om-code-review

Repo-local extension of om-code-review for money-tracker.online - the context sources, the severity map and the suppression ban every review here applies.

SKILL.md

GitHub

om-code-review (money-tracker.online)

This file extends the shared om-code-review skill installed at .claude/skills/om-code-review/SKILL.md. Follow the shared skill's full workflow (validation gate, breaking-change gate, checklist, output format, verdict) with the repo-specific rules below layered on top.

Context sources

Read before the first finding, per touched area:

  • AGENTS.md (Always / Ask First / Never) and CODE_REVIEW.md.
  • docs/SPEC.md §6 (layering), §7 (definition of done), §8 (one concept, one owner), §9 (localisation), §10 (accessibility claim).
  • docs/MONEY.md in full when the diff touches an amount, a rate, a currency or a date bucket.
  • The ADR for the area in docs/adr/: 0002 tenant boundary, 0006 minor units, 0007 charts, 0010 ledger, 0011 auth, 0012 billing, 0008 no bank connectors.
  • docs/agents/reference.md: the clean-room rule. A string, a number, a name or a layout copied from the reference product is a blocker.
  • .ai/review-checklist.md (the configured reviewChecklist) and BACKWARD_COMPATIBILITY.md.

Severity map

Finding Severity
Float arithmetic on an amount; a hard-coded scale of 2; a recomputed ref_amount_minor; a cross-currency amount without its rate_e10 blocker
SQL outside src/server/repos; a removable tenant predicate; a write to account_balances outside src/server/ledger blocker
A server-function input not Zod-validated or not re-parsed in the repository; a permission enforced only client-side blocker
A new repository function without a tenant-isolation case; a money-touching write path without a money-invariant case blocker
A key added to en.ts and not pl.ts, or the reverse: the parity test fails the build blocker
A hand edit to src/routeTree.gen.ts or drizzle/; a migration that drops, renames or narrows in the same change that stops writing blocker
An added @ts-ignore, @ts-expect-error, eslint-disable, .skip, .only, .todo, .fixme, xit or xdescribe; a deleted test file blocker
A clean-room violation; a bank connector or credential capture (ADR 0008); a new dependency (SPEC §3); a broken surface in BACKWARD_COMPATIBILITY.md blocker
A WCAG 2.2 AA regression: unlabelled control, no keyboard path beside a drag, a chart without its text table, colour as the only channel, a target under 44px major
A blanket "WCAG AAA" claim in copy, comment or commit major
Slow work (import, export, email, backup, ledger rebuild, notification fan-out) run inline instead of enqueued on TASKS major
A second implementation of an owned concept: rounding, formatting, conversion, newId(), dispatch, a dialog primitive (SPEC §8) major
A missing loading, empty or error state on a screen major
TODO, console.log, commented-out code, placeholder copy minor

Review personas

CODE_REVIEW.md -> Review personas holds the dispatch table and the boundary: four subagents in .claude/agents/ (manual-tester, accessibility-specialist, security-tenancy-reviewer, product-owner), each dispatched on the paths it matches, each reporting only. Read the table there rather than a second copy here.

Two rules apply to this skill when it dispatches one:

  • A persona's finding enters the report at the severity of the map above, not at the severity the persona felt. Re-rank it; do not paste it.
  • A persona that could not run - no browser, no environment, Task not granted - is named in the report as not run. Silence reads as a pass, and a review that claims a lens it never used is the one failure mode this whole arrangement can produce.

Rules

  • The gate's two drift commands run first. A dirty generated file is a finding; never regenerate it silently in the review worktree.
  • A count or measurement in the review report came from a command whose output you read (docs/SPEC.md §7.7). When the number is 0, say what the command was aimed at.
  • Secrets never enter the report: no value from .dev.vars, .env.local or .ai/qa/test-env.env.