Workspace View – Interface Guide

Once a project is created and its repository is connected, day-to-day development happens inside the workspace. This is where you collaborate with agents, explore the codebase, implement changes, run commands, and review results, all scoped to a single project and trial.

The workspace is organized into three main areas:

• Left sidebar: the task list for the project
• Center: the chat interface and active session
• Right panel: a context-sensitive panel that shows the file tree, terminal, context documents, sub-tasks, or other views depending on what is active

The status bar at the bottom shows the current sandbox state (TDX Sandbox), Create PR shortcut, indexing status (Indexed), active MCP count, and the current project name.

How Work Is Structured: Tasks, Trials, and Sessions

Work is organized into three levels:

Task: what you want done. For example: implement a feature, fix a bug, refactor a module, add tests, or perform a security review.

Trial: an attempt to complete that task. You can create multiple trials for the same task to explore different approaches, restart from a clean state, or compare implementations. The current branch shown in the chat header reflects the active trial (for example, trial-96acac35 (branched from main)).

Session: the conversational workspace inside a trial. It captures instructions, file reads and edits, terminal commands, and agent responses.

This structure lets you iterate safely. Earlier attempts remain preserved and reviewable.

The Chat Interface

The center of the workspace is a conversational interface. You describe what you want in natural language and the agent reads relevant files, proposes and implements changes, runs commands, generates diffs, and prepares pull requests.

The input bar at the bottom contains several controls:

• Model selector: choose the model for this session. Opens the same model browser as Project Settings. Selecting a text-only model will convert any attached images or documents to text. See Models for the full model list.
• Effort selector: controls how much reasoning effort the agent applies. Three options are available: Default Behavior, High Effort, and Max Effort. Higher effort increases depth of reasoning but may increase response time and token usage.
• Enhance Prompt: improves your instruction before sending. Adds technical clarity, structures requirements, and expands vague requests into precise implementation steps. The enhanced version is shown for review before submission.
• Scan & Review: triggers a structured quality review of all changes made in the session. See Scan & Review below.
• Attach: opens a file picker to attach files from your local machine.

Two shortcuts are available directly in the input field:

• Type @ to open the repository file picker. Search and select files to attach them as context for the current message.
• Type ! to send a shell command directly to the terminal without switching panels.

Press Shift + Enter to add a newline. Press Enter or the send button to submit.

Effort Levels

The effort selector lets you control how deeply the agent reasons before responding:

• Default Behavior: standard reasoning, suitable for most tasks.
• High Effort: increased reasoning depth. Useful for complex multi-file changes or tasks requiring careful analysis.
• Max Effort: maximum reasoning. Best reserved for critical tasks, architectural decisions, or situations where accuracy is more important than speed.

Scan & Review

The Scan & Review button triggers a systematic review of all changes made in the session before you mark work complete or create a pull request. It applies a structured internal prompting framework that checks for consistency with project architecture, potential security issues, edge cases, and production readiness.

At completion it provides a summary of files reviewed, issues found by severity, and whether the changes are ready to merge.

Attaching Context

You can provide context in several ways:

Repository files: type @ in the input field to open the file and folder picker. Navigate with arrow keys, press Enter to select, and Esc to close. Selected files are attached to the current message as context.

Local file upload: use the attach button to upload files from your machine, including PRDs, design documents, API specs, logs, or screenshots.

Context panel: the right panel's Context view lets you attach project documents and recent chat sessions as persistent context. See Context Panel below.

Context Panel

The Context panel in the right panel area gives you control over what the agent reasons with.

It is divided into two sections:

Project Documents: lists the documents from Project Context. Each document shows its name and type (for example, Product Requirements or Documentation). Click + Add to Context to include a document in the current session.

Recent Chats: shows previous sessions from the project, including Code Mode sessions and AI Task Discovery runs, with timestamps and project names. Click + Add to pull a past session into the current context. This is useful when continuing work from a previous session or providing the agent with relevant prior reasoning.

Sub-tasks

The sub-tasks panel shows a checklist of smaller steps associated with the current task. It displays a completion counter (for example, Tasks 1/3) and lets you:

• Check off completed sub-tasks
• Add new sub-tasks using the + Add a sub-task input at the bottom

Sub-tasks help break down larger tasks into trackable steps without creating separate top-level tasks.

Linking a Task

From within the workspace, you can link the current trial to a specific task using the task selector panel. It shows all available tasks for the project with their status and priority, and a search field to find a task quickly. Linking a trial to a task keeps execution traceable back to the work item that initiated it.

Terminal

Each trial includes a terminal running inside the sandbox. Access it from the right panel icon bar.

You can open multiple terminal tabs using the + button. From the terminal you can:

• Run development servers
• Execute test suites
• Install dependencies
• Run build scripts
• Execute migrations

All output is captured in the session. You can also send commands directly from the chat input by typing ! before the command.

Changes Panel

The Changes panel tracks all file modifications made during the current session. The panel icon in the right panel icon bar shows a badge with the number of changed files.

The panel header shows the total files changed and a combined additions and removals count (for example, 4 files changed +9 -12). Each file in the list shows its name, path, and a per-file breakdown of additions and removals.

Click the expand icon to open the full Session Changes diff view.

The diff view shows all modified files in a left sidebar. Selecting a file loads its diff in the main area. A file navigation control (for example, 1/4) lets you move between changed files without returning to the sidebar.

Each file header shows the filename, a Modified badge, and the full file path. The diff can be viewed in Unified or Split mode using the toggle in the top right. Removed lines appear in red and added lines in green. Unchanged regions are collapsed and can be expanded inline.

Nothing is committed automatically. You review all changes here before creating a pull request.

Running Services & Port Forwarding

The Running Services panel shows any active sandbox services. When a development server is running, it appears here and can be opened in a browser tab.

If no services are detected automatically, you can forward a port manually using the Forward a port section. Select the protocol (HTTP or HTTPS), enter the port number, and click Open to access the service.

MCP Servers

The MCPS panel shows connected Model Context Protocol servers. The panel header shows how many are currently enabled (for example, "3 of 3 enabled").

Each MCP server can be toggled on or off individually. Connected servers show a green dot. You can search the list and add custom servers using + Add MCP server.

The Marketplace tab lists available MCP servers you can install:

• Websearch (Exa): real-time web search via Exa MCP. An optional API key enables authenticated access.
• Context7: fetches up-to-date official library documentation.
• Grep.app: searches public GitHub repositories.
• Vercel: Vercel platform integration.

Installed servers show an Installed badge. Click Add to install any server from the marketplace.

Creating a Pull Request

Once changes are ready, click Create PR in the status bar to open the pull request modal.

The modal shows:

• The repository name and branch direction (for example, main ← trial-96acac35)
• Title field, type a title manually or click Generate with AI to auto-generate one based on the changes
• Description field, supports Markdown. Pre-populated with a ## Summary and ## Testing structure to guide what to include.
• Create as draft checkbox, submit the PR as a draft if it is not ready for review

Click Create Pull Request to submit, or Cancel to return. If no changes are detected in the trial, a warning appears before submission.