Agent System

SmartCursorX features a pluggable agent system with cognitive steering, taxonomy-aware context assembly, and fine-grained tool permissions.

Architecture

The agent system follows a plan-execute-review loop:

  1. Plan: The AI model analyzes the user request and generates a structured plan with file-level modifications
  2. Execute: Each modification is presented in a side-by-side diff editor for user review
  3. Review: Users accept or reject individual file changes before they are written to disk

Cognitive Steering

The Taxonomy Engine classifies the development context to optimize model performance:

Example

When working on a Python Django project, the steering engine:

  1. Sets the taxonomy domain to python.django
  2. Prioritizes urls.py, models.py, views.py, settings.py in search results
  3. Includes Django-specific conventions in the system prompt
  4. Restricts file operations to the project's virtual environment scope

Tool Permissions

The agent has access to a curated set of tools, each with workspace containment:

ToolDescriptionContainment
search_filesRegex search within workspace filesWorkspace root
read_filesRead file contentsWorkspace root
edit_fileApply edits to filesWorkspace root
write_fileCreate new filesWorkspace root
bashExecute shell commandsConfigurable

All file operations validate paths against the active workspace to prevent directory traversal attacks.

Context Assembly

The agent assembles context through multiple channels:

Multi-Profile Terminal

The integrated terminal supports multiple shell profiles:

Process state is cleaned up on component unmount to prevent orphaned processes.

Abort & Error Handling