.

Flow¶

graph LR
  A[Request with image blocks] --> B{Model supports image?}
  B -->|No| R1[Reject: UNSUPPORTED_CONTENT]
  B -->|Yes| C{Attachment service mounted?}
  C -->|No| R2[Reject: INVALID_REQUEST]
  C -->|Yes| D{Within pixel and byte budgets?}
  D -->|No| R3[Reject: INVALID_REQUEST]
  D -->|Yes| E[Read bytes, encode base64 data URL]
  E --> F[Send image_url part to Kimi]

• Text-only requests never touch the attachment service.
• Image references are collected from every message, including nested tool results.
• Each image is budgeted from its durable reference metadata before any bytes are read.

Budgeting is internal¶

The plugin enforces its own budget from the durable attachment metadata (width, height, bytes) rather than depending on a request-scoped projection from the attachment service. An image is rejected when:

• its pixel count (width * height) exceeds imagePixelBudget (default 8294400, Kimi's 4K ceiling), or
• its encoded byte length exceeds imageMaxBytes (default 1048576, 1 MiB).

The byte budget keeps the base64-expanded payload within Kimi's per-message size limit. Both budgets are configurable in cordis.yml. See Configuration.

Rejection cases¶

Wire format¶

Each image becomes an ordered content part beside any text:

{
  "role": "user",
  "content": [
    { "type": "text", "text": "What single color fills this image?" },
    { "type": "image_url", "image_url": { "url": "data:image/png;base64,..." } }
  ]
}

Text-only messages keep the compact string form. Supported media types are PNG, JPEG, WebP, and GIF.