Workspace

The Workspace is where you and the agent get things done. It runs in your browser on top of the open-source DeepSeek Harness (Apache-2.0), and everything the agent does happens inside an isolated Linux container that belongs to your account.

Layout

• Sidebar. On the left, the sidebar lists your projects and the sessions inside each of them. This is where you switch context, start new work and find earlier conversations.
• Conversation. The current session: your messages, the agent's replies, its plan, the commands it ran and the files it changed.
• Message box. Where you write to the agent, attach files, reference files with @ and type slash commands. On Pro and Max plans it also has a microphone button for voice input.

Sessions

A session is one conversation with the agent, attached to one project. The project folder is the agent's working directory, so file paths, scripts and outputs all stay in one place.

What the agent can do

Within a session the agent can:

• Plan a task and explain its approach before starting.
• Write code and run it.
• Execute shell commands in your container.
• Read and write files in your project folders.
• Browse the web to look up documentation or check a page.
• Call tools, including any connectors you have configured.

It tells you what it is doing as it goes, so you can step in at any point.

Questions and approvals

The agent is not a one-shot command runner. When a request is ambiguous it asks you a question and waits for the answer. When an action is risky, such as deleting files or changing something outside the project, it requests your approval first.

How often it stops to ask depends on your permission preset. A stricter preset means more confirmations and fewer surprises; a more permissive one lets the agent move faster on its own. Pick the level you are comfortable with for the task at hand, and review the agent's actions either way.

Managing sessions

• Rename a session so its title reflects what it is about.
• Archive finished sessions to keep the sidebar tidy. Archiving hides a session from the main list without deleting it.
• Fork a session to branch off from its current state. The original keeps its history; the fork continues independently. This is useful for trying an alternative approach, or for reusing a well-prepared context for a new task.
• Search across sessions to find the conversation where you solved a similar problem before.

Slash commands

Slash commands are typed into the message box and start with /. Two you will use often:

• /compact summarizes the conversation so far and replaces the long history with that summary. Use it when a session has grown large and you want to keep going without carrying every earlier message. Windlass also compacts long sessions automatically.
• /export downloads the session so you can keep it outside Windlass or share it with someone.

Referencing files with @

Type @ in the message box and pick a file from your project. The agent receives the exact path, so there is no ambiguity about which config.json you mean. This is the quickest way to point the agent at a file you uploaded, an attachment from an earlier message or a file it created itself.

Long conversations

Sessions support a context window of up to 1M tokens, which is enough for very long working sessions. As the conversation approaches its limit, Windlass compacts it automatically. Compaction keeps the important facts, decisions and file references while dropping repetitive detail. If you would rather control the moment yourself, run /compact at a natural checkpoint, for example after a feature is finished and tested.

Tips

• Give the agent a goal and constraints rather than a list of keystrokes. "Add input validation to the signup form and write tests for it" works better than a step-by-step script.
• Keep one topic per session. A new session starts with a clean context, which makes the agent more accurate and uses less of your allowance.
• Fork before an experiment. If it does not work out, the original session is untouched.
• Attach the files that matter and reference them with @ instead of pasting long contents into the message.
• Read the Models and effort page to decide when a task deserves DeepSeek V4 Pro or High reasoning effort.