Improve interface design

[ErikBjare]

• Improve how tool previews/confirmations work
  • inline instead of modal?
  • show pending tool while tool execution waiting to confirm/in progress (so that user doesn't wonder "what's happening?" while waiting for a tool response)
  • enter to confirm e3cdfae
  • esc to skip (already works?)
• Make the assistant always reply ("auto-step") to its tooluses Auto-step conversation #9 (not only if auto-confirm is on)
  • Might need backend changes, but could also just be just a frontend thing?
• Disable scroll-to-bottom while reading past messages Disable auto-scrolling during generation if user scrolls up #27
• Fix bug introduced in feat: rewrite chat rendering with LegendState #26 where messages are no longer correctly chained for streamed messages (margin between messages, rounding is wrong).
  • Screenshot:
    
  • Reloading the conversation makes it look correct again
  • Fixed in fix: chatmessage styling not updating on new messages #31
• Fix bug where you cannot open contents of <details> while streaming (instantly closes as component re-renders on each streamed token)
  • Example: ask it to output a long "```txt" block, try to open it
  • "Ideal" design/UX:
    1. show a 3-5 line scrolling preview of block contents as they stream in
    2. collapse as done generating
    3. click to expand
  • Fixed in: feat: streaming markdown parsing #33
• Improve general design/style
  • Make light mode default? (I think yes)
  • Improve contrasts/visibility in dark mode
  • Warning: unbounded task (timebox it!)
• "Stop" is unreliable/doesn't work
  • interrupts the conversation correctly but button remains in "stop" state and doesn't switch back to the "send" state.
  • It worked before so it was either a legendstate PR issue, or I messed it up afterwards when trying to fix something else.
  • Fixed here: f1d3a17
• Add button to preview markdown code blocks
  • Currently blocks of markdown ( ```markdown ) will not render the markdown as html. Would be nice to have a button to toggle between showing the actual markdown contents and the rendered document.
• Remove gptme-webui menu bar from gptme-landing chat and put connection state somewhere else.
  • Should we even show the current connection button with the modal that shows URL and user token? For the hosted service perhaps the user should not be bothered with this info? Might be confusing. But would still be good to show the connection state somewhere and maybe have a reconnect button if connection fails?
  • Perhaps we should have an instance-switcher on the chat page so you don't have to navigate back to the dashboard?