CodeNib

Coding agents repeatedly search, navigate, and retain context from repositories that keep changing underneath them. In practice that context comes from disconnected pieces — a lexical index here, a language server there, a task-local scratchpad that dies with the session. The result is repeated discovery work and lifecycle costs nobody measures.

CodeNib compiles a repository into lexical, dense, and structural views keyed per commit, maps every output back to repository-relative source ranges, and repairs those views incrementally as edits land instead of rebuilding from scratch. One runtime then serves ranked search, symbol navigation, and bounded context to agents over MCP and LSP-shaped APIs.

Across 100 repository snapshots:

• graph and vector updates are 8.7× and 25.4× faster than independent rebuilds that produce matching output
• static navigation requests that match live-server locations see a 4.7× median latency improvement
• context selection policies cut trajectory tokens by 50–87% versus grep/read baselines

The project also ships a local Wiki UI, an Ask view, and a dependency map so a human can inspect exactly what the agent would have been served.

pip install codenib
codenib wiki /path/to/repository

Developer preview, Apache 2.0. Built under SysEvol.