db-connector

In the DeepSeek Harness (dsh) framework, agents that touch databases directly often lack meaningful safety boundaries — writes go unapproved, injection risk is unmanaged, queries run without time limits, and operations leave no audit trail. db-connector closes that gap by wrapping SQLite, PostgreSQL, and MySQL behind five model-facing tools, so agents can access structured data under strict, auditable policies without hand-rolling connection management, transaction control, or logging pipelines.

The design centers on read-first access, explicit write approval, and full audit coverage. db_query permits only SELECT-class statements and rejects everything else before a driver is invoked. db_exec demands an allowWrite confirmation and executes inside a transaction with automatic rollback on failure. db_schema provides TTL-cached introspection of tables, columns, indexes, and foreign keys. db_audit appends every call — denials and errors included — as a JSONL record. Parameters are always bound, credentials are referenced by name only and never written to logs, reducing injection and leakage surfaces at the source.

Ideal for dsh workflows where agents need to read from or cautiously write to a database: report generation, data inspection, ops triage. If you want your model to collaborate with three mainstream databases inside controlled boundaries — with security defaults, a clear tool contract, and a durable operation trail — db-connector is a ready-to-use layer that eliminates the need to build that infrastructure from scratch.