What problem does it solve? Coding agents forget what they learn about a codebase between sessions, and they cannot distinguish relationships they merely observed in code from ones verified as correct. This Skill installs a persistent SQLite-backed world model that remembers entities, interactions, and constraints across sessions, tracks observed-vs-normative confidence for each fact, and detects contradictions with proposed fixes. ## Core Features & Use Cases - Persistent codebase memory: Records symbols, files, modules, and external referents plus their interactions in a per-project SQLite store, auto-captured by four Claude Code lifecycle hooks with zero manual effort. - Two-axis confidence tracking: Separates observed confidence (we saw this in code) from normative confidence (an oracle like a test, CI run, doc, or human verified it), so observed-but-unverified facts are flagged instead of trusted. - Contradiction detection and fixes: Evaluates constraints each turn, opens contradiction rows with located, human-readable proposed fixes, and converts contradicted mass to validated mass as the code is corrected. - Ontology-guarded writes: Every triple is validated against a closed predicate vocabulary with domain/range rules before insert, rejecting hallucinated verbs and impossible pairings. - Use Case: An agent editing services/report.py is warned before the edit that it imports a deprecated config module, because a prior session recorded that constraint — knowledge not visible in the file itself. ## Quick Start Ask the agent to install the world-model-ledger skill into this project and restart Claude Code so the lifecycle hooks begin capturing codebase knowledge automatically.
