[Im-Beast]

• Requests over a provider's image limit keep the most recent images and send each older one as a short text placeholder in its place
  • Uploaded files and images returned by tools both count; PDF pages do not
  • Tool calls and their results keep their pairing
  • Images are omitted in batches of a fifth of the limit, so the prompt prefix stays cacheable across turns
• Limits per adapter:
  • Azure OpenAI: 50
  • OpenAI: 1,500
  • Anthropic: 100 on 200k-context models, 600 on the rest
  • Gemini and Vertex: none, since their 3,000+ is out of reach
• openrouterModel and openAICompletionsModel take a maxImages option, because OpenRouter applies whichever limit its upstream provider has