v0.5.8

• A tab marked while you were away stays marked until you look at it
• GPT models now describe their work beside the spinner instead of adding thinking items
• Each provider now lets you hide models from the list that is shown
• Tool lists and MCP errors now say why a tool never reached the model
• Pressing Send repeatedly on a slow connection no longer posts the message several times
• A message sent while the connection is down stays in the box instead of vanishing
• A Windows build from source no longer leaves a console window open behind the app
• The wails submodule clones over HTTPS, so building from source needs no SSH key
• Instances reached at a <project>.localhost hostname now work instead of refusing every request
• LM Studio models now report the context window they are loaded with, not a conservative 8K
• A reply cut off at the output limit now says so, instead of being retried three times
• A background command's properties now show its live output and exit code, and can stop it
• A premium serving speed now shows on the model button and can be set as the default
• The composer and Settings now share one model picker, with type-to-filter and keyboard navigation
• A message can now be scheduled for the end of the current turn, not just a delay
• A stalled tool engine is now noticed and revived, instead of failing every tool until restart
• The Docker image runs the engine in Node, so it needs no virtual display
• The Node engine host now starts on every supported Node 22, not just 22.7 and later
• A sub-agent is no longer offered the tools it could only have failed to call
• A resting conversation now says when it was last updated, at the end of the transcript
• Find no longer counts hidden text, and reveals the match itself rather than the message holding it

Full Changelog: v0.5.7...v0.5.8