[ztcs-software]

What's new

• Library. A new sidebar page for the texts, files (PDF, Word, code, images) and web links you give your models, saved once and used anywhere. Group them into datasets, then attach an item or a whole dataset to a chat (+ → From Library, or type #) or to a Space. Sources are linked, not copied, so an edit reaches every chat on its next message. A reply, or a message's attachments, can be saved to the Library from its menu.
• Sources sized to the context window. Small sources go to the model in full. Once they outgrow half the window, only the passages most relevant to each message are sent. Choose Auto, Full text or Excerpts from the Sources chip, and every reply shows which sources it used. Search works by keyword out of the box, and by meaning once you add an embedding model in Settings → Library: one click downloads one for Ollama, which keeps everything on your machine, or use your OpenAI-compatible endpoint.
• Spaces use the Library too. A Space's Sources tab takes Library items and datasets alongside uploads, shows how much of the window they take, and lets you pause any one of them. A new Settings tab brings together the Space's default model, memory auto-save and source mode. Spaces made before 1.6 keep sending their sources in full until you switch them to Auto.
• HTML preview. HTML code blocks and .html attachments get a Preview in the code canvas. The page renders offline in a sandbox with scripts off, and Run scripts turns them on for that page.
• Image scaling. Images are now scaled down before they're sent, to 1024 px on the longest side by default. A full-size photo costs thousands of tokens, and many local models read it worse, not better. The original stays in your chat. Change it or turn it off in Settings → Features, for Library images in Settings → Library, or per Space and per chat.
• More slash commands. /stop, /find, /import, /label, /folder, /tone, /set, /unload, /theme and /lock join the palette, which now completes model, Space, snippet and persona names. Bare /model opens the model picker.
• Drafts stay with their chat. Unsent text and attached files belong to the chat (or Space page) you wrote them in, so switching chats no longer carries them along.
• At home on macOS. Native window buttons, a menu bar icon that shows the fish, Cmd-only shortcuts, and a Dock click that reopens the window. Closing the window hides Loach instead of quitting it, and locks it when an app lock is set.
• Smaller touches. Launching Loach while it's running brings the open window forward instead of starting a second copy. Search (Cmd/Ctrl+K) covers the Library, with in:library to narrow it. The model picker lists every model an OpenAI-compatible endpoint offers.

Fixes

• OpenAI reasoning models work. newer GPT models that refuse max_tokens or a custom temperature are retried with the fields they accept.
• Steadier app lock. The lock screen at launch no longer reads the system keychain, so there's no keychain prompt before you've typed anything. Failed unlock attempts survive a restart, and locking closes open dialogs and ends a Private Chat.
• Private Chat stays private. It no longer offers built-in tools or Ollama cloud models, and it tells you when Ollama runs on another machine.
• Safer workspace edits. Changes to git hook folders and bare repositories always ask first, paths using Windows short names (IMPORT~1.MD) are refused, and an edited file keeps its Windows permissions.
• Sturdier MCP. A long tool call that reports progress is no longer cut off after 30 seconds, a hung server restarts on the next call, and tokens in a server's arguments or URL are left out of backups.
• Keyboards and text files. Shortcuts work on Cyrillic and Greek layouts, Cmd/Ctrl + / works where / needs Shift, and UTF-16 and Windows-1250/1252 text files open correctly as attachments.
• Closer counts. The context bar includes a Space's instructions and sources and estimates Chinese, Japanese and Korean text properly. The count tool treats an emoji or accented letter as one character.
• The Models editor asks before saving over an existing tag, the base model included.
• Many smaller fixes and hardening throughout, plus dependency updates including Tauri 2.12.

Good to know

• Backups made by 1.6.0 include the Library, so Loach 1.5 and older refuse to restore them. 1.6.0 still restores older backups.