1.42.3 — Image input on guarded Bedrock routes
Editorial identity incomplete
2026-11-09
A guarded LLM call can now send images to Nova 2 Lite on Bedrock — in llm.call(), llm.stream() and inside tool loops. Images are validated bytes, never URLs, and a route admits them only when its policy adds the images capability.
Added
llm.image_block(data, media_type)turns PNG, JPEG, GIF or WebP bytes into a validated image block;llm.ImageInputErrorcarriesimage_invalidorinput_too_large.- Bedrock routes may add
"images"to their capabilities, with or without"tools". A user-message image then reaches Nova 2 Lite in plain calls, streams and every round of a tool loop. - New refusal code
image_invalid.
Changed
- Image input on every provider is checked before any budget, ledger row or dispatch: at most 5 images per request, 5 MiB each and 20 MiB in total, each side at most 8000 px, and only as top-level blocks of a user message. The assistant's file analysis tool now refuses images outside these limits up front rather than at the provider.
- Each image reserves 1600 tokens against the budget, and image data no longer counts toward
max_input_byteson any provider — including Anthropic routes that already admitted images. The framework limits above replace the byte cap for images.
Fixed
- A tool argument shaped like an image block no longer forces the
imagescapability or refuses the rest of a tool loop.
Upgrade notes
- Adding
"images"to a route changes the policy hash: updateLLM_SAFETY_POLICY_EXPECTED_HASHand activate the deployed policy again. Routes without it are unchanged and refuse images withcapability_unsupported. - The 5-image limit counts the whole transcript, so a chat that resends earlier screenshots must drop old ones. Nova bills 230 input tokens per image on every round it is resent.
- Images inside a tool result, and PDF or other document blocks, are not supported.