Datamimic – don't let your coding agent invent its own test world

Datamimic is a Python-native, MIT-licensed deterministic data engine designed for generating synthetic datasets and performing PII-aware pseudonymization for regulated enterprises. The community edition provides deterministic (seeded) and privacy-maximized (non-seeded) transformations, model-driven synthetic generation without source data, and execution pipelines against PostgreSQL, MySQL, Oracle, MS SQL, SQLite, MongoDB, CSV/JSON/XML/XLSX, DbUnit and fixed-width formats. It supports complex modeling constructs - weighted state machines, composite multi-field references, control flow, and a scriptable memstore - while emitting append-only provenance logs and per-output content hashes for audit re-execution. The commercial platform layers on PII scanning, role-based governance, multi-system coordinated execution, industry message templates (EDIFACT/SWIFT/HL7/FHIR), a Rust/ML performance core, and on-prem/air-gapped deployment used in regulated EU banking.

Authoring and automation are driven by a strict CLI agent contract: preserve intent as model.dm.json, scaffold to get a bounded XML model, repair from structured diagnostics, declare explicit expectations (counts, uniqueness, ranges, foreign keys), and stop on verified=true. The project documents agent prompts and workflows, enforces single compile/lint/bounded-run transactions per attempt, and offers an optional MCP adapter that maps four canonical operations (reference, scaffold, check, dry-run). A concrete example shows seeding a relational hierarchy - four customers with exactly two orders each - using scripted foreign keys (parent.id) and per_parent_count expectations to guarantee referential integrity.