What problem does it solve? Subagents in multi-agent workflows are spawned with limited context and cannot predict which files or patterns they need upfront, leading to context overflow, missing information, or failed tasks. ## Core Features & Use Cases - Four-Phase Retrieval Loop: Implements DISPATCH, EVALUATE, REFINE, and LOOP phases that progressively narrow search criteria over a maximum of three cycles. - Relevance Scoring: Scores candidate files on a 0-1 scale and identifies missing context gaps to drive the next query refinement. - Terminology Discovery: Learns codebase-specific naming conventions (e.g., discovering "throttle" instead of "rate limit") to improve subsequent searches. - Use Case: When fixing an authentication token expiry bug, the loop starts with broad keyword searches, evaluates relevance of auth.ts and tokens.ts, refines with discovered terms like "jwt" and "refresh", and returns only high-relevance files. ## Quick Start Use the iterative-retrieval pattern to gather the right codebase context for this task, refining the search over up to three cycles.
