| name | audit |
| description | Audit a database schema, ORM model, or migrations for design and performance — produces two independent 0-100 scores (Design & Integrity + Performance & Scale) plus a prioritized, evidence-backed report. Read-only; never writes or migrates. Use when the user asks to audit, analyze, review, check, or score a database, schema, data model, migrations, indexes, keys, referential integrity, types, constraints, query patterns, partitioning, or connection pooling — for Postgres, MySQL, Mongo, DynamoDB, Cassandra, pgvector, ClickHouse, Neo4j, and friends. |
| argument-hint | <path|connection-target> [--paradigm auto|relational|document|key-value|wide-column|vector|time-series|graph] [--tier 0|1|2] |
| allowed-tools | Read, Grep, Glob, Bash, Task |
/claude-db:audit
A full, read-only database audit. Never writes files, never mutates the database — even at Tier 1/2 it runs only read-only verification queries.
$ARGUMENTS = <path|connection-target> [flags]. The target is a repo path (schema/ORM/migration files) and, optionally, a live database via $DATABASE_URL for Tier-1/2 verification. If no target is given and no schema/ORM/migration artifacts are found in the working directory, do not fabricate a pass: say so plainly and suggest /claude-db:start (guided wizard, zero artifacts) or /claude-db:design (greenfield).
What to do
- Invoke the db-orchestrator skill with the target and flags. It detects the stack (
scripts/detect-stack.mjs — paradigm/engine/ORM/platform), parses the schema (scripts/parse-schema.mjs / parse-orm-python.py), records the data tier reached (0/1/2), dispatches the read-only auditor subagents in parallel, merges findings, and runs db-score.
- Present:
- The two scores — Design & Integrity and Performance & Scale — each with a band (A–F) and a one-line interpretation. Show the uncapped
computed alongside any capped:true (a fail+severity:5 on that axis caps it at F). Never blend the two.
- A per-category breakdown for each score (value, weight, active?), the detected paradigm/engine, the tier reached, and the count of
needs_api checks (score confidence — never a silent pass).
- If multiple datastores were detected, the worst-of-across-stores roll-up per axis with the flooring store named, plus the per-store breakdown.
- A prioritized action list sorted by impact: each item with status, severity, evidence (
evidence.observed quoting real DDL/query with secrets redacted), recommendation, fixability (auto/proposed/advisory), and expected_impact (axis + confidence + magnitude, banded high|medium|low — never a fabricated %/latency/row-count).
- End by offering: "Run
/claude-db:fix to apply the safe, reversible changes (you confirm each one), /claude-db:migrate to lint a migration, or /claude-db:next to see what to tackle first."
Two scores, never blended. Every finding conforms to schema/finding.schema.json with reproducible evidence. Respond in the user's language (EN/ES).
