architecture-review

name	architecture-review
version	2.2.0
description	[Code Quality] Use when reviewing architecture compliance for layers, messaging, service boundaries, CQRS, repos, entity events, and data/consistency/tenancy boundaries. Universal architecture laws, coupling taxonomy and the anti-pattern catalog live in `.claude/docs/architecture-knowledge.md` (project docs always outrank it).

[BLOCKING] Execute skill steps in declared order. NEVER skip, reorder, or merge steps without explicit user approval. [BLOCKING] Before each step or sub-skill call, update task tracking: set in_progress when step starts, set completed when step ends. [BLOCKING] Every completed/skipped step MUST include brief evidence or explicit skip reason. [BLOCKING] If Task tools are unavailable, create and maintain an equivalent step-by-step plan tracker with the same status transitions.

Quick Summary

Goal: Ensure changes preserve architecture boundaries, ownership, message flow, and generated artifact integrity before handoff — validating changed code against layers, service boundaries, message flow, CQRS, repositories, entity events, frontend architecture, generated artifacts, recorded architecture decisions (ADRs), and quality tooling.

Summary:

• Purpose: validate a changeset against architecture rules the project records in its OWN reference docs; classify every finding PASS/WARN/BLOCKED with file:line proof; self-validate before handoff — one reviewer in the workflow-review-changes pipeline.
• Main phases — run in order: Phase 0 load architecture rules → Phase 1 determine scope → Phase 2 blast radius (if graph.db) → Phase 3 architecture review (13 categories) → Phase 4 finalize compliance report → Phase 5 /why-review self-validation gate → Next Steps AskUserQuestion.
• The 13 Phase-3 categories — review EVERY applicable one, serially: 0 quality-tooling baseline · 1 clean-architecture layers · 2 message-bus patterns · 3 CQRS compliance · 4 repository patterns · 5 service-pattern era (legacy vs modern) · 6 entity event handlers · 7 service boundaries · 8 frontend architecture (frontend files only) · 9 ADR / recorded-decision conformance · 10 spec-loop discipline (property-TC + dual-feedback) · 11 scalability & coupling regression (diff-scoped; BLOCKED/WARN) · 12 data, consistency & tenancy boundaries (dual-write, idempotency, isolation level + write skew, unfenced lock, breaking migration, tenant isolation, dataset writer ownership; BLOCKED/WARN). Per category: Think: derivation → doc rule → source evidence → file:line proof + grep 3+ counterexamples → verdict. NEVER scan categories in parallel; codebase convention wins over a suspected violation. — why: skipping a category silently drops the violation class it uniquely covers.
• Phase 0 is non-negotiable and first: load the project architecture docs (backend-patterns-reference.md, project-structure-reference.md, frontend-patterns-reference.md, code-review-rules.md) — every rule and base-class/symbol name comes from those docs, NEVER general knowledge; the framework names in Categories 2–8 are illustrative only.
• Universal reasoning comes from .claude/docs/architecture-knowledge.md (coupling taxonomy + four coupling dimensions, distributed-monolith signature, module-design principles §4, isolation levels + coordination primitives §8-§9, ~100-entry anti-pattern catalog, symptom→root-cause triage, judgment checklists §20) — use it to RECOGNIZE a defect class, then prove it with file:line. The project's own reference docs and accepted ADRs OUTRANK that catalog on every conflict — NEVER flag a deviation from the catalog as a project violation. An anti-pattern match is a HYPOTHESIS until evidence plus the damaged quality attribute are both named. — why: pattern-shape matching without project grounding is exactly the guess-as-fact failure this skill exists to prevent.
• Stay in lane: deep-review only what this skill OWNS (layers, messaging/CQRS/repos/service boundaries, entity events, frontend architecture, quality tooling, generated artifacts, ADRs); record a one-line → route to {sibling} pointer for security/performance/DDD/UI/test findings instead of expanding them. — why: duplicated findings across reviewers inflate severity counts and bury issues each reviewer uniquely owns.
• Read-only until validated: self-audit every draft finding against the 11 thinking red flags (architecture-knowledge.md §20.3) FIRST — a finding whose sacrifice/trade-off you cannot name, or that rests on "best practice", is demoted or deleted, never reworded — then run the Phase 5 /why-review self-validation gate before handoff; fixes happen only in the validated fix loop, and every fix restarts a full review from Phase 0. Write findings to plans/reports/arch-review-{date}-{slug}.md.

Default scope: All uncommitted changes (staged + unstaged). Override: specify files, directories, services, or full codebase.

MANDATORY MUST ATTENTION Plan tasks to READ architecture docs BEFORE reviewing:

1. docs/project-reference/backend-patterns-reference.md — CQRS, messaging, repos, validation, entity events, layer rules (READ FIRST — primary rules source)
2. docs/project-reference/project-structure-reference.md — service map, layer structure, DB ownership
3. docs/project-reference/frontend-patterns-reference.md — component hierarchy, store, API patterns (frontend files only)
4. docs/project-reference/code-review-rules.md — anti-patterns, conventions

Not found → search: "architecture documentation", "service patterns", "messaging patterns". Rules come from docs — NOT general knowledge.

