| name | code-review |
| description | Perform a project-wide full-scope code review covering correctness, security auditing, test coverage, locale sync, documentation quality, code smells, UI/UX accessibility, and project conventions, then report findings and fix critical issues. |
When performing a code review, conduct a full project-wide sweep — do not limit scope to recent changes. Read broadly across the codebase and apply every check below.
Step 1 — Orient and Plan
Before reviewing, understand the system's current shape:
- Read
CLAUDE.md for architecture, conventions, and known gotchas
- Skim
lib/ directory structure: all contexts, LiveViews, controllers, components, schemas
- Check
doc/TODOs.md for known issues that may overlap with findings
- Prioritise: public endpoints, federation boundary, auth flows, content handling, admin routes
Step 2 — Correctness
- Logic errors, off-by-one mistakes, incorrect pattern matches, missing
nil/{:error, _} guards
- Ecto: missing
Repo.preload, N+1 risks, missing FOR UPDATE locks on counter updates (use Ecto.Multi), unescaped ILIKE inputs (require Repo.sanitize_like/1)
- LiveView: missing
handle_info clauses for all subscribed PubSub topics, stale socket assigns after redirects, race conditions between mount and handle_params
- Pagination: uses
Baudrate.Pagination (paginate_opts/3 + paginate_query/3) — no hand-rolled LIMIT/OFFSET
- Schema timestamps:
published_at (original feed entry date, bot posts only) vs inserted_at (local creation) — usage must match intent
- Federation: AP objects stamped with
ap_id post-insert; Publisher falls back to Federation.actor_uri/2 only when stored ap_id is nil
- Soft-delete: articles and comments use
deleted_at; queries must filter is_nil(deleted_at) where appropriate
Step 3 — Security Audit
Information security is the top priority. Treat every finding as potentially exploitable.
Input and injection
- No
String.to_atom/1 on any external or user input (atom table exhaustion)
- No user input in file paths (path traversal)
- All user-generated and federated HTML sanitized via
Baudrate.Sanitizer.Native (Ammonia NIF) before storage
- All dynamic assigns in HEEx rendered with
{@var} (auto-escaped); {:safe, ...} never wraps untrusted content (XSS)
- All user values parameterized in Ecto queries; no raw interpolation into
fragment() (SQL injection)
Network and federation
- Remote HTTP fetches: SSRF-safe (reject private/loopback IPs, HTTPS only); use only
Req (never HTTPoison, Tesla, httpc)
- AP payload ≤ 256 KB, content body ≤ 64 KB enforced at inbox
- Remote actor display names sanitized (strip HTML, control chars, truncate)
Authentication and authorization
- Every public LiveView route has the correct
on_mount hook (:require_auth, :require_admin, :require_admin_or_moderator, :require_admin_totp, etc.)
- Admin sudo mode (
:require_admin_totp) enforced on all admin-only actions — 10-minute TOTP re-verification window
- Bot accounts (
is_bot: true) rejected by authenticate_by_password/2; managed exclusively via Baudrate.Bots context
- Rate limiting applied to all public endpoints (auth, registration, AP inbox, API)
Key material and secrets
- Federation private keys encrypted via
KeyVault; TOTP secrets via TotpVault; no plaintext secrets in DB or logs
KeyStore.ensure_user_keypair/1 called before enqueuing any signed outbound activity
HTTPSignature.sign/5 / sign_get/3 must not return a "host" header (causes signature verification failures)
File handling
- File uploads: magic bytes validated; avatars re-encoded as WebP (EXIF stripped)
- No user-supplied filenames used directly on the filesystem
OWASP Top 10 checklist — cross-check each category:
A01 Broken Access Control · A02 Cryptographic Failures · A03 Injection · A04 Insecure Design · A05 Security Misconfiguration · A06 Vulnerable Components · A07 Auth Failures · A08 Software Integrity · A09 Logging Failures · A10 SSRF
Step 4 — Test Coverage
- Every public context function and LiveView action has a corresponding test in
test/ mirroring lib/
- LiveView/controller tests use
BaudrateWeb.ConnCase; context tests use Baudrate.DataCase
- Security-sensitive paths (auth, federation inbox, file uploads, rate limiting) have dedicated negative-path tests
- No
Process.sleep for timestamp ordering — use Repo.update_all with explicit timestamps instead
- All queries with user-visible ordering include a deterministic tiebreaker (e.g.,
desc: :id)
- Rate limiter stubbed via
BaudrateWeb.RateLimiter.Sandbox.set_global_response({:allow, 1}) in tests that hit rate-limited paths
- Federation delivery tests run synchronously (
federation_async: false in config/test.exs) — no async Task races
- Tests pass deterministically across all 4 partitions under concurrent execution — no shared mutable state, no order dependencies
Step 5 — Locale Sync
- Every user-visible string is wrapped in
gettext() — no bare English strings in templates, flash messages, HTML attributes (title, aria-label, placeholder), feed metadata, or error messages
