The code tool is not quite ready for beta release, however, we have working demos for chat to architecture (determining the structure of a project and automatically generating all of the pre-requisite files) and Fully or Semi automated generation of each file for the project working, we were even able to one-shot a working pong game:
Most of the testing/troubleshooting time this week went to making sure that changes to the description, pseudocode, links, and raw code cause the correct cascade of updates to the PML files representing the architecture. This is working reliably now in the lab, so we'll move on to finishing the tool.
The features that still need to be completed are:
- Automated generation of PML from an existing codebase without making changes to the code to trigger the workflow
- "Troubleshoot" tool that allows you to pop open an analysis pane next to your code and discuss error messages, design changes, etc with an AI assistant
- "Recon" tool that determines which files are most likely involved in an error message or glitch you're seeing and prompts you to add them to the troubleshooting tool
I expect to have the full code tool in the hands of at least a few beta testers by the end of next week.
Bug Fixes
I got a surprising amount of cleanup done this week including:
- removing unused fields from the ACL Rule editor
- fixing the cancel button on the ACL Rule editor
- fixing the delete all button/backend functionality for digests
- general mobile-friendly CSS adjustments in several pages
Known Issues
- Long running requests through the ingest tool return 500 errors instead of a "check digests for long running requests" warning (will fix this week if time allows, otherwise first sprint after code tool is done)
- Comma-separating multiple tags in the ingest tool does not work (will fix next sprint - sorry!) you can still manually add tags in the digests page as a workaround
- Video ingestion does not work in Qwen mode because of a relatively simple tokenizer mistake I didn't catch on the backend - this should also be fixed this or next sprint depending upon workload for the code tool
Infrastructure Issues
- Only one infrastructure issue this week, a GPU node experienced a driver crash and needed to be restarted, health checks dropped this node's endpoints out of the cluster, there was no impact to users.
That's all for this week - Happy testing!
Member discussion: