Repository Context

The function you are writing depends on types defined in another file, a helper three directories away, and a convention nobody wrote down but every file follows. A model shown only the current file is missing most of what it needs, and the difference between a completion that compiles and one that invents a plausible-looking method is usually somewhere else in the repository.

Nobody fits a repository into a context window, so the work is selection: which files to pull in, and in what order. Retrieval by textual similarity is the baseline and is beaten by anything that uses structure — imports, call graph, the file the cursor's symbol is defined in. Repository-level benchmarks exist because file-level ones were saturating while the tools were still visibly failing at the thing people used them for.

The share of a repository that fits in context falls as the repository grows and the window does not, and repositories grow faster. A medium project is already millions of tokens against a window of hundreds of thousands, so the fraction visible at once is a few percent and shrinking — which means selection quality, not window size, is what a completion depends on. Doubling the window moves a few percent to a few more percent.