What
When a prompt includes an image, Claude Code now checks the image before resizing it. If the image is malformed, it is replaced with a text note reading "[Image could not be processed: ...]" and the turn carries on. This covers an image that:
- has no source data attached
- has source data that is not in the expected form
- has a file type that is not written as text
- contains data that cannot be decoded
- fails during resizing
Before, most of these cases stopped the turn with an error.
Why
This matters most for programs that drive Claude Code and send it messages directly, for example through the SDK (the toolkit developers use to build on Claude Code). A bad image block from such a program now costs one image rather than the whole turn.
Something disagreesSomething we can check disagrees with this entry, or the writer said they could not settle it.
The writer flagged doubt
It is not known whether images arriving by other routes than prompt submission get the same check.