Commit

Shipped phpgraph 1.5.8. AI agents on a PHP codebase burn most of their tokens on grep and reading files just to learn who calls what. phpgraph hands them that as facts: a knowledge graph built by static analysis (no LLM, no network), served over MCP to Claude Code, Codex, Cursor… Same agent on a 4,800-file Symfony/DDD app, 3 runs each, with vs without: - impact / refactoring: −42% tokens, −50% time, −59% tool calls - explaining a feature: −25% tokens, −34% time, −49% tool calls Same or better answers. Against graphify on Kimai (Symfony, 1,800 files, not in its test corpus): - 20,195 caller → callee pairs vs 7,615 - callers of methods declared once: 2,407 of 2,408 found vs 902, no false link - routes and event subscribers linked to their code: 266 routes, 26 subscribers vs none - build in 11 s vs 27 s Every edge says how sure it is (EXTRACTED / INFERRED / AMBIGUOUS), and the graph lists what it could not link, so the agent knows when to go read the code. https://github.com/hbenabdallah/phpgraphhbenabdallah/phpgraph