Workflow:

1. Phase 0: Load Architecture Rules — Read project architecture docs (rules come from docs, NEVER general knowledge)
2. Phase 1: Determine Scope — Changed files (default) or user-specified scope
3. Phase 2: Blast Radius — Run /graph-blast-radius if graph.db exists
4. Phase 3: Architecture Review — Check each file serially against all 13 applicable categories (0 tooling → 12 data, consistency & tenancy)
5. Phase 4: Finalize — Generate compliance report with PASS/BLOCKED/WARN verdicts
6. Phase 5: Why-Review Self-Validation Gate — Adversarially validate own findings via /why-review before handoff (MANDATORY when any finding exists)
7. Next Steps — AskUserQuestion: /code-simplifier / /code-review / skip

Key Rules (top 3 critical first):

• MUST ATTENTION read project architecture docs in Phase 0 BEFORE reviewing — rules come from docs, NEVER general knowledge.
• Every violation needs file:line proof + grep 3+ counterexamples before flagging — NEVER speculate.
• MUST ATTENTION review one category at a time: doc rule → source evidence → verdict — NEVER scan categories simultaneously.
• Write findings to plans/reports/arch-review-{date}-{slug}.md.
• BLOCKED = must fix before merge | WARN = review and decide | PASS = compliant.
• Review is read-only until /why-review --validate-findings confirms findings; fixes happen only in the validated fix loop or downstream plan/feature-implement, and every fix restarts a full architecture review from Phase 0 with a fresh task breakdown.

Your Mission

$ARGUMENTS

First Principle — Easy to Change

Success metric: future change cost. DRY, SRP, abstraction, patterns, naming, layering, tests exist to make next change cheaper.

Before applying any rule, ask: does this lower or raise future change cost?

• Reject "best practices" raising cost: premature abstraction, speculative generality, leaky indirection, ceremony without payoff. — why: cost added now with no payoff is debt, not quality.
• Name real enemies: coupling, hidden state, duplicated knowledge, unclear intent, irreversible decisions exposed too early.
• Prefer simple reversible design over sophisticated rigid design. — why: reversible decisions cost less to undo when wrong.
• If downstream rule raises change cost, this principle wins.

---

Quality Tooling Principle — Tech-Stack Adaptive

Architecture review includes automated quality guardrails. Without stack-appropriate linting, formatting, type checks, static analysis, dependency/security-review scanning, CI enforcement, defects depend on reviewer memory.

Evaluate detected stacks, not fixed tool list:

• Detect stacks from project-reference docs, manifests, lock files, build files, CI before recommending tools.
• Per production stack, verify formatter/style config, linter/code analyzer, compiler/type-check strictness, dependency/vulnerability scanning, tests/coverage, CI/pre-commit enforcement.
• Prefer official or ecosystem-standard tooling; local docs absent/stale → check current official docs before recommending setup.
• MUST ATTENTION recommend enforceable best practice only: installed but unwired tool = WARN; production source with no relevant automated quality gate = BLOCKED.
• Identify missing capability first; map to local equivalent before prescribing new tooling. — why: prescribing a tool that duplicates an existing gate adds noise, not coverage.

---

Review Mindset (NON-NEGOTIABLE)

Skeptical. Every claim needs traced proof, confidence >80%.

• NEVER flag violations without reading actual code + tracing dependency — READ the code, trace the import chain, then flag.
• Every finding MUST include file:line evidence.
• Before flagging pattern violation: grep 3+ existing examples — codebase convention wins.
• Question: "Actually a violation, or an established exception?"

Ownership & Handoff (own vs delegate)

This skill = one reviewer in a multi-reviewer pipeline — workflow-review-changes runs it beside the siblings below. Review ONLY what this skill owns; route the rest so findings are not double-reported across reviewers.

This skill OWNS (deep-review here)	Delegate to sibling (one-line pointer only — do NOT deep-review)
Layer boundaries, dependency direction, business-logic placement	—
Messaging patterns, CQRS structure, repository patterns, service-pattern era, entity event handlers, service boundaries	—
Frontend ARCHITECTURE (base classes, store/effect, API-service base, subscription teardown, CSS-class presence)	Visual/SCSS/responsive/z-index quality → ui-review
Quality-tooling baseline, generated-artifact integrity, ADR / recorded-decision conformance	—
Architecture-level auth PLACEMENT (a gate exists at the boundary)	OWASP, secrets, dependency/supply-chain, authz-matrix depth → security-review
Structural soundness of a hot path (no obvious N+1 introduced by the diff)	Query plans, indexing depth, latency/throughput budgets → performance-review
—	Domain entity / value-object DDD design quality → domain-entities-review
—	Integration-test assertion quality, coverage, traceability → integration-test-review
—	Runtime production-readiness of service/API changes (observability wiring, rollback) → production-readiness-review

When a finding clearly belongs to a sibling, record one-line → route to {skill} pointer and move on — NEVER expand it. — why: duplicated findings across reviewers inflate severity counts and bury issues each reviewer uniquely owns.

Phase 0: Load Architecture Rules (MANDATORY FIRST)

MUST ATTENTION: Read project docs BEFORE reviewing. Rules come from docs, NEVER general knowledge.

• read docs/project-reference/backend-patterns-reference.md — extract messaging naming, layer rules, CQRS patterns, repo rules, entity event handler patterns, validation patterns
• read docs/project-reference/project-structure-reference.md — extract service map, layer structure, DB ownership
• frontend files in scope → read docs/project-reference/frontend-patterns-reference.md
• read docs/project-reference/code-review-rules.md — extract anti-patterns + review rules directly

Universal reasoning layer (secondary, never authoritative): consult .claude/docs/architecture-knowledge.md for the coupling taxonomy + four coupling dimensions, the distributed-monolith detection signature, the ~100-entry anti-pattern catalog, and the symptom→root-cause triage matrix. Use it to RECOGNIZE a defect class the project docs do not name explicitly. Honor its provenance markers in §3/§8/§9/§10: a row or section banner marked — VERIFY is an UNVERIFIED assertion — it may seed a hypothesis, but NEVER quote it as the authority for a finding; confirm against the named source (or the project's own docs) first.

