quality

---

name: quality description: "AI code quality checks. The Big 5: input validation, edge cases, error handling, duplication, complexity. Triggers: quality, big 5, ai code, review, validate." allowed-tools: Read, Grep, Bash kernel: kind: methodology version: 1 side_effects: none confirmation: none

<quick_checks>

# 1. Missing validation
grep -r "req\.body" --include="*.ts" --include="*.js" | grep -v "parse\|validate\|z\." | head -5

# 2. Empty catch blocks
grep -r "catch.*{}" --include="*.ts" --include="*.js" | head -5

# 3. String concat in queries (SQL injection)
grep -rE "SELECT.*\$\{|INSERT.*\$\{" --include="*.ts" --include="*.js" | head -5

</quick_checks>

<data_correctness> For any pipeline that extracts or transforms figures (financial, metrics, counts):

• Parse deterministically (a real parser, regex, typed loader). The LLM never generates, transforms, or "fixes" numeric values.
• Units are explicit at parse time (percent vs fraction, counts vs currency); a value never crosses unit categories through arithmetic.
• Tie-out gate: derived aggregates must reproduce the source's own totals before any output is shown downstream. A delta between your output and the source is assumed to be YOUR normalization bug until proven otherwise.
• Silent-empty guard: "no findings" produced from an empty parse is a failure of the parse, not a finding. </data_correctness>

<on_complete> agentdb write-end '{"skill":"quality","big5_checked":true,"violations":N}' </on_complete>