PostgresGuard: Production Safety & Query Guard

AI agents often lack the situational awareness to distinguish between a local dev container and a live production database. A single command like prisma db push or TRUNCATE can wipe years of data if an agent misinterprets a connection string or falls back to a production DSN after context compaction.

What it does

• Destructive Operation Gate blocks commands like TRUNCATE, DROP, and DELETE without WHERE clauses.
• Migration Safety intercepts risky tool-specific commands including prisma db push, alembic downgrade, and terraform destroy.
• Target Verification independently flags any destructive action targeting non-localhost hosts, specifically cloud-managed endpoints like RDS, Supabase, and Railway.
• Security Auditing checks for missing Row Level Security (RLS) on sensitive tables and unsafe search_path configurations.
• Health Monitoring identifies table bloat (>20% dead tuples) and missing indexes on foreign keys that cause lock contention.

How it works

1. The agent intercepts any SQL DDL/DML or migration tool command before execution.
2. It parses the active connection string to extract the hostname, ignoring credentials to prevent secret leakage.
3. If the command is destructive and the host is production-adjacent, the skill halts execution and presents a confirmation gate.
4. The user must provide a specific authorization string to bypass the block.

Frameworks & tools

Works with raw PostgreSQL, Prisma, Alembic, Drizzle, and Terraform. Compatible with managed services including AWS RDS, Railway, Aiven, Supabase, and Neon.

Why this beats prompting it yourself

Manual prompts are easily ignored during context overflow or agent hallucinations. This skill provides an architectural safety rail that independently verifies connection strings and command intent, ensuring a production wipeout is never one "Yes" away.

Use cases

• Prevent an agent from accidentally resetting a live Railway database during a schema sync.
• Audit production tables for missing RLS before deploying new data access layers.
• Detect stalled autovacuum processes and high dead tuple ratios on high-traffic tables.
• Stop infrastructure-as-code agents from executing terraform destroy on RDS instances.

Known limitations

Does not inspect actual table rows for privacy. Requires read access to project environment files and config to verify DSN patterns.