AI code review of SD source code

[Big, River Guy]

In case anyone is interested:

I had AI  (Cursor/Anthropic Claude) do a review of all the C , Basic and Bash code that makes up SD.  The resullt is in the repository sdscripts_ai.

I did make a few changes after the review to the installsdai.sh script  changing the install methodology. When you download this repository you get the scripts and the source code in one bundle.  After your manual download, the install script always uses the local source code for installation.  

The results of the code review are stored in the AI_Modification_Notes folder.  Under that there are separate folders: Bash_Scripts, Basic_Code and C_Code.  Each of those directories includes two files ChangesApplied.txt and UnsafeProposed.txt.

There were three separate review passes for each group of code: Basic, Bash and C.  The first pass was the code review., the second was to make safe changes, the third was try to fix the remaining issues.. The second pass (safe) changes are in the ChangesApplied.txt file. .  The result of the third pass is in the UnsafeProposed.txt file along with the status of each suggested change.  The result can be RESOLVED, MOSTLY RESOLVED, PARTIALLY RESOLVED, DEFERRED or UNCHANGED.

WARNING: this process was experimental.  While the system does compile, install and run, additional testing has been minimal.  It can not be run along side a normal SD install.  Use a separate virtual machine or computer.

The changes files are fascinating. Some of the fixes repair rather serious bugs. Others are more cosmetic.  For example, in the C code there were 208 safe changes made, and nine changes that might require human review.  In the Basic code especially, there are several files that the AI would not touch because of their complexity and the fact that they are core to the Basic compiler and other deep system level functions.

To download do "sudo git clone  https://www.codeberg.org/stringdatabase/sdscripts_ai". 

NOTE: This was my first time using AI to do code reviews and modifications.  I did learn a lot and might make some changes to my methodology should I try this again.

If you do experiment with this system, let us know your experiences in this thread.

[Steven Martin Trimble]

thanks for this!

CDMI

Steven Trimble

(501) 772-3450 cell/text

cdm...@gmail.com

--
You received this message because you are subscribed to
the "Pick and MultiValue Databases" group.
To post, email to: mvd...@googlegroups.com
To unsubscribe, email to: mvdbms+un...@googlegroups.com
For more options, visit http://groups.google.com/group/mvdbms
---
You received this message because you are subscribed to the Google Groups "Pick and MultiValue Databases" group.
To unsubscribe from this group and stop receiving emails from it, send an email to mvdbms+un...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/mvdbms/91faaa31-367c-4533-973e-1c98e480b5f2n%40googlegroups.com.

[Big, River Guy]

I have switched to an AI agent called Kun and the Deekseek AI.  It actually seems as good as Claude and is much cheaper than Cursor. You do pay for API calls, but even the best Pro model that they have, I would have to really work to spend more than $10 per month.

An the SD-AI side of things.  I have changed how the install and remove process works for now.  It doesn't try to download anything, just works locally.  If someone wanted to upgrade an existing system they would have to clone the repository, rather than depend on the script to do the download.  The AI has made many improvements to the install script. For example, a user no longer has to reboot to update users and groups. The user can install, remove, install, remove over and over without ever leaving a terminal session.  In addition the changes make it so that that SD is immediately available for use after each install.

Continuing code reviews on the SD basic code.  Deepseek is identifying issues that Claude did not.  It is automatically updating the change log files that were started by Claude.  So far what I have found is that Deepseek is very good at responding and finding ways to implement ideas provided to it.  It does not volunteer next steps like some other AI's do. However, it will make suggestions if you ask.