iterative-retrieval

Iterative retrieval refines repository context by running a four-stage loop of dispatch, evaluate, refine, and loop to progressively narrow down relevant files and patterns. This ensures subagents receive high-value context without noise.

Searching a large codebase for bug diagnosis works best by running up to three cycles of dispatch and refinement with broad initial keywords. The system ranks retrieved files by relevance and returns a concise set of high-value sources for evaluation.

Adaptive term discovery expands keywords and patterns based on evaluation feedback during the retrieval cycles. This uncovers domain terms and codebase conventions that initial broad search queries might miss.

You need iterative retrieval for agent workflow context when orchestrating multiple subagents across a large repository. It prevents agents from being overwhelmed by noise or missing critical details by progressively refining retrieved materials.

Run up to three cycles of dispatch and refinement to gather highly relevant codebase context. Starting with broad patterns and keywords, these cycles progressively narrow down the most relevant files for subagents.