What
You can point Claude at a file by typing @ followed by its name. When that file is too large to include in the message, Claude Code now handles it differently.
- Before, the mention was silently dropped. Now it becomes an
at_mention_referencemarked unread for beingtoo_large, and it includes the file's size (fileSize). - The file is also tagged
readAllOfIt: 'over_size_limit'. A file that fails the size check in tokens and is also over the byte limit is now listed as unread rather than skipped. Tokens are the small chunks of text the model counts. - The note sent to Claude says: "It was too large to attach, so you have not seen any of it."
- Claude is told to read the whole file in portions using
offsetandlimit, unless your request needs only part of it. The note says how many reads that will take, at roughly 60 KB or 25,000 tokens per read. - Before, the note told Claude to read in portions, or to read the whole file in one call if it truly needed all of it.
- Claude is also told to tell you when its answer comes from searches or a partial read.
- A new notice tells Claude that project docs and PDFs were read before its reply, for example "The PDF file that came with this message was read before your reply".
Why
Claude now knows a large mentioned file exists instead of never hearing about it. You should see answers based on the whole file more often, and be told when an answer comes from only part of it.