iterative-retrieval

To find relevant files in an unfamiliar codebase, use a structured dispatch-evaluate-refine loop that executes targeted keyword searches and glob patterns, scoring file relevance to narrow down results systematically.

The best way to search a large codebase for bug fixing is to analyze naming conventions and import chains iteratively, using relevance scoring to prioritize files based on their functional importance rather than performing blind searches.

The dispatch-evaluate-refine loop works by dispatching targeted keyword searches, evaluating file relevance on a 0-1 scale, and refining search criteria based on discovered imports and naming conventions to maintain high-precision context retrieval.

Yes, you need access to standard grep and glob-based file system traversal tools to perform iterative codebase retrieval, which are required to execute targeted searches and maintain high-precision context retrieval.

Agents get overwhelmed by irrelevant files due to the context-window dilemma where they either lack sufficient information or are flooded with files, which systematic iterative retrieval solves by scoring and refining search criteria.

You should use systematic context retrieval instead of a blind search when working in large or unfamiliar repositories, as it guides agents to discover correct files by analyzing import chains and naming conventions iteratively.