mcpField
A
audit

audit

github.com/hainrixzData & databases2273 installsUpdated 6d ago
Repository

About audit

audit is a read-only database auditing skill for projects assessing design integrity and performance. It coordinates parallel checks and reports separate scores, evidence, and prioritized actions.

Key features

  1. 01Detect the database stack and parse schemas
  2. 02Run read-only audits in parallel
  3. 03Score design and performance separately
  4. 04Roll up findings across multiple data stores
  5. 05Provide a prioritized action list with evidence

Use cases

  1. 01Review repository schema or migration files
  2. 02Verify findings against a live database
  3. 03Compare audit results across data stores

Install

Copy the command and run it in a terminal.

$npx skills add https://github.com/hainrixz/claude-db --skill audit

After installing, type/auditin Claude Code to confirm it loaded.

Details

From the project README

/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

  1. 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.
  2. 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).
  3. 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).

Comments

Markdown supported