Chat & Streaming

​

The input bar

The input box itself is multi-line and grows upward as your message wraps to more lines — up to about 92 extra pixels (roughly five lines) — before an internal scrollbar takes over, so the rest of the chat layout never gets pushed around by a long prompt.

• Enter sends. Shift+Enter inserts a newline.
• / opens slash-command autocomplete.
• @ opens autocomplete for instances and scripts in the place, and for skills via @skill:name.
• The + button opens the tray; its image-attach button is described in Vision & Images.

For the full list of commands and a deeper look at how mentions resolve, see Mentions & Commands.

​

Markdown rendering

Assistant replies render through a small markdown engine built for this UI: bold, italic, inline code, fenced code blocks, lists, headings — and tables, rendered as a real grid of Frame instances rather than plain text, since Roblox’s RichText has no <table> tag.

​

Streaming, live

Both halves of a turn stream now, not just the text:

​

Stop, for real this time

Now:

• The underlying SSE connection is polled for cancellation roughly every 50ms while it’s open, and closed immediately the moment you hit Stop — instead of waiting up to two minutes for the response to finish on its own.
• If a permission prompt (Allow / Deny / Always Allow) is on screen when you hit Stop, it’s automatically resolved as Deny so the agent loop isn’t left blocked waiting on a button you were never going to click.

You can trigger it from the Stop button (it replaces Send while the agent is busy) or by typing /stop.

​

Reliability, under the hood

• Transient network failures and 5xx responses now retry automatically — up to 3 attempts with backoff — instead of failing the turn outright. A 4xx never retries, since the request itself was the problem.
• Fixed a bug where a response could start streaming and then stall permanently, looking like the agent “reads a skill, thinks, then never responds.” The stream timeout increased from 120s to 180s, and now tells apart partial content that already arrived (shows a real timeout error) from nothing having arrived at all (falls back to a non-streaming request instead of hanging).

​

Reading the transcript

• The transcript auto-follows new content, but a genuine scroll on your part (wheel, touch, or dragging the scrollbar) pauses that for 3 seconds before it resumes — so reading back through history doesn’t get yanked back to the bottom mid-scroll.
• Click one of your own past messages to copy it: it swaps to a selectable-but-not-editable text box, so a stray keystroke can’t corrupt your own history while you copy it.
• A cost/cache footer can appear under a finished response (e.g. ~0.000123 credits · 42% cached), controlled by a Show Cost Info toggle in Settings (on by default).
• A subagent finishing shows up right in the transcript, as a small gray “Subagent done: label” line — in addition to the toast and the rail badge covered in Subagents.

​

Hitting the tool-call limit

Each turn is capped at 25 tool-call rounds, to keep a confused agent from looping forever. If it hits that cap mid-task, you get an amber card with a Continue button that just sends the literal message "Continue" — letting the agent pick up exactly where it left off instead of you having to retype anything.

​

History

Every chat is saved per-game via ChatStore. Opening it lists past chats for the current place, with a + New chat shortcut at the top.