ROADMAP.md

CPE (Chat-based Programming Editor) Roadmap

This document outlines the planned features, improvements, and notes for the CPE project. Items are organized by priority and status.

Features

Agentic flow

• Move from disparate mulit-agent to single-agent, will reduce necessary calls, as we can remove the needs codebase function call
• Experiment with idea of sub agent creation on the fly?
  • Models like Gemini such at editing files with function calling. Maybe just have the model utilize the bash tool to give specific edit instructions to a file-editing function calling capable model? like gpt-4o-mini?

Token Analysis and Visualization

• Implement basic token counting per file
• Show total tokens for each directory (tree)
• Model specific token counting
  • Use anthropic token counting endpoint
  • Default to using gpt-4o token counting

Code Map

• Some repos are too large to fit in a context window, even without function bodies, so we should process the codebase in chunks if it exceeds the context window
  • Or maybe we can create further levels of fidelity hierarchy. Maybe something like:
    1. File paths (for extremely large repos, or maybe just in arbitrary directories?)
    2. Only type, function and method names (and class names for object-oriented langs)
    3. Code comments, global variables, full type definitions (like all struct fields, tags, field comments, etc.) (this is the current lowest fidelity level today)
    4. Full file contents
• Instead of detecting file extensions, try to detect if file of text context using magic bytes
• Include code from dependencies if possible in file overview
• Use model token counting to return an error to the model if given level of file overview will exceed to context window (or maybe given threshold? 75% of context window to allow for some room?)

Getting related files

• Transition from using tree sitter for go to using native ast pkg in stdlib
• Explore using python in wasm runtime to resolve related files (RustPython?)
• Explore using java in wasm runtime to resolve related files

Tooling

• Add support for bash execution tool

LLM Integration

• Support for more LLM providers
  • Mistral
  • Deepseek
  • Nous
• support multimodality
  • images
  • videos
• Use official sdks instead for openai, gemini
  • openai
  • gemini
  • anthropic
• Use structured outputs for openai and gemini to ensure strict following of tool schemas.

User Experience

• Command auto-completion
• Add support for continuing a conversation if user chooses to do so
• Support sending requests to multiple models and picking the best one
• Support sending requests to multiple models and picking the best one

Performance

• Parallel processing for large codebases

Documentation

• Comprehensive user guide
• Example use cases and tutorials
• Contributing guidelines
• Architecture documentation

Goals

• Test against extremely large mono-lang codebases
  • cert-manager
  • kubernetes
  • SWE bench?
• Test against extremely large multi-lang codebases
  • ???