Chat & Conversations

Chat is where you work with the assistant. It reads your workspace context, inspects governance signals, and proposes or carries out actions. Its actions are checked against your policies and recorded to the audit trail. Open Chat in the sidebar, under the Operate group (or go to /workspace/chat).

Live execution against external ad platforms is connector-gated and not enabled by default. See Operations.

Writing a Message

Type in the composer and press Enter to send, or Shift+Enter for a new line. Stop generating halts an answer in progress. An empty conversation offers three starters — Review queue context, Check policy risk, and Plan an automation — that fill the composer for you to edit and send.

The bar above the composer shows what the assistant has in context: the current page, project memory, active skills, available tools, and a warning when a connection is unhealthy. Select it to expand each part. See Knowledge and Skills for what feeds it. The header names the conversation and carries a Governed agent label.

Reading an Answer

Hover over a message for Copy message, Delete message (permanent), and, on the latest answer, Retry message. Save to Knowledge saves an answer as a knowledge record.

Good response and Bad response rate an answer. A good rating is sent at once; a bad one opens an optional comment (What went wrong? (optional)) and Submit. The rating is stored with the conversation and the skills that produced the answer.

Governed answers carry a footer with the skill that handled the turn, a policy summary such as "All 3 passed" (hover for each rule), and a confidence score. The hover detail marks each rule as "blocked the call" (an error-severity rule stopped it), "ran anyway, recorded" (a warning let it run and logged it), or "advisory" (information only).

How Proposed Actions Are Governed

When a tool call needs a decision, the thread shows a card instead of running it silently:

Rules, severities, and what each one does are described in Blocking rules. Everything the assistant does is recorded in the Audit log.

Conversation History

On screens 1280 pixels wide or more, a history rail docks beside the thread; the History button collapses it, and your choice is remembered on that device. On narrower screens the button opens it as a panel.

The rail lists the current project's conversations, newest first under Today, Yesterday, and Earlier, thirty at a time with Show more. Type in Search conversations to search the whole history (up to 50 recent matches). Hover a row for Rename conversation and Delete conversation; deleting asks for confirmation and removes the conversation and its messages for good. All conversations opens the full list.

The Conversations Page

/workspace/conversations is the searchable record of what the rail shows. It is not in the sidebar; reach it from All conversations in the rail. Its New chat button opens Chat.

• Four figures: conversations, active today, this week, and total messages
• A date range, the last 30 days by default. An empty range offers to widen to include the most recent activity
• The chips All and Needs attention — the second lists conversations with a turn that failed a policy check or had low confidence, marked with a dot
• A search box that matches title, message text, or a conversation ID
• Each row shows the title, a snippet of the latest message, the message count, and last-active and started dates. Its menu has Open, Rename, Export (Markdown), and Delete. Tick rows to delete several, and use Load more for older ones

The list follows the active project, chosen in the project switcher at the top of the sidebar (see Projects).

A Conversation's Page

Select a conversation (/workspace/conversations/[id]) to review it. Continue with the agent loads it back into the assistant on the conversation's project, and Export copies it as Markdown or downloads a .md file. Three tabs:

• Thread — the messages, with each answer's policy footer
• Agent reasoning — per turn: confidence, the policy checks, and the tools called with their status. This is the trace an auditor reads
• Timeline — the messages in time order

Side cards show the Current handler (the skill that last answered) and Governance: policies applied, the lowest confidence seen, and any failing checks. On each answer, Take action offers Seed rule (opens Policies with that turn as the starting point), Add to knowledge, and Flag for skill tuning, which needs a skill to have been active on the turn. What you did is shown under the answer as Rule seeded, Added to knowledge, or Flagged for tuning.