MUST ATTENTION — precedence is absolute. Project reference docs and accepted ADRs > the knowledge catalog > general knowledge. NEVER report a deviation from the catalog as a project violation, and NEVER let a catalog entry override an established, grepped codebase convention. Every catalog-derived observation is a HYPOTHESIS until you have BOTH file:line/config/topology evidence AND the named quality attribute it damages — otherwise record it as INFO or drop it. — why: universal patterns applied as project rules generate confident false positives, the most expensive output this skill can produce.

Phase 1: Determine Scope

Default (no override): Review all uncommitted changes.

git status          # List changed files
git diff            # Staged + unstaged changes
git diff --cached   # Staged only

• Collect file list to review.
• Categorize: backend (.cs), frontend (.ts/.html), config, docs, other.
• Filter to architecture-relevant files (skip pure docs, configs, tests unless architecture-relevant).

Phase 2: Blast Radius (if graph.db exists)

• .code-graph/graph.db exists → call /graph-blast-radius skill.
• Record: impacted file count, cross-service impact, risk level.
• Prioritize review by highest-impact files first.
• Graph unavailable → note "Graph not available — skipping blast radius" and proceed.

Per changed file with downstream impact:

python .claude/scripts/code_graph trace <changed-file> --direction downstream --json

Flag MESSAGE_BUS consumers or event handlers impacted by changes.

Phase 3: Architecture Review

Create report: plans/reports/arch-review-{date}-{slug}.md

Per file in scope, evaluate against ALL applicable categories. Skip categories not applicable to file type.

MUST ATTENTION review serially. Per applicable category: read docs/source evidence → derive risk with Think: → grep 3+ examples/counterexamples → record PASS/WARN/BLOCKED. NEVER scan categories simultaneously — why: parallel scanning collapses per-category evidence into one undifferentiated pass and drops findings.

Portability note (MUST ATTENTION): Framework symbols, base-class names, directory conventions in Categories 2–8 below are illustrative examples — authoritative form comes from Phase 0 reference docs (backend-patterns-reference.md, frontend-patterns-reference.md, project-structure-reference.md); verify code against those docs. On any stack, map each example to project's equivalent as named in its own reference docs, flag deviations from project's actual convention — NEVER from these literal names. Same discipline as Category 5: read project docs at review time; NEVER treat a hardcoded name as universal.

---

Category 0: Quality Tooling Baseline — Severity: BLOCKED/WARN

Think: Can project automatically catch style, type, complexity, security, dependency, boundary regressions for detected stacks?

• Detect production stacks via docs/project-config.json, relevant docs, manifests, lock files, build files, CI.
• Inventory gates: formatter, linter, code/static analyzer, compiler/type checker, dependency audit/SCA/SBOM, SAST, test/coverage, architecture/dependency-boundary checks, pre-commit, CI/build.
• Verify stack-appropriate coverage: .editorconfig/language analyzers, JavaScript/TypeScript linting, UI template linting when supported, formatter config, dependency vulnerability scans, semantic security analysis.
• BLOCKED when production stack lacks runnable lint/static-analysis/type-check command and equivalent enforced gate, or CI/build references missing/broken quality command.
• WARN when tooling local-only, not wired into CI/build/pre-commit, partial for active production code, broadly/unexplainedly suppressed, unclear on generated-code exclusions, or stale for stack.
• Scope to the change (MUST ATTENTION): On normal change-level review, pre-existing tooling gap unrelated to diff is WARN with single note — NEVER BLOCK whole review on standing, change-unrelated condition. Reserve BLOCKED for: new stack/service introduced by this change with no gate, change itself removing/breaking existing gate, or explicit full-codebase/greenfield audit scope. — why: change review that BLOCKs on unrelated standing gap produces noise that buries regression the diff actually introduced.
• Before recommending tools, find current official/ecosystem setup and cite it; recommend capabilities first, tools second.

Fitness-function enforcement (architectural rules must be EXECUTABLE): an architectural rule not automatically verified is a SUGGESTION and will be violated within a quarter. Check whether the project's own recorded architectural rules have a machine check — and whether THIS change adds a rule with no check:

Rule the project records	Fitness function expected (any equivalent counts)
Layer / dependency direction	Architecture test in CI (ArchUnit / NetArchTest / dependency-cruiser / import-linter / lint boundary rules)
No module cycles	Cycle detection failing the build
Domain purity	Assertion that the domain package imports no framework/ORM/HTTP namespace
API / event schema compatibility	OpenAPI-or-protobuf diff gate · schema-registry backward-compat check · consumer contract tests
Multi-tenant isolation	Test asserting a cross-tenant query returns zero rows (see Category 12)
Resilience	"Every outbound call has a timeout" lint or test
Performance / bundle budget	Latency-or-size assertion in the pipeline

• WARN when a recorded architectural rule has NO machine check and relies on review discipline alone. BLOCKED when this change REMOVES or disables an existing architecture/boundary check, or introduces a new enforced-by-prose-only boundary while the project already has a fitness-function mechanism available.
• Existing violation backlog is fine if it is a RATCHET (new violations blocked, count only decrements) — a "cleanup later" comment with no gate is WARN.

Violation format:

BLOCKED: {stack} has no enforced lint/static-analysis/type-check quality gate ({evidenceFile}:{line})
WARN: recorded architecture rule "{rule}" has no fitness function — enforced by review discipline only ({docFile}:{line})

---

Category 1: Clean Architecture Layers — Severity: BLOCKED

Think: What layer is this file in? What layers can it legally import from? Does any import break inward-only flow (Service/API → Application → Domain ← Persistence)?

