1. A new approach Using ASTs to reliably merge LLM generated code in to existing codebases. Line diffing approaches are sooo 2023

2. A little about me I was really tired of manually copying and pasting snippets from chat GPT or other tools in to my code editor. I tested other tools and they seem to be a little too eager to delete my code or mangle my source files. (mmiscool) on github. I make seagulls a bit cross.

3. The project https://github.com/mmiscool/aiCoder A tool for making changes to existing code files reliably using LLMs.

4. The normal workflow today You know you do it. Copy Copy and paste existing code file in to ChatGPT conversation Paste Manually copy code snippets one at a time from LLM response and paste them in to the correct location in source code file Lorem ipsum Write instructions about the desired changes or new code to be generated and submit along with the existing code.

5. Ideal workflow. Select target file AST merge tool: Test snippet is syntactically correct. Describe desired modifications or new functionality to be added. Surgically replace duplicate nodes in AST. Discard stubs if existing function or method has code. Extract snippets for merging from LLM response Regenerate codefile with modifications made.

6. File being edited. Code snippet produced by LLM

7. Merging the code together using an AST File being edited. Code snippet produced by LLM Classes and methods that are duplicated need to be merged. The existing method is replaced with the new method in the snippet.

8. Results. Existing file modified to replaced only the method that changed.

9. Rules we are using when merging LLM snippets in to existing files. ● Empty stub methods or functions in snippets only get added to existing classes if they do not currently exist in the target. ● If target contains non stub function or method it is not replaced with an empty one. ● Existing location of class or function in file is not disturbed. Example implementation for javascript files: src/intelligentMerge.js

10. LLM instructions to coax snippets to be easy to merge. ● Only provide the classes and methods being modified. ● Exclude existing code. ● Provide examples to the LLM that show the format of the desired snippets. ● Instructions to avoid creating global variables and testing or example code. Full markdown file with LLM instructions for snippet production: src/prompts/snippet-production-prompt.md

11. aiCoder Chat interface ● Chat interface to edit a selected file ● Auto apply mode (Pops up a timeout to review changes and reject them otherwise it automatically apply snippets from conversation. ● Buttons on each snippet to manually apply code from that snippet to code file.

12. aiCoder Tools ● List methods in current target file. ● Mark stubs as red. ● Clicking on gree functions or methods generates a new conversation about that specific bit of code. ● Clicking a red method launches a chat to automatically implement the method. ● Merge and format. You might have some code that follows the rules for automatic merging produced by an outside LLM. Just paste this code at the bottom of the code file and hit this button to make the magic happen.

13. aiCoder Project Settings ● Change premade prompt templates for the specific project you are working on ● This can tweak the style of javascript produced. ● Provide refinement about how you want snippets formatted or any other tweaks needed to enhance the LLMs output using prompt engineering.

14. aiCoder LLM Settings ● Select the LLM provider and specific model you want to use. ● OpenAI and Anthrompic both provide best quality responses. ● Ollama provides local LLM support. Best results for javascript and snippet production compliance come from IBM granite 3.1 based models. ● Groq has a very small context window and code generation quality is poor with the older models they have available.