Experiment · EXP-002

Can the context window stand in for the repository?

Observational study of coding agents on codebases larger than the window.

I asked whether a coding agent’s usefulness on large repositories comes from stuffing more files into context, or from indexing, retrieval, and iterative discovery. Public systems and my own use of coding agents point to the second. Loading “the repo” is not an architecture.

Hypothesis: If the model could see more of the repository, it would know where to change authentication.

Research question

When a coding agent works inside a repository larger than its context window, does usefulness come from seeing more of the repository at once, or from repeatedly discovering the right parts of it?

Observation

In real coding-agent use, the repository is not the context window. Useful agents search, open files, follow symbols, inspect references, run commands, observe failures, and retrieve additional context as the task develops. Their working context changes over time.

From repository to model

Repository
  → Chunks
  → Representations
  → Retrieval
  → Selected context
  → Model

Cursor publicly describes splitting code into syntactic chunks and producing embeddings that later support semantic retrieval. Instead of sending 300,000 lines to the model, the system can ask which pieces of the repository are most likely relevant to what the developer just asked. Indexing and retrieval are what make a large repository usable to a model.

Not: memorise everything. But: know how to find the next useful thing.

Why more context is not enough

A request such as 'change how authentication works' may have obvious semantic matches in a login component while the behaviour is actually controlled by middleware, a policy module, an API handler, or an imported utility. Loading more related files can increase token usage without revealing the controlling code.

Result

The stronger mental model is iterative context discovery: index the repository, retrieve likely candidates, inspect structure, follow relationships, execute tools, and update context from evidence. The task is not to fit the codebase into the window. The task is to repeatedly assemble the smallest useful context for the next decision.

Implication

For coding agents, retrieval quality, symbol relationships, tool access, and feedback loops can contribute as much to usefulness as raw context-window size. A larger window is helpful, but it is not a substitute for repository navigation.

Limitations

This is an observational architecture study based on public system behaviour and my own use of coding agents. I have not published a controlled benchmark that isolates context-window size from retrieval and tooling.

Abubakar, Imam. "Can the context window stand in for the repository?." Imam Abubakar AI Research, 2026. https://imamabubakar.com/research/exp-002-repository-in-the-window

This archive contains independent research, experiments, architecture studies, and technical observations. Some findings are exploratory and should not be interpreted as peer-reviewed scientific conclusions unless explicitly stated. Architecture studies of third-party products are based on publicly observable behaviour, published technical information, and engineering analysis. They are not claims of access to private source code or infrastructure.