• Read docs/project-config.json → architectureRules.layerBoundaries for project-specific rules.
• Determine layer from file path: Domain/, Application/, Persistence/, Service/.
• Scan configured language's import/include statements — flag imports from forbidden layers.
• MUST ATTENTION verify business logic in correct layer: Entity/Domain > Service/Application > Controller/Component.
• NEVER allow direct infrastructure access from Domain — keep repo interfaces in Domain, implementations in Persistence. — why: Domain depending on infrastructure inverts the dependency rule and couples core logic to a swappable detail.
• NEVER allow business logic in API/Controller layer — push it down to Entity/Domain or Application.
• Module cycles (BLOCKED): a cycle in the module/package dependency graph means the two modules ARE one deployable unit whatever the folder structure says. Detect with the graph (trace --direction both) or the project's dependency tool; flag any NEW cycle the diff introduces. — why: an unbroken cycle makes independent testing, release and extraction impossible, and it never gets easier to cut later.
• Domain purity (BLOCKED): the domain layer MUST NOT import ORM attributes, HTTP types, SQL, serialization or framework namespaces. Grep the changed domain files for the infrastructure namespaces named in backend-patterns-reference.md.
• Shared/infra layer domain leak (BLOCKED): a generic/shared/infrastructure layer MUST reference NO consumer-specific domain concept — tenant/customer/product IDs, business entities, feature rules. This leak compiles, runs, and passes functional tests while silently coupling the "reusable" layer to one consumer. Fix by keeping the shared type domain-free and pushing domain fields down into the consumer via subclass/composition. — why: a layer coupled to one consumer's domain is no longer reusable, and every later consumer inherits the wrong abstraction.
• Cohesion / dumping ground (WARN): a new or growing Utils/Common/Shared/Helpers/Managers module that everything imports is coincidental cohesion — it becomes the coupling hub and the cycle source. Test: "how many DIFFERENT reasons would make me edit this file?" More than one actor ⇒ split.
• Pass-through layer (WARN): a layer that only forwards calls unchanged (sinkhole) adds cost with no responsibility — collapse it or give it a real job.
• Shallow module / pass-through method (WARN): a new type whose public interface is nearly as large as its implementation, or a method that only forwards to the next layer with no added responsibility, earns nothing — it is interface cost with no hidden complexity. Judge module VALUE as functionality hidden ÷ interface surface: many tiny one-method classes ("classitis") raise total complexity while looking modular. Prefer pulling the complexity DOWNWARD into one deep module over spreading it across N call sites. — why: reviewers count classes and read it as modularity, so this defect is the one that survives review and then makes every future change touch five files.
• DIP placement (WARN, BLOCKED when the project's docs require it): dependency inversion is only real when the interface lives in the domain/policy package and the adapter package depends inward. An interface declared beside its single implementation in the infrastructure package is a naming convention, not inversion — grep where the changed port/interface is DECLARED, not merely where it is used. — why: an interface in the infra package leaves the dependency arrow pointing the wrong way while the code reads as clean architecture.
• Wrong-abstraction extraction (WARN): a diff that MERGES two code paths that look alike but change for DIFFERENT reasons creates a shared module with two actors. Duplication is cheaper than the wrong abstraction — require three real occurrences sharing the same reason to change (rule of three) before extracting. Verify against the project's own strategic-DRY decision before flagging. — why: a premature abstraction is defended by everyone who depends on it, so its cost compounds while duplication's cost stays linear.

Violation format:

BLOCKED: {layer} layer file {filePath}:{line} imports from {forbiddenLayer} layer ({importStatement})
BLOCKED: {filePath}:{line} introduces module cycle {A} → {B} → {A}
BLOCKED: shared/infra {filePath}:{line} references consumer domain concept {concept} — shared layer must stay domain-free

---

Category 2: Message Bus Patterns — Severity: BLOCKED/WARN

Think: Does this message correctly name its type (event vs request)? Does it extend the right base class? Is producer/consumer relationship correctly oriented — does the leader service own the event?

Naming (BLOCKED):

• Event messages + request messages MUST follow project's bus-message naming convention — encode owning service + feature + action, with distinct suffix distinguishing event-kind from request-kind messages. Resolve exact convention + suffixes from backend-patterns-reference.md.
• Grep existing examples in source for current stack's message-naming pattern before flagging — codebase convention wins.

Base classes (BLOCKED): Verify against bus base types named in backend-patterns-reference.md (Phase 0); concrete names are illustrative examples.

• Bus messages MUST extend project's trackable/payload bus-message base — see backend-patterns-reference.md.
• Consumers MUST extend project's message-bus consumer base — see backend-patterns-reference.md.
• Producers MUST extend project's event-bus-message producer base — see backend-patterns-reference.md.

Upstream/Downstream (BLOCKED):

• Leader service owns entity data → defines event message for it.
• Follower services consume events — NEVER produce events about data they don't own. — why: producing events about non-owned data forks the source of truth across services.
• NO circular listening: A→B + B→A for same data = boundary violation.
• Consumers MUST implement project's cross-message dependency-wait primitive for cross-message data dependencies — see backend-patterns-reference.md.

Ordered delivery (WARN):

• Messages requiring ordered processing MUST set project's ordered-delivery / sub-queue partition key to meaningful value (resolve concrete API from backend-patterns-reference.md).
• Unordered messages leave it unset / null.

