Interactive Conversation Guide

Overview

This chapter covers all the interaction details once you’ve entered a session in the CLI. It applies to the following scenarios:

• You’ve completed configuration and login, and want a deeper understanding of the session prompt, conversation modes, and runtime switching.
• You want to know how files generated by the agent’s tool calls are opened automatically.
• You want to familiarize yourself with the input experience, including command history, Tab completion, and / commands.

This chapter covers: launch and the prompt, conversation modes, runtime switching, tool-call approval, streaming status and cancellation, input experience, automatic opening of result files, file management with /files, and exiting.

Launch and the Prompt

How to Launch

The following two commands are exactly equivalent:

unitarylab            # No subcommand; automatically enters `start`
unitarylab start      # Explicit form

Both commands run the startup wizard before entering the session. If an LLM is already configured, the wizard adds an extra confirmation step that lets you choose between “continue with the current configuration” and “reconfigure,” so that you don’t start a conversation having forgotten what was last configured.

Startup Banner and Self-Check

Once started, the following appear in order:

1. Brand banner: Displays the UnitaryLab name.
2. Startup self-check: A single-line metadata display.
3. Wizard steps: See the “Start a Session” section of the Quick Start chapter.

The Session Prompt

Once you’re in a session, the prompt in the terminal looks like this:

[agent│model] ›

Meaning:

• mode: The current conversation mode. Possible values are agent, ask, and plan.
• model: A short name for the current model.

Conversation Modes

The CLI has three built-in conversation modes, each applied to the current session:

• agent (default): Full tool chain. The agent can use all available tools, including file reading and writing, remote workspaces, and skills.
• ask: Retrieval-based Q&A mode. Geared toward retrieval-based question answering, with more restrained tool use.
• plan: Planning-oriented mode. The agent tends to produce an execution plan first and then carry it out step by step.

Switching modes does not lose conversation history, and the change takes effect immediately on the next turn.

Runtime Switching

You can use the following commands to dynamically change parameters between any two turns, with no restart required:

After typing /mode, /model, /models, or /session, press Space to see the available values.

Tool-Call Approval

In Agent mode, when the agent is about to call a non-safe tool (such as writing files or creating tasks), an approval confirmation dialog appears. You decide whether to allow the operation.

What the approval dialog shows:

• A Chinese-language action description such as “Loading skill description” or “Creating project folder,” making the agent’s intent clear at a glance.
• Key parameters displayed in a structured way, such as skill file paths and task IDs.

Key bindings:

Approval behavior:

• Denying a tool does not prevent the same type of tool from being confirmed again later — a previous denial will not silently skip future prompts.
• Approved tool-and-parameter combinations are remembered for the current session, so you won’t be asked repeatedly for the same operation.
• Any unfinished input remaining in the input box is automatically cleared before the approval dialog appears, preventing an accidental Enter key from rejecting the dialog.

Pressing Ctrl+C during tool approval only rejects the current tool call — it does not exit the CLI. Destructive tools are never cached, so you will be asked to confirm each time.

Streaming Status and Cancellation

In an interactive terminal, the CLI shows the current task phase in real time so you know exactly what the agent is doing:

• Preparing request
• Analyzing
• Generating
• Executing intermediate steps
• Waiting for user confirmation
• Continuing
• Organizing final answer
• Canceled

Cancellation:

• Press Esc at any time to cancel the current inference.
• Long-running tasks display a cancellation hint.
• During multi-turn tool-call scenarios, the CLI suppresses intermediate bursts to present mostly the final answer, reducing noise.

Input Experience

The session provides the following input enhancements in the terminal:

• Command history: Use ↑ / ↓ to scroll through history.
• Slash-command completion: Typing / triggers command and skill completion; press Tab to confirm a candidate, and Backspace refreshes the candidates in real time.
• Second-level completion: See the runtime switching commands in the previous section.

/ is only used in the input box to trigger slash-command completion. Once you’re inside the fuzzy model selector opened by /model, /models, and similar commands, type keywords directly to filter — there is no need to type / again.

File Management: /files

/files is used to view and reopen files in the current working directory. Common forms are:

/files operates on the current working directory (see the Session and Working directory Guide), not the process’s current directory.

Help: /help

Running /help in a session opens a help panel that is color-coded by category:

• Switching (blue): /mode, /model, and so on.
• Skills (orange): /skills and every /<skill_name>.
• History (purple): Session-related commands.
• Files (green): File and working-directory commands.
• Session (gray): Exiting the session, agent-interaction commands, and the like.

Exiting

You can exit a session in any of the following ways:

• Type exit, quit, or :q and press Enter.
• Press Ctrl-D.
• Press Ctrl-C.

After you exit, the session and working directory information are kept locally so they can be used the next time you launch the CLI.

Recommendations

• Stay within the same session context: For day-to-day work, complete continuous tasks within a single session so the agent has full context; start a new session only when you begin a new task.

• Switch modes carefully: After switching from agent to ask or plan, the agent’s tool-use capabilities become more limited — check that the current task can still run as expected.

• Make good use of the help command: When you come across an unfamiliar command, run /help first to view the help panel, then consult the relevant chapter of this manual for more detail.