What problem does it solve? String-assertion tests on generated templates cannot prove that scaffolded Nodefony code actually compiles, runs, or is discoverable by AI agents. This Skill provides three complementary benches that measure what the scaffold truly produces: whether generated code stands up (compilation, tests, real HTTP), whether an agent dropped into a fresh app finds the tooling instead of guessing, and whether a real-world data model is expressible with the field grammar. ## Core Features & Use Cases - Truth bench (verify-generated.mjs): Builds an isolated witness application from tarballs and exercises generation, compilation, lint, real HTTP (201/422/409/pagination), production mode, and multi-database dialects (SQLite, PostgreSQL, MySQL/MariaDB). - Discoverability bench (bench-discoverability.mjs): Runs real agents (Claude, Vibe, Codex, Gemini) through frozen tasks in a reset decor, judging them with objective probes and gates rather than LLM judges, and measures turns, duration, and cost. - Transcript analysis (analyse-transcript.mjs): Parses transcripts from six agent dialects to report turns, cost, MCP calls, failed gestures, and full timelines. - Use Case: After modifying a Nodefony generator template, run the truth bench to confirm the generated entity compiles and its REST routes respond correctly before declaring the change done. ## Quick Start Run the devkit bench to verify that the Nodefony scaffold still generates code that compiles and passes real HTTP checks.