Reliable publication — dual write (BLOCKED):

• NEVER write the database and publish a message as two independent operations. Any changed flow that commits state AND publishes MUST go through the project's inbox/outbox mechanism or CDC (verify the project's inbox/outbox enablement config — see backend-patterns-reference.md). — why: dual write fails silently in both directions — DB commits + publish fails ⇒ downstream never learns and diverges forever; publish succeeds + DB rolls back ⇒ phantom downstream data referencing a row that does not exist. Neither failure appears in tests or logs.
• Consumer idempotency (BLOCKED): exactly-once DELIVERY is impossible, so at-least-once is what the bus gives you. Every changed consumer MUST be idempotent — dedup on message ID with a TTL, a version/sequence check that discards backward transitions, or a naturally idempotent write (SET status = 'paid', never balance += x). Grep the handler for the project's dedup primitive. — why: a non-idempotent consumer produces duplicate side effects — double charge, double email, double shipment — only under redelivery, which is exactly when nobody is watching.
• Poison message + queue bounds (WARN): capped retries with exponential backoff and jitter, then DLQ; DLQ depth monitored. Flag unbounded queues, uncapped/infinite retry, and unjittered retry (synchronizes into a thundering herd). A DLQ nobody watches is data loss with extra steps.
• Command-in-event-costume (WARN): an "event" with exactly ONE permitted consumer that MUST handle it, whose failure means the business flow failed, is a command misnamed as an event — you pay async debugging difficulty AND keep the sync coupling. Flag the naming, route the redesign to domain-analysis.
• Raw-row events (WARN): publishing internal DB rows/columns as the event contract (CDC with no mapping layer) freezes your schema as a public contract by Hyrum's Law — consumers then depend on columns you can never rename.
• Durability acknowledgement weakened (BLOCKED when a diff lowers it, WARN when newly introduced): a change to producer ack mode, replication factor, or min-in-sync-replicas that accepts data loss on leader failure. acks=1 loses the message when the leader dies before replication; acks=all waits for every replica CURRENTLY in the ISR, so min.insync.replicas is the FLOOR that decides whether a shrunken ISR rejects the write or accepts it with no error — with min.insync=1 a healthy 3-replica ISR still waits for 3, but durability DEGRADES to acks=1 the moment the ISR shrinks to the leader alone. Check BOTH values together, never one alone, and flag the degradation path rather than asserting an unconditional equivalence. — why: the config reads as durable while the guarantee is not, and the loss appears only during the failover nobody rehearsed.
• Ordering/replay assumption unmet (WARN): a changed consumer that assumes global ordering when the broker gives ordering only per partition key, or that assumes replay is available on a consume-and-gone queue. Verify the required ordering scope (global / per-entity / none) and replay window against the project's broker config. Queue-vs-log semantics → .claude/docs/architecture-knowledge.md §10.

Also verify:

• NEVER direct cross-service DB access — MUST use message bus. — why: direct DB reach couples services and bypasses ownership boundaries.
• last-sync-timestamp field on message used for conflict resolution in consumers (resolve concrete field from backend-patterns-reference.md).
• Event schema changes are backward-compatible (a rolling deploy runs old and new consumers simultaneously) — a breaking schema change with no versioning or compatibility gate is BLOCKED.

Violation format:

BLOCKED: {filePath}:{line} writes DB then publishes without outbox/CDC — dual write, silent divergence on partial failure
BLOCKED: {filePath}:{line} consumer is not idempotent — no dedup/version check under at-least-once delivery

---

Category 3: CQRS Compliance — Severity: BLOCKED/WARN

Think: Is Command+Result+Handler in one file? Is validation using fluent API (not exceptions)? Does DTO own mapping, not the handler? Are side effects in event handlers, not command handlers?

File organization (BLOCKED):

• Command + Result + Handler MUST be in ONE file under the command folder for the feature (resolve concrete folder from project's structure reference / docs/project-config.json; e.g. {command-folder}/{Feature}/)
• Query + Result + Handler MUST be in ONE file under the query folder for the feature (resolve concrete folder from project's structure reference / docs/project-config.json; e.g. {query-folder}/{Feature}/)

Validation (BLOCKED):

• MUST use project's validation-result fluent API — NEVER throw exceptions for validation; return validation result instead — verify exact type + method names in backend-patterns-reference.md. — why: exceptions for expected-invalid input conflate control flow with errors and skip the validation pipeline.
• Sync validation in command's validate hook, async in request-validation hook — see backend-patterns-reference.md for hook names.

DTO mapping (BLOCKED):

• DTOs MUST own entity mapping via project's DTO base mapping methods — NEVER map in command handlers; map in the DTO instead — see backend-patterns-reference.md for method names.

Side effects (BLOCKED):

• NEVER put side effects (notifications, sync, cascade updates) in command handlers — place them in Entity Event Handlers instead. — why: side effects in the handler couple the command to downstream concerns and cascade failures.
• Side effects go in Entity Event Handlers under project's event-handler folder (resolve from project's structure reference / docs/project-config.json; e.g. {event-handler-folder}/)
• Each handler = one independent concern (failures don't cascade).

---

Category 4: Repository Patterns — Severity: BLOCKED

Think: Is this using a service-specific repo interface, not the generic one? Are complex queries extracted to RepositoryExtensions?

• MUST use project's service-specific repository abstraction — NEVER the generic root-repository base directly; per-service naming scheme defined in backend-patterns-reference.md. — why: the generic base leaks unbounded query surface and erases per-service boundaries.
• Complex queries MUST use project's repository-extension pattern with static expressions (e.g. RepositoryExtensions)
• All query filter/FK/sort columns MUST have database indexes.

