Repository navigation
Releases
Releases · stonega/cusco
v0.5.56
Added
- Add Kimi K3.1 Pro with a 1M-token context window, image input, and Low, High, and Max reasoning efforts.
Fixed
- Recover damaged conversation index markers or quarantine unrecoverable ones to stop repeated parse errors during saves, and publish asynchronous writes only after they are complete.
v0.5.55
Added
- Add DeepSeek Search as a web search fallback, sharing the DeepSeek API key and returning cited web results.
v0.5.54
Removed
- Remove desktop notifications after 10 seconds without provider updates.
v0.5.53
Added
- Add Take over to the background desktop preview so users can click, drag, scroll, and type after stopping the agent's current turn.
- Add a Remove All button to Archived Chats; hold it for three seconds to permanently delete all archived chats.
v0.5.52
Added
- Run Computer Use in an independent background GNOME desktop by default, with a separate browser profile, application launch tool, and a visible stop control.
- Preview the live agent desktop by hovering over Stop agent desktop, and open the preview in a separate window.
Changed
- Start each Computer Use turn on Background; agents can select Current desktop when the user requests the foreground or an already open window, without a fixed mode setting.
- Keep the agent desktop hover preview compact, remove its heading, and use an icon button to open the separate window.
Fixed
- Animate the chat transcript to the top after loading earlier messages.
- Start Cusco from source when the installed GSettings schema lacks newer Computer Use keys by using the fallback settings store.
v0.5.51
Changed
- Show Agent working time in tenths of a second for runs shorter than one minute.
v0.5.50
Fixed
- Open Cusco and connect local or unauthenticated MCP servers when GNOME Keyring's Secret Service is unavailable.
- Fix the computer-use GNOME Shell extension failing to enable on GNOME 51 after removal of the legacy Clutter backend API, while retaining support for GNOME 45–50.
v0.5.49
Fixed
- Save bundled plugin removal state per user so removing and reinstalling plugins works in RPM and other read-only system installations.
v0.5.48
Added
- Render inline and display LaTeX math in chat messages, tables, and Markdown document previews, with offline native GTK display and copyable TeX source.
- Add a predefined Chrome DevTools plugin with the official local MCP server, seven bundled skills, and a helper for creating a named browser tab group per task.
Changed
- Reserve plugin Connect buttons for authentication; configure plugins without authentication automatically before tool discovery.
- Use the running Chrome profile and its existing logins for Chrome DevTools tasks, retaining named task tab groups. Upgrade saved original preset arguments while preserving custom configurations.
Fixed
- Keep the transcript in place after scrolling up during a response, including later streamed updates and response completion; resume following when returning to the bottom.
- Warn about response delays only after provider inactivity, resetting on streamed updates and suppressing warnings during tool execution or approval waits.
- Preserve bundled plugin sources when removing a plugin, so its skills and MCP configuration remain available for reinstalling, and reliably save removal state on Ubuntu.
- Configure and start local MCP servers when installing plugins, with startup errors available under Plugins → MCP.
- Include local skill directories in agent context so bundled references and helper assets can be found from any chat working directory.
v0.5.47
Added
- Manage the default chat and image models in a validated JSON catalog, with daily GitHub updates, offline fallback, and manual refresh controls in provider settings.
Changed
- Replace the two retired DeepSeek Flash entries with
deepseek-flash, including vision support, and migrate saved selections to the canonical model. Flash uses the configured search fallback because its API ignores native web search.
Fixed
- Preserve complete model parameters during catalog updates and prevent stale discovery results from overriding built-in capabilities or hiding new models.