Automated Debugging
Debugging loop:
‘generate’ and test\
Creation of the baseline reproduction scenario; and then iterative hypothesis and test.
-
requires a fixed testable defect reproduction: Case A - ie a spec/description for the senario when done that causes the defect to show up. Case B - a spec/description for the scenario where the defect does not show up.
-
The LLM by default shoul be able to keep repeating and experimenting on Case A; ‘B’ becomes the case.
This is quite similar to how humans debug: they reproduce the erroneous scenario (Case A). Figure out what the correct scenario would look like. And then come up with reasons on why Case A might be happenin and how to resolve it to Case B, and then try various hypothesis/ideas until B becomes the C.
Reasoning/Grounding
For the LLM to become aware of the architecte, several approaches can be used:
- Specific skill files which ‘teach’ the LLM about the connector.
- A pre-computed typed knowledge graph of the architecture of the plugin, stored in the DB or a relational format.
- Other structured knowledge such as DB tables with useful information.
- Others … ?