feat(ui): add moderated group conversations

Let users give installed models individual turns or ordered rounds in a shared conversation. Attribute completed responses and exclude interrupted output from subsequent prompts.

Add streaming and cancellation tests, browser coverage for CI, and a guide for the session-only page.

Assisted-by: Codex:GPT-6

docs/content/features/group-chat.md

@@ -0,0 +1,56 @@
++++
+title = "Group chat"
+weight = 23
+toc = true
+description = "Moderate a shared conversation between installed chat models"
+categories = ["Features"]
++++
+
+Group chat lets you moderate a text conversation between 2–6 participants.
+Each participant uses an installed chat model and has a unique display name.
+You can use the same model for multiple participants.
+
+## Start a conversation
+
+1. Open **Chat** in the web UI and select **Group chat**. The page is also available at `/app/group-chat`.
+2. Select a model under **Model to add**, then select **Add participant**.
+3. Add at least two participants. Give each participant a unique, nonempty name.
+4. Enter an **Objective**, such as “Compare three ways to reduce our application's startup time.”
+5. Optionally enter a **Moderator message** and select **Add message**.
+
+Participants and the objective remain editable until the first model turn.
+After that turn starts, select **New conversation** to unlock setup and clear the transcript.
+The page uses the existing chat permission and chat completions API.
+
+## Control turns
+
+Select **Give [name] a turn** to request one response from that participant.
+Select **Run one round** to give every participant one turn, in the displayed order.
+To repeat that order, set **Number of rounds** to 1–10 and select **Run rounds**.
+You can add moderator messages between runs.
+
+Each request includes the objective, the participant's identity, and all completed contributions with their speakers' names.
+Other participants' responses appear as shared context, not as the current model's previous assistant responses.
+Private reasoning is neither displayed nor included in later requests.
+
+The page sends one request at a time. It does not generate responses concurrently.
+Sequential requests avoid simultaneous generation, but do not unload models or guarantee that only one model remains in memory.
+Configure model loading and memory limits on your LocalAI server as needed.
+
+## Stop or recover from an error
+
+Select **Stop** to cancel the active request and all remaining turns in the run.
+A server error, malformed or truncated stream, empty response, or output limit also stops the run.
+Partial responses remain visible as **Incomplete** and are excluded from future requests.
+You can resume with another turn or start a new conversation.
+
+## Limits and history
+
+{{% notice note %}}
+History exists only while this page is open. Leaving the page or reloading clears it and cancels any active request.
+Group chat does not save conversations or support attachments, tool calls, or automatic speaker selection.
+{{% /notice %}}
+
+Each request includes the full completed transcript. Long conversations can exceed a model's context window.
+Start a new conversation when the discussion becomes too long for your selected models.
+Select **Back to Chat** to return to normal chat.