Violation format:

BLOCKED: {filePath}:{line} uses the generic root-repository base instead of the service-specific repository — see backend-patterns-reference.md for the required naming

---

Category 5: Service Pattern Era (Legacy vs Modern Split) — Severity: BLOCKED (new services) / WARN (existing)

Think: When project distinguishes legacy vs modern service patterns (e.g., auth scheme, telemetry stack, permission model, language-version syntax), is this a new service (must follow modern) or an existing legacy service (expect legacy patterns)? Is the modern pattern partially mixed into a legacy service without full migration?

New services — BLOCKED if any legacy-only pattern used. Identify project's modern-pattern checklist from injected reference docs (e.g., project-structure-reference.md, ADRs, scaffolding templates) and verify every item.

Existing legacy services — WARN if modern patterns partially mixed without full migration. Flag legacy patterns only when partial mixing creates inconsistency; in their own consistent context they are expected, NOT violations.

Determining era: Read project's reference docs at review time — service-pattern era assignments are project-specific and listed authoritatively there. NEVER hardcode service names in this skill. — why: hardcoded service names rot the moment the project renames or adds a service, and break portability to other repos.

---

Category 6: Entity Event Handlers — Severity: BLOCKED/WARN

Think: Are side effects defined inline in command handlers (wrong) or in project's event-handler folder (correct)? Does each handler have a single concern?

Location (BLOCKED):

• Entity event handlers MUST be in project's event-handler folder (resolve from project's structure reference / docs/project-config.json; e.g. {event-handler-folder}/)
• NEVER inline side effects in command handlers — move them to a dedicated entity event handler. — why: inline side effects couple the command to downstream concerns and cascade failures.

Implementation (BLOCKED):

• MUST extend project's entity-event application-handler base — see backend-patterns-reference.md.
• MUST implement CRUD-action filter hook — see backend-patterns-reference.md for hook name.
• One handler = one independent concern.

Naming (WARN):

• Convention: {Action}On{Trigger}EntityEventHandler
• Grep existing examples before flagging.

Producer patterns (BLOCKED):

• Bus message producers MUST extend project's event-bus-message producer base — see backend-patterns-reference.md.
• MUST implement message-build + action-filter hooks — see backend-patterns-reference.md for hook names.

---

Category 7: Service Boundaries — Severity: BLOCKED

Think: Does any code reach directly into another service's database or project reference? All cross-service data flow MUST go through the message bus.

• NEVER direct DB access to another service's database — route through the message bus. — why: direct DB reach couples services and bypasses ownership boundaries.
• NEVER using reference to another service's domain/persistence project — depend on shared message contracts instead.
• Cross-service communication via message bus only (event bus or request bus).
• Shared data through shared message projects, NOT direct references.
• Verify service-to-DB mapping from project-structure-reference.md.
• One writer per dataset (BLOCKED): every dataset has exactly ONE owning service/module that writes it. Many READERS are fine — via API, replica, or published event stream. Flag any change that adds a second writer to a dataset owned elsewhere. — why: shared write access is shared coupling with NO contract; the data then diverges and no layer owns correctness.
• Shared domain library (WARN): a change making many services depend on one shared domain library forces lockstep deploys — that is deployment + semantic coupling, not reuse. Small duplicated DTOs are cheaper; share only genuinely generic technical utilities. Verify against the project's own strategic-DRY decision before flagging.
• Entity-shaped boundaries (WARN): a new service/module named for a NOUN (UserService, ProductService) doing CRUD signals boundaries drawn around data instead of capabilities — it guarantees real use cases must synchronously traverse many services. Flag the smell; route the re-modeling to domain-analysis.

Violation format:

BLOCKED: {filePath}:{line} references {otherService} domain/persistence directly — must use message bus
BLOCKED: {filePath}:{line} writes {dataset} owned by {otherService} — second writer, no contract

---

Category 8: Frontend Architecture (if frontend files in scope) — Severity: BLOCKED/WARN

Think: Are components extending the right base class? Is state going through the store? Are subscriptions properly cleaned up?

Verify against frontend-patterns-reference.md (Phase 0, frontend files); concrete names are illustrative examples.

• Components MUST extend project's component base classes (BLOCKED) — see frontend-patterns-reference.md.
• State MUST use project's view-model store + reactive-effect pattern — NEVER manual signals or direct HTTP client (BLOCKED); route state through the store — see frontend-patterns-reference.md.
• API services MUST extend project's API-service base (BLOCKED) — see frontend-patterns-reference.md.
• All subscriptions MUST use project's auto-teardown operator — NEVER manual unsubscribe (BLOCKED). — why: manual unsubscribe is forgotten on early-return paths and leaks subscriptions — see frontend-patterns-reference.md.
• All template elements MUST carry project's CSS-naming-convention classes (WARN) — see frontend-patterns-reference.md.
• Logic in lowest layer: Model > Service > Component (WARN).

Boundary with /ui-review: This category owns frontend ARCHITECTURE — base classes, view-model store / reactive-effect pattern, API-service base, subscription teardown, layer placement, CSS-naming-class presence. VISUAL/styling quality — long-content overflow, responsive multi-screen flex, flex-vs-fixed sizing, z-index discipline, SCSS/CSS detail — owned by /ui-review, which /changes-review invokes as its UI dimension when frontend changes present. Flag missing base classes / store / teardown here; defer SCSS-quality depth + visual-layout findings to ui-review to avoid double-reporting.

