Message threads, streaming states, composers and citations.
Chat UI for a model is not messaging UI with a robot avatar. The differences are structural: responses stream token by token, they can be stopped, they carry sources worth linking, and a regenerate is a first-class action. These handle the streaming cursor, the stop button, the error-mid-stream case and the thread that has to stay scroll-anchored while text arrives.
A scrolling conversation with per-turn actions, announced as a polite log restricted to additions — so a streaming reply is not re-read on every token — and auto-scroll that only pins when the user is already at the bottom.
The trigger menus done properly: @ and / open only at a token boundary, the query is derived from the caret so the menu closes itself, and focus never leaves the textarea — the combobox pattern, not a floating div.
Token-by-token text with linked citations, follow-up chips and a caret that stops the moment the stream does. The status is announced once; the prose is deliberately outside every live region.
The answers a regenerate throws away — siblings rather than replacements, an edited question that branches in place, and a plain sentence about what happens to the replies below.
The "thought for 5s" panel: a native details/summary outside, individually expandable steps inside — buttons rather than nested disclosures, which are announced as a disclosure within a disclosure.
The trace after something broke: retries counted in the open with what each one cost, a step that was silently worked around drawn in amber rather than green, and the raw error kept apart from the agent’s reading of it.
The blank thread, answered: grouped starter prompts as real lists of buttons, each group labelled, so the highest-leverage screen in an AI product is not a centred logo.
The two-pane assistant layout, with the mismatch nobody else states: the canvas names the version it is showing and says so when that is not the one the conversation is about. Versions link back to the turn that made them.
Model selection with the three numbers the choice actually turns on — capability, latency and cost per message — and locked models explained rather than greyed out.
Files staged against a prompt with a running context-token budget, unreadable files flagged before send, removal allowed mid-upload, and a send button that names its one blocker.
Dense tool invocations that open into their arguments and result, with status carried by icon and word rather than colour, and errors marked as alerts only once the user opens them.
Run status as rows — done, running, failed, queued — with native progress elements, per-task retry buttons that name their task, and progress announced from one small status region instead of every row.
The gap before the first token: a shimmer grid, a phase label that advances as the wait lengthens, and a running counter — which is the only honest answer to "is it stuck". The timer is hidden from the live region on purpose.
What the agent tried, what failed, and what it did about it — the run history a user needs before they trust a second attempt.
What one run cost, per step, so an expensive answer is attributable rather than a monthly surprise.
Tool-call traces, approval gates, context panels and inline actions.
Thread lists, conversation views, composers and presence.
Dialogs, side panels, bottom sheets and confirmations.
Month grids, availability pickers, booking flows and time slots.
List-edit-delete screens, inline editing, bulk actions and audit rows.
Article layouts, post grids, author bylines and reading progress.