AI Chat Panel Generator.

No — one is a widget, the other is the pattern the widget hosts. A conversation is the organizing model: who spoke, in what order, how streaming appends, what a fork means. A chat panel is the embedded surface that implements that model inside someone else's page, with the constraints embedding adds: narrow width, host-page context, focus rules, panel behaviors like collapse and close. MeDo has a separate Conversation pattern page for the organizing model; use this page when the question is where the dialogue lives and how it behaves as a guest in your interface.

When the assistant stops advising and starts doing. A panel is one region; if the AI needs to edit files, run commands, show diffs, manage approvals, or drive a live preview that the user watches change, you have crossed into an agent workspace — a multi-region environment where the conversation is one column among code, files, and results. Adding a diff viewer and a terminal inside a 380px rail is not an upgrade, it is a workspace crushed into a panel. Keep the panel for asking, explaining, and summarizing around the page the user is already in.

That job belongs to a task input, not a panel. A landing-page prompt box is single-shot: one task in, one result out, no history, no thread. A chat panel implies an ongoing exchange — prior turns visible, context accumulating, a log that persists. Dropping a full panel onto a marketing page adds scroll weight and state you do not need. The common pattern is a single-shot entry on the landing page that hands off into a panel once the user commits, so the dialogue starts exactly when there is something to continue.

Three named pieces: a labelled panel region, a message area with role="log" — which politely announces newly appended messages to assistive technology without stealing focus — and a real textarea for the composer. Mark only new content for announcement, keep the stop and send controls as named buttons, return focus to the trigger when the panel closes, and support Escape to dismiss. Streaming complicates this: announce once when a reply starts rather than per token, which is exactly what the log role's polite behavior gives you.

Numbered superscripts in the reply, resolving to a compact source list directly beneath the message — title, source type, and a link. Skip inline URL previews at panel widths; they eat the column. Make each citation durable: if the panel closes or the thread reloads, the link still resolves. In a docs context, clicking a citation should scroll the host article to the cited section, which turns the citation from decoration into navigation and quietly proves the answer was grounded.

Keep the partial text visible and mark it — 'interrupted' or 'failed' — rather than deleting it, because the half answer often contains useful work. A Retry row resends the same request without the user retyping. Distinguish the failure causes in the copy: a network drop says reconnect, a rejected request says why, a failed tool says which step. And never lose the draft: if the user had typed a follow-up while waiting, it is still in the composer after the failure, untouched.