Product
Attached files reach research runs, DeepSeek tool calls no longer break chats
Files you attach to a chat now reach Deep Research and Ultracode runs, and raw DeepSeek tool calls run instead of printing markup.
Until now a Deep Research or Ultracode run saw only the short text you typed. If you attached a brief and wrote "start", the planner got the word "start" and never the file. Both kinds of run now take the files of the chat along. They appear in the task as an attached brief. A file that is too large for the prompt shows up as an outline, its first lines and its path under /work/uploads, and the agents read the rest inside the container. Only files from your own chat are used.
Text version
Before: a research or Ultracode run got only the typed task, so an attached file was never seen. After: the run gets the task and the files of the chat as an attached brief, with outline, first lines and path for large files. Before: a DeepSeek reply in raw DSML text showed the markup and the turn stopped. After: the call runs as a tool call and the markup is hidden.
Some DeepSeek V4.1 replies announce a tool call in a raw format marked DSML instead of the normal call. The chat printed that markup as text and the turn stopped. ZeroLog now reads these calls, runs them and hides the markup, also while the reply is still streaming. A tool the model has not loaded gets an answer that tells it how to load the tool.
Two smaller fixes: the Ultracode tools now stay available for the whole turn, even after the agent has loaded several other tools. And tool output no longer carries credentials: passwords in addresses and proxy settings are replaced by three asterisks before the output is stored or sent to a model.
Images in a chat are named in the brief but not described, because the planning models read text only.