---

Category 9: ADR / Recorded-Decision Conformance (if docs/adr/** or recorded ADRs exist) — Severity: BLOCKED/WARN

Think: Does any changed file contradict a binding decision recorded in an accepted ADR — a rejected library/technology, a forbidden dependency direction, a recorded quality-attribute/NFR budget, a banned pattern — without a superseding ADR?

This category closes design→review loop: /architecture-design emits ADRs + fitness-function choices; this category verifies changed code still conforms to them. Checks CONFORMANCE only — NEVER re-runs deep performance or security analysis (those route to siblings in the Ownership & Handoff matrix).

• Locate recorded decisions: docs/adr/** (or ADR location named in project's reference docs). Read only ADRs with Status: Accepted — skip Superseded/Proposed/Rejected.
• Extract each accepted ADR's binding constraints: chosen vs rejected options, layer/dependency rules, NFR targets (latency/throughput/availability/RPO-RTO), banned patterns.
• Per changed file, check conformance against those constraints — grep the diff for reintroduced rejected options or forbidden references; cite file:line.
• BLOCKED when change contradicts an accepted ADR's binding decision and no superseding ADR exists. Correct way to change a recorded decision = new superseding ADR (per docs/adr/0001 lifecycle), NEVER a silent violation in the diff. — why: silent ADR violations erode the decision record and let rejected options creep back unreviewed.
• WARN when change drifts from a recorded guideline, or is NFR-impacting against a recorded budget — flag it and route depth check to performance-review/security-review.
• New one-way door with NO ADR (WARN, or BLOCKED when the project mandates ADRs): when the diff makes a decision that is expensive to REVERSE and no ADR records it, flag it. One-way doors: a new data model or primary-key strategy · a tenancy-model change · a consistency-model change on a read path · a NEW sync-vs-async choice at a boundary · a new public API or event contract · a new service boundary · a new cloud-primitive lock-in · an auth/identity model change · a data residency or retention decision. — why: an unrecorded irreversible decision cannot be enforced by this category later, and the next engineer relitigates or silently breaks it.
• No ADRs exist → record "No recorded ADRs — conformance N/A" and skip (this category NEVER blocks a project that has chosen not to keep ADRs). Do NOT retroactively demand ADRs for pre-existing decisions on a change-level review — scope to what the diff decides.

Violation format:

BLOCKED: {filePath}:{line} contradicts {adr-id} ("{decision}") with no superseding ADR
WARN: {filePath}:{line} makes a hard-to-reverse decision ({decision}) with no recorded ADR

---

Category 10: Spec-Loop Discipline (applies across all categories) — Severity: BLOCKED/WARN

Think: Does each behavior-affecting architecture finding feed BOTH the spec and a guarding test, or only the code? Does any [HARD] rule or cross-boundary invariant ship with no property TC?

• BLOCKED when a [HARD] architecture rule or cross-boundary invariant (layer contract, message-ownership rule, CQRS/repo invariant, service-boundary guarantee) has no universally-quantified property TC + boundary counter-case — an example-only test does not guard a rule that must hold for ALL inputs.
• Every behavior-affecting architecture finding MUST carry a Dual-Feedback row (spec axis + test axis): the spec NAMES the changed contract/invariant AND a test GUARDS it — blank either axis = INCOMPLETE; NEVER record an architecture finding as code-only.
• Review the whole package — spec + tests + structural diff — NOT the structural diff alone; loop until zero new spec-loop gaps remain, each cycle enriching the spec. — why: a boundary change that compiles but is never asserted regresses silently the next time a sibling service is touched.

Violation format:

BLOCKED: {filePath}:{line} [HARD] {rule/invariant} has no property TC (spec axis: {present/blank} | test axis: {present/blank})

---

Category 11: Scalability & Coupling Regression — Severity: BLOCKED/WARN

Diff-scoped regression guard, NOT a project audit. This category catches architecture/scalability regressions a change introduces; it does NOT re-grade the whole system. Init/on-demand grading of all 10 scorecard areas is owned by architecture-scalability-review. Cross-cutting spec-loop discipline (Category 10) also applies to findings from this category. Cross-references Category 7 (service boundaries) and Category 9 (ADR / recorded-decision conformance).

Think: Does this diff introduce a new sync cross-context call, shared-DB reach, statefulness, or copy-pasted cross-context logic that regresses module isolation, loose coupling, or horizontal scaling — turning a clean boundary into distributed-monolith coupling?

Judge the FOUR coupling dimensions SEPARATELY — low code coupling proves nothing on its own:

Dimension	Question	Evidence to pull
Code	Who imports whom? Any cycle?	Import graph / trace --direction both
Temporal (runtime)	Must both be UP at once? A sync chain makes your SLO the PRODUCT of the chain's	Call graph across the boundary, sync vs async
Semantic (contract)	Does a change in their MEANING force a change in mine?	Shared schema, shared entity/domain lib, shared enum
Operational (deployment)	Must these ship TOGETHER? If yes they ARE one service	Release coordination, shared version pin

Distributed-monolith signature (BLOCKED when the diff moves the system toward it): LOW code coupling + HIGH temporal + semantic + deployment coupling. Concrete detection: services that must release together · a shared database or shared entity/domain library · a request synchronously traversing ≥4 services · one team blocked by another's deploy · "we roll back all services together." — why: it pays every distributed cost and buys none of the benefits; it is the most expensive and most common modern architecture failure, and every diff that deepens it makes the exit harder.

