Fitting a 300k-Line Repository into a 200k-Token Window

Your coding agent has a 200,000-token context window. The repository it works in has ~300,000 lines of code, and a typical task takes 30–60 tool calls. In early testing, sessions routinely hit the limit halfway through a task, and answer quality degrades well before that.

1. Explain what is filling the window and why quality drops before the hard limit.
2. Design the context-management strategy: what goes into the prompt, in what order, and the techniques you use to keep usage low.
3. Describe how you would implement compaction so that it does not destroy the information the agent needs to finish the task.