Qwen Code Documentation

The Virtual Viewport implementation (introduced in PR #4146) addresses performance issues (flickering, lag, and unbounded scrollbar growth) caused by rendering entire conversation histories in ink's append-only <Static> component.

Key Features included in V.2:

• Core Viewport: Efficiently renders only visible items.
• ASCII Scrollbar: Includes auto-hide animations.
• SGR Mouse Support: Enables mouse-wheel interaction.
• Keyboard Navigation: Supports scroll keys.
• ui.useTerminalBuffer Gate: Allows switching between the legacy <Static> path and the new ScrollableList path.

Planned Features (V.3+):

• Scrollbar dragging and click-to-position.
• In-app / search.
• Alternate-buffer mode (full alt-screen takeover).

The acp-sdk is a Java implementation of the Agent Client Protocol (ACP), designed to facilitate standardized communication between AI agents and client applications. It enables developers to implement features such as session management, file system operations, terminal command execution, and tool call handling within Java applications.

Key Capabilities:

• Session Management: Create, load, and manage conversation sessions.
• File System Operations: Read and write text files.
• Terminal Access: Execute commands via the terminal.
• Tool Integration: Handle tool calls and manage permissions for sensitive operations.
• Rich Content Support: Process text, images, audio, and resources.
• MCP Integration: Works with Model Context Protocol (MCP) servers.

Qwen Code hooks allow you to extend and customize application behavior by executing custom scripts or programs at specific lifecycle points (e.g., before/after tool execution, session start/end). Hooks are enabled by default. You can globally disable all hooks by setting disableAllHooks: true in your settings file.

Common use cases include:

• Monitoring and auditing tool usage
• Enforcing security policies
• Injecting context into conversations
• Integrating with external systems

{
  "disableAllHooks": true,
  "hooks": {
    "PreToolUse": [...]
  }
}

The agent uses the cua-driver to interact with your desktop. It is recommended to use element-addressed actions (targeting element_index from the accessibility tree) rather than raw pixel coordinates for better stability.

Key Tool Categories:

• Mouse: click, double_click, right_click, drag, move_cursor, scroll
• Keyboard: type_text, press_key, hotkey
• Windows / UI: list_windows, get_window_state, get_accessibility_tree, set_value, zoom
• Apps: launch_app, list_apps, bring_to_front, kill_app (Note: bring_to_front is Windows-only)
• Browser: page (execute JS, read text, query DOM, click elements)
• Screenshots: get_window_state (captures PNG), page
• Recording: start_recording, stop_recording, replay_trajectory
• Sessions: start_session, end_session

In Qwen-Code, Artifacts are explicitly registered, structured outputs within a session that are reusable, clickable, previewable, or downloadable.

Key distinctions:

• Artifacts: Explicitly declared outputs like generated reports, HTML files, PDFs, images, or specific business resource URLs (e.g., a dashboard link from a tool).
• Not Artifacts: Standard source code changes (these are file changes/diffs), markdown links in assistant responses, URLs found in shell stdout, or URLs from web_fetch (these are considered noise or inputs).

AutoSkill is a mechanism for automatically extracting procedural memory from agent sessions. When an agent completes a task involving high tool-call density, the system evaluates the session for reusable operational workflows and saves them as project-level skills.

Key Differences from Memory Extract

The /review skill uses an ensemble of specialized agents to perform code reviews. To prevent human-reported blockers from being ignored, the system uses a semantic recognition pattern called carriesBlockerSignal rather than relying on a literal [Critical] marker.

Any body (inline thread or issue comment) asserting a blocking defect—using phrases like [Critical], (blocker), is a blocker, must fix, still reproducible, broken, fails, or the Chinese 阻塞项/问题/点—is automatically promoted to a "Blockers to re-check" section. This section is rendered in full to ensure agents can see the complete context, preventing blockers from being lost due to character limits or snippet truncation.