iterative-retrieval

Progressive codebase context refinement supplies subagents with just-in-time files by running a four-phase loop: DISPATCH, EVALUATE, REFINE, and LOOP. This iteratively narrows large repositories into high-relevance file sets without overwhelming the agent's context window.

Multi-agent context management is handled by pruning irrelevant files through guardrails during an iterative retrieval loop. The loop evaluates dispatched context and refines it across up to three cycles to ensure only high-relevance code is supplied to subagents.

The iterative retrieval loop handles dynamically emerging context by cycling through evaluate and refine phases. When a subagent discovers it needs different files during execution, the loop prunes irrelevant files and dispatches newly identified high-relevance codebase context.

Iterative context refinement applies directly to multi-agent tasks like bug fixes, feature implementation, and code exploration. It orchestrates subagents by progressively assembling the specific codebase context each task requires as it emerges during execution.

The four-phase loop for codebase context management is limited to a maximum of three refinement cycles. After three cycles, the loop stops refining and pruning irrelevant files, meaning highly complex tasks requiring deeper exploration may not fully resolve their context gaps.