[ztcs-software]

What's new

• Work in a folder. Add a folder to a chat from the + menu next to the message box and the model can list, find, read, search and edit the files in it. It stays sandboxed to that folder and has no shell. Before anything is written, edited, moved or deleted, Loach shows you the change itself (the new contents, a line-by-line diff, from → to, or a deletion warning) and asks, unless you've allowed that tool for the chat. Line endings and non-UTF-8 encodings are kept as they were. With a cloud provider, files the model reads are sent to it like any other message.
• Project instructions. A LOACHFILE.md at the root of a chat's folder is added to every turn: how to run the tests, what not to touch, house style. It's read fresh each time, so an edit applies on your next message. There's a sample to start from.
• Global memories. Turn this on in Settings → Features (or in setup's Pick your defaults step). Loach then notes lasting facts about you from chats outside a Space and brings them into every chat, Spaces included. It's off by default, and Private Chat never reads or writes memory.
• Memory that keeps up. Space and global memory now update a fact that changed, or drop one that no longer holds, instead of stacking up contradictions. Every change shows an Undo. The memory editor adds search, shows when each fact was saved and links back to the chat it came from.
• Local MCP servers. Settings → MCP can now run servers as local programs (npx, uvx, node or a full path) as well as over HTTP. A native dialog shows exactly what will run and asks before any local program starts. Environment variables, which often hold keys, are left out of backups.
• Approve MCP tool calls. An MCP tool call now pauses the reply and shows the server, the tool and its arguments, and you choose Allow once, Always allow for that tool, or Deny. This is on by default, so servers you already had will start asking too. Switch Ask before each tool call off on a server you trust.

Fixes

• Compacted chats keep their summary. In a Space with its own instructions, compacting a chat used to lose the summary, so the model saw neither the summary nor the turns it replaced. A compacted chat with no instructions of its own also stopped following your global Custom instructions. Both are fixed.
• Private Chat stays private. A file dropped onto Private Chat no longer lands in the regular chat behind it, and keyboard focus stays in Private Chat. Private Chat also sends the settings its panel shows, instead of leaving the unchanged ones to Ollama's defaults.
• /remember respects the memory switch. It no longer saves a fact in a Space whose memory is turned off.
• Customize Hugging Face models. Models pulled as hf.co/user/repo:tag can now be saved as new models from the Models editor.
• Clearer errors. Running out of memory while loading a model now says so and suggests a smaller context or model. An API key withheld over plain http:// is explained instead of looking like a wrong key.