wright-index
A code intelligence engine for AI coding agents. It parses a repository once — symbols, call graph, git-derived coupling — and serves it as queryable tools, so agents stop rediscovering the same structure every session.
- ~4,000 LOC Python, 57 tests — indexes repositories into symbols, a call graph, and co-change coupling via tree-sitter across Python, Go, and TypeScript, requiring no compiler or build environment.
- A confidence-graded call resolver: every edge records how it was proven — receiver type, package scope, import trace, repo-unique name — and unprovable targets are stored as unresolved rather than guessed, keeping every downstream query trustworthy.
- Association-rule mining over git history (support / confidence / lift), surfacing files that change together 17–25× beyond chance despite sharing no import or symbol.
- An MCP server (8 token-bounded tools) queried directly by Claude Code — measured 33% fewer agent turns and up to 2.3× fewer tokens to the same answers; incremental reindex in 0.64s on a 69k-line repo via content-hash diff and symmetric-difference invalidation.
- The open-source spillover: a merged fix in CNCF HAMi's GPU admission webhook, an O(N²) lock-contention benchmark + fix in HAMi-core (C/CUDA), and a scope-tree fix in DeepSource Globstar.