iterative-retrieval

Iterative context retrieval progressively narrows down relevant files for subagents through a four-phase loop of dispatching, evaluating, refining, and looping. This structured method selectively includes high-value material, reducing waste while preserving essential information needed for accurate decisions.

The best way to gather candidate files for bug diagnosis is using an initial broad query to collect candidates, evaluating relevance scores, and refining search conditions. This iterative query refinement repeats up to a configurable limit, returning a concise, high-signal context set.

Context scoring evaluates retrieved files for relevance during the EVALUATE phase, enabling selective inclusion of high-value material. Files are scored to identify relevant patterns and terms, allowing the system to prune irrelevant ones and assemble a concise context for the subagent.

Yes, you can configure the maximum number of retrieval cycles for subagent context refinement. The system enforces a controlled cycle limit, defaulting to three iterations, to prevent overfitting or endless search while progressively narrowing down relevant context.

Subagents make poor decisions because standard approaches like sending everything or nothing overwhelm the context window or leave critical details missing. Applying an iterative retrieval loop with high-relevance scoring incrementally refines knowledge, ensuring essential information is preserved.

You should avoid iterative query refinement when a single broad query sufficiently captures necessary context, or if the codebase is small enough that overfitting could occur. The controlled cycle limit prevents endless search, but simple tasks may not require progressive refinement.