Qwen3.6-27B hits small-bug problems as a codebase grows
A small hand-written chatbot grew much larger after was used to add many quickly. The early results felt strong, but the larger started to collect many tiny bugs.
Some mistakes were basic enough that a junior developer might have caught them, so manual review and fixing became necessary. The earlier asked the model to read the current project before each feature or bug fix.
Reading the project used about half of a 128K token context, and the conversation was compacted once context use reached about 80%. The then shifted to a new conversation for each change, with the model reading only exact or line ranges, confirming specific bugs, and fixing them in the requested way.
Key points
- helped add quickly, but the became much larger.
- The larger developed many small bugs that needed manual review.
- Reading the whole project used about half of a 128K token context.
- The conversation was compacted when context use reached about 80%.
- A narrower used fresh conversations and focused only on exact or line ranges.