data-structure-protocol

LLM coding agents lose context between tasks. On large codebases they spend most of their tokens on "orientation" — figuring out where things live, what depends on what, and what is safe to change. DSP solves this by externalizing the project's structural map into a persistent, queryable graph stored in a .dsp/ directory next to the code.

DSP is NOT documentation for humans and NOT an AST dump. It captures three things: meaning (why an entity exists), boundaries (what it imports and exposes), and reasons (why each connection exists). This is enough for an agent to navigate, refactor, and generate code without loading the entire source tree into the context window.

DSP models the codebase as a directed graph. Nodes are entities, edges are imports and shared/exports.