OptionalonProgress: (message: string, data?: Record<string, unknown>) => voidOptionaldriver: BaseAutoDocDriverDatabase access for checks the LLM's proposals must pass (composite PK uniqueness). Optional: without it, composite PK proposals are rejected.
Use CheckConvergence.
Execute backpropagation
Lock interim ground truth: FKs with confidence ≥ threshold become immutable. Call this AFTER the iterative analysis completes but BEFORE the pruning pass.
Lock high-confidence PK candidates as interim ground truth.
Perform cross-schema sanity check Validates consistency across all schemas
Perform dependency-level sanity check Checks consistency across tables at the same dependency level
Perform schema-level sanity check Holistic review after entire schema is analyzed
ProtectedpersistPersist enum/value-list verdicts from LLM response onto ColumnDefinition. Keeps the highest-confidence verdict across multiple iterations.
Use ProcessLevel.
Process a single dependency level
Use PruneForeignKeys.
Two-pass FK pruning using a potentially stronger model. Pass 1: Per-table — evaluate each table's unlocked FKs, propose removals. Pass 2: Holistic — review all proposed removals at once for final decision. Locked FKs (interim ground truth) are never touched.
Use PrunePrimaryKeys.
Two-pass PK pruning using a potentially stronger model.
Supply the probe used to verify LLM-proposed foreign keys. Injected rather than constructed here because the engine has no driver of its own; the orchestrator owns the connection and passes ONE verifier so the per-run probe budget is shared with the organic-key path instead of each path getting its own allowance.
Exports for core module