Connection strings are not change control
Credentials open the door. Receipts govern the write.
SQLGuard field notes
Notes from the production write boundary.
Practical distinctions we keep finding while building permission controls for database agents. Mechanisms, evidence, and limits—from one engineering team to another.
Written by SQLGuard·hello@sqlguard.io·Follow by RSS·12 notes
Human-in-the-loop is not a receipt
A chat yes is a moment. A statement-bound receipt is evidence.
Dry-run is not permission
Sandbox and EXPLAIN are taste. Production execute still needs authorize.
Audit logs are after the write
Logs explain what ran. Receipts decide whether it was allowed.
RLS is not statement authorize
Tenant isolation and statement permission solve different failures.
CI gate is not production authorize
A green PR check is a rehearsal. Runtime still needs PASS + verify.
Readonly flags are not receipts
mode=readonly is a kill switch. The intentional write still needs a decision.
Which agent_id authorized this?
The pager question after an agent mutates production.
Access mode is not authorize
Unrestricted / write modes set capability. Permission binds one statement.
Client confirm does not travel
Local [y/N] helps one client. Other hosts need a statement-bound receipt.
Why Probe is free
Risk guidance is useful before a write, but guidance is not permission.
Authorize is not lint
Lint inspects a statement. A receipt answers who permitted the write.