%{var} interpolation used inside gettext() — never Elixir string interpolation ("#{}")
- After scanning
lib/ and priv/gettext/en/, verify that both priv/gettext/zh_TW/ and priv/gettext/ja_JP/ .po files contain translations for every msgid
- Terminology is consistent across all locales:
Board → zh_TW: 看板 (not 版面/板塊), ja_JP: 掲示板 (not ボード)
User → zh_TW: 使用者 (not 用戶)
- No stale/orphaned
msgid entries remaining in .po files after string removal
Step 6 — Documentation Quality
@moduledoc present and accurate for every public-facing module; @doc on every public function with non-obvious behaviour
doc/development.md reflects current architecture, contexts, auth flow, federation mechanics
doc/sysop.md reflects current installation, configuration, and maintenance procedures
doc/api.md documents all public AP and web API endpoints
doc/TODOs.md contains no items already completed
CLAUDE.md — stack table, key gotchas, and project conventions match the current codebase
README.md — feature list, prerequisites, and acknowledgements are current
- No commented-out dead code left in place of proper documentation
Step 7 — Code Smells
- Duplication: repeated logic across modules that should be extracted into a shared helper or context function
- Bloated functions: functions doing more than one thing; complex
with chains that should be broken into named steps
- Primitive obsession: raw strings/integers used where a well-named type, struct, or enum would be clearer
- Feature envy: a module reaching deeply into another context's internals instead of calling its public API
- Unnecessary complexity: over-engineered abstractions for one-time operations; speculative generality (YAGNI)
- Stale code: unused functions, dead branches, obsolete modules, leftover
IO.inspect / dbg calls, commented-out blocks
- Inconsistent naming: functions or variables that don't follow Elixir conventions or contradict surrounding code
- No nested modules in a single file — one module per file, always
- OTP paths: no
:code.priv_dir/1 in module attributes (@var); use Application.app_dir(:baudrate, "priv/...") in functions (runtime resolution)
- Cache coherence: settings and board mutations go through context functions; direct
Repo writes to settings/boards must call SettingsCache.refresh() / BoardCache.refresh() manually
- Avatar sizes: integers
[120, 48, 36, 24] only — never string names
Step 8 — UI/UX and Accessibility (a11y)
Semantic HTML
<section aria-labelledby="…"> for headed content areas
<article> for self-contained content items in lists
<aside> for supplementary content
<nav> for navigation landmarks
- Semantic
id attributes on content containers; unique id + semantic CSS class on each list item
WAI-ARIA compliance
- Interactive elements (buttons, links, inputs) have descriptive labels — either visible text,
aria-label, or aria-labelledby
- Dynamic regions that update without a page navigation use
aria-live appropriately
- Modal dialogs trap focus and restore it on close;
role="dialog" + aria-modal="true" present
- Icon-only buttons and icon links always have
aria-label or visually-hidden text
- Form fields have associated
<label> elements (via for/id or wrapping); error messages linked via aria-describedby
Keyboard and focus
- All interactive elements reachable and operable via keyboard alone
- Focus order is logical and follows visual reading order
data-focus-target present on primary content container in list/browse pages for post-navigation auto-focus
- No keyboard traps outside intentional modal dialogs
- Focus-visible ring visible on focused elements (check
app.css focus-visible styles)
Responsive design
- Layouts adapt correctly across mobile, tablet, and desktop breakpoints
- No horizontal overflow on small viewports from fixed widths
- Touch targets ≥ 44×44 px on interactive elements
Colour and contrast
- Text contrast ratio ≥ 4.5:1 (normal text) / 3:1 (large text) per WCAG 2.1 AA
- Information is not conveyed by colour alone — always paired with text, icon, or pattern
Layout correctness
{@inner_content} (not @inner_block) in layouts
- Templates not wrapped in
<Layouts.app> (causes duplicate flash IDs)
Reporting
Present all findings grouped by severity:
| Severity | Criteria |
|---|
| Critical | Security vulnerabilities, data loss, auth bypass, key material exposure — fix immediately |
| Major | Logic errors, missing test coverage for observable behaviour, broken conventions that cause runtime failures, a11y barriers that block screen-reader or keyboard users |
| Minor | Style, clarity, missing docs, i18n gaps, cosmetic a11y issues, minor code smells |
For each finding: cite the file and line number, describe the issue, explain the impact, and provide a concrete fix.
Fixing
After reporting, apply fixes for all Critical and Major findings directly. Then run the full test suite:
for p in 1 2 3 4; do MIX_TEST_PARTITION=$p mix test --partitions 4 --seed 9527 & done; wait
Do not consider the review complete until all tests pass. Diagnose and resolve any failures before finishing.