dsh-fovea

Working across hundreds of files inside a large repository, an agent is caught between two poor options: stuffing the entire workspace into context (expensive and noisy) or repeatedly grepping text fragments that carry no awareness of symbol-level causality and structure. dsh-fovea reframes the problem in two steps. First it compiles the repository into a typed, weighted graph. Then it turns a question or change set into an interest field that diffuses relevance through that graph, so a bounded context budget lands precisely where the agent is active while the rest of the workspace stays as a low-cost structural skeleton.

Four model-facing tools do the heavy lifting. fovea_sketch surveys feature basins, routes, and hubs in an unfamiliar codebase. fovea_focus centers on a symbol, file, or concept and reveals its semantic neighborhood. fovea_dwell widens the current lens without re-exposing the periphery already disclosed. fovea_impact ranks the blast radius and a sensible review order before a cross-file edit lands. A continuous-sync layer quietly re-baselines when repository drift is structurally surprising and appends only a sliver of context to the model when a signal is actually present, so every turn is not a full rescan. The grep.mode setting lets bare symbol queries navigate the graph natively (replace) or augments native grep results with a graph section (augment), while regex and path filters always fall back to the exact native definition.

dsh-fovea is built for teams and individuals who drive DeepSeek Harness on medium-to-large repositories and want their agent to think map-first, not grep-first. It does not displace precise text search or file reads; it answers the prior question of where to look and why, so the context budget is spent on the right lines with the right surrounding structure.