Agent Chat

Agent Chat is the window that opens when you click an agent on the World Map or in the Activity Bar — a streaming conversation with one specific specialist, with full controls over slash commands, mentions, tool approvals, mode settings, and conversation history.

How It Works

Every agent in Viberia has its own chat window, and the window has the same shape regardless of role. The agent has a system prompt, a model, a permission mode, an optional thinking mode, a list of skills it can invoke, and a list of MCP servers it can call. You see a header with the agent’s name and portrait, a scrollable transcript, an input box at the bottom with mode cells beneath it, and a History sidebar you can toggle.

When you send a message, the agent’s reply streams in token by token. Mid-stream you’ll see assistant text, tool call cards, and structured prompts (permission requests and ask-user questions) interleaved in the order they happened. Streaming is real — you can read along as the agent thinks, and you can interrupt with a Stop control if you want to cut a reply short.

The composer supports two power-user shortcuts: slash commands for invoking skills and @-mentions for referencing other agents or buildings. Both autocomplete as you type.

Composer features

Slash commands. Typing / at the start of the input opens a menu of every skill the agent can run, sourced from its skills list. Picking one inserts the command — for example /draft-a-prd — and you can type arguments after it. The full set of user-invocable slash commands is documented in Skills (Slash Commands).
@-mentions. Typing @ opens a picker of every project, building, and agent in your world. Selecting one inserts a structured reference; when the agent runs the message, it can route a message or task to the mentioned target.
Newlines. Shift-Enter inserts a newline; Enter sends.
Send-while-busy. If the agent is mid-reply, your new message is appended to the Queue and delivered when the agent is idle.

Mode cells

Three small cells sit below the input — they’re called Mode Cells and they configure the next message you send:

Model — pick the model the agent will use (provider-specific list; populated from your CLI Configuration).
Thinking — toggle thinking mode and pick a thinking budget. Off for fast turns; higher budgets for harder problems.
Permission — switch between Standard (per-tool prompts) and YOLO (auto-allow). Defaults are inherited from Settings → Agent Defaults but you can override per message.

Each cell pops a small selector when clicked. Changes apply to the next message; you can change them between turns.

Tool call rendering

When the agent uses tools, each call renders as a card in the transcript. Cards are grouped by intent so a sequence of reads doesn’t pollute the conversation:

Reads (open a file, list a directory) collapse into a single group labeled by count.
Writes (create a file, append) render with a preview of what was written.
Edits (modify a file) render with a colored diff.
Deletes render with a small confirmation card showing the target path.
Bash and shell commands show the command, exit code, and a truncated output you can expand.
MCP tool calls show the tool name, its arguments, and the response.

Each card is collapsible. Click the header to fold/unfold; long outputs are truncated with a “show more” affordance.

Permission requests

When the agent calls a tool that needs approval and you’re in Standard mode, a permission request card appears inline in the transcript with three buttons:

Allow — approve this call only.
Allow always — approve this tool for this agent for the session.
Deny — block this call. The agent gets the rejection and can choose another path.

The card includes the tool name, the arguments it’s asking to use, and a short rationale when the agent provides one. While the card is pending, the agent’s stream pauses; the moment you click a button, the stream resumes.

Ask-user questions

Some agents and skills surface structured questions back to you via an ask-user question card. The card has a question, optional context, and either free-text or multiple-choice answers. Replying advances the agent; ignoring it leaves the agent waiting in a paused state, surfaced as a status dot in the Activity Bar.

History sidebar

A toggleable History sidebar lists prior conversations with this agent. Click a row to load that conversation back into the transcript; click New chat to start a fresh one. Conversations are persisted per agent and survive restarts.

Context badge and compaction

The window’s header shows a Context Badge — a small gauge of how much of the model’s context window is in use. As the conversation grows, the gauge fills. When it gets near full, the badge offers a one-click /compact shortcut that runs the /compact skill: summarizing the conversation so far and starting fresh with the summary as primer. Old conversations that pre-date the context tracking show a — placeholder until the next completed turn fills it in.

Per-agent settings

A gear icon in the window header opens Agent Options, organized into five tabs:

Prompt — the agent’s system prompt. Edit and save to change the agent’s behavior.
Model & Tools — model defaults, tool allowlists, and tool defaults for this agent.
Skills — pick which skills this agent can invoke via /.
MCP — per-agent MCP servers, added on top of the building-level and global lists. See MCP Servers.
Directories — additional directories the agent can read from and write to beyond its building’s defaults.

UI Indicators

Streaming dot trails the cursor at the end of an assistant message while it’s still streaming.
Status dot in the window header — green idle, amber thinking, blue streaming, red waiting on permission or question.
Context badge gauge fills as the conversation grows.
Queued strip appears at the top of the composer when one or more messages are waiting in the Queue for this agent.
Mode cell highlight — a cell glows when its current value differs from your defaults.
Permission card pulse — pending permission cards pulse softly until you respond.
Question card pulse — pending ask-user cards pulse with a different accent color.
Tool group counter — collapsed read groups show a count, e.g. “Read 6 files”.
History badge on the sidebar toggle when there are unread updates to a prior conversation.

Use Cases

Hand a task to one agent. Open the Planner or Developer, drop a request, and watch it work in the open.
Run a skill end-to-end. Type /draft-a-prd on the Planner, then /handoff-prd-to-developer to ship the result.
Approve a tool call with confidence. Read the permission card’s exact arguments before clicking Allow, or pick Allow always to stop being asked for safe tools.
Bring another agent into the loop. @-mention a Reviewer or Dean and let the current agent route work to them.
Reduce context cost mid-conversation. Click /compact when the context badge crosses the threshold to summarize and continue without losing the thread.
Switch models per message. Use the Model mode cell to escalate a hard turn to a heavier model, then drop back to a lighter one for follow-ups.

Limitations

Closing the window does not pause the agent. The agent keeps streaming until done; you’ll see the result when you reopen the chat.
Editing the system prompt mid-conversation takes effect on the next message; it doesn’t retroactively change earlier turns.
Mode cells apply to the next send only; they don’t change the agent’s defaults. Adjust defaults in Settings → Agent Defaults.
Tool call output is truncated in the UI for performance — full output is still available to the agent and to history.
The History sidebar lists conversations for this agent only. Cross-agent search is not available from this window.