iterative-retrieval

You retrieve codebase context through a structured four-phase loop: DISPATCH, EVALUATE, REFINE, and LOOP. This progressively narrows the search space and applies relevance scoring to determine when subagents have sufficient context for their task.

Iterative context refinement prevents context explosion by capping retrieval cycles and applying relevance scoring. The approach stops fetching once sufficient high-relevance files are identified, avoiding over-fetch in code exploration and debugging scenarios.

Start with broad file patterns and iteratively refine your search criteria. The pipeline dispatches an initial search, evaluates the relevance of returned files, refines the search parameters, and loops until enough high-relevance files satisfy the task.

Yes, iterative retrieval supports debugging and feature implementation with context-limited agents. The multi-phase loop progressively reveals unknown codebase context, ensuring subagents receive only the relevant files needed without exceeding their context limits.

Avoid iterative retrieval when the required codebase context is fully predictable upfront. The four-phase DISPATCH, EVALUATE, REFINE, and LOOP cycle adds overhead, making it unnecessary for tasks where context is already known and does not require progressive narrowing.