• BLOCKED when a change adds a NEW forbidden cross-context dependency, a circular context dependency, or a direct sync cross-context call where the recorded architecture requires an event/message or an owned contract (producer calling consumer directly — "I call you" instead of "you listen to me").
• BLOCKED when the diff adds a NEW outbound network call, lock or long query with no timeout — a missing timeout is the single most common resilience defect, and the dangerous dependency failure is SLOW, not down: it exhausts your pool while every health check stays green. Also flag uncapped or unjittered retries on the new call, and retries on a NON-idempotent operation.
• BLOCKED when the diff deepens a synchronous chain to ≥4 hops, or adds a sync hop to a path the recorded architecture requires to be async. Availability MULTIPLIES down a sync chain (five 99.9% deps ⇒ ~99.5%) and latency SUMS.
• WARN for a new connascence of Value or Timing ACROSS a boundary (two services that must change a value together, or that depend on execution order). That is a BOUNDARY error, not a bug to patch — either the invariant belongs in one transactional boundary, or it must be made eventual ON PURPOSE with a designed compensating action. Route the redesign to domain-analysis.
• BLOCKED when a change makes a previously stateless/scalable path stateful in a way that breaks horizontal scaling — in-memory session/cache assumed node-local, sticky-instance state, a new SPOF, or unbounded fan-out on a hot path — where the ADR/scale budget requires statelessness.
• WARN for cross-context duplication (copy-pasted domain rule/util across contexts — a DRY regression), a new shared-DB read across a boundary, or a distributed-monolith smell with weaker evidence; record the smell and route the deep fix.
• Detect-only smells → route, do NOT deep-analyze here: local hot-path/query/N+1 latency → performance-review; rollout/capacity/SRE/runtime readiness → production-readiness-review; bounded-context / aggregate re-modeling → domain-analysis; auth/secret/tenant-boundary coupling → security-review. This category flags the regression at file:line; the sibling owns the depth. — why: a diff reviewer that re-runs full capacity/DDD analysis blows its context and duplicates the sibling's job.
• Behavior-affecting findings carry a Dual-Feedback row (spec axis + test axis) per Category 10 — a coupling/scaling regression that changes a contract MUST enrich BOTH the spec AND a guarding test, never code-only.

Violation format:

BLOCKED: {filePath}:{line} new sync cross-context call to {context} — recorded architecture requires event/message (route deep coupling design to domain-analysis)
WARN: {filePath}:{line} cross-context duplication of {rule/util} — DRY regression (route shared-lib decision to architecture-scalability-review / scaffold)

Technique applicability (advisory — INFO, does NOT alter this category's verdict): When the diff touches a scale-sensitive surface, invoke SYNC:scale-technique-gate — derive the scale tier from evidence, then emit the Technique Applicability Matrix as an INFO/advisory block noting warranted-but-missing techniques and any OVER-ENGINEERED ones. This is guidance only: it is NEVER a BLOCKED/WARN finding, does NOT change the Category 11 severity, and does NOT feed the Phase 4 verdict table. A MISSING-WARRANTED technique is advice to consider, not a regression. Full catalog → .claude/docs/scale-technique-catalog.md.

---

Category 12: Data, Consistency & Tenancy Boundaries — Severity: BLOCKED/WARN

Diff-scoped, and the LEAST reversible category. Data outlives every service, framework and team: a defect here is a future data migration, a silent divergence, or a cross-tenant breach — never a simple refactor. Skip ONLY when the diff touches no persistence, no consistency boundary, no message consumer and no tenant-scoped data. Overlaps by design with Category 2 (dual write / idempotency at the messaging layer) and Category 7 (dataset writer ownership) — record the finding ONCE in the category that owns the mechanism, and cross-reference.

Think: Does this diff write state and publish without atomicity? Can it be replayed safely? Is the isolation level named, and is every check-then-act invariant actually protected? Does any lock or leadership claim carry a fencing token? Does every query and cache key carry the tenant? Does a schema change survive a rolling deploy? Is any read path newly stale with no declared budget?

Consistency & atomicity (BLOCKED):

• Dual write — state committed to the DB and a message/webhook/second-store write performed as two independent operations, with no outbox/CDC. Cross-reference Category 2. — why: it diverges silently in both directions and no test catches it.
• Non-idempotent consumer or mutation — a changed message handler or unsafe endpoint with no dedup key, version check, or naturally idempotent write, under at-least-once delivery or client retry. Cross-reference Category 2.
• Cross-network transaction — a database transaction held OPEN across an HTTP/RPC/queue call. Locks held during I/O convert one slow dependency into a database-wide stall.
• 2PC/XA introduced across services — flag it; the correct shape is saga + outbox + compensation. And a compensation must be a NEW BUSINESS FACT (RefundIssued), not a pretend rollback of a real-world side effect.
• Unprotected check-then-act (write skew / lost update) — a changed path that READS state, decides, then WRITES based on that read, inside a transaction at Read Committed or snapshot isolation, with no protecting mechanism. Neither Read Committed nor snapshot isolation prevents write skew: two concurrent transactions both pass the check and both write. The change is correct only if it uses ONE of: a DB constraint (unique/check/exclusion) · SELECT … FOR UPDATE on the rows read · Serializable isolation with app-side handling of serialization failures · a single atomic conditional write (UPDATE … WHERE version = n / WHERE stock > 0). Grep the changed handler for the project's transaction/isolation primitive and 3+ existing examples before flagging. Anomaly table → .claude/docs/architecture-knowledge.md §8. — why: this is the classic on-call data-corruption bug — it passes every single-user test and fires only under production concurrency.
• Range invariant guarded by row locks — an invariant over rows that DO NOT YET EXIST (no double-booking, no overlapping interval, at most N per tenant) cannot be protected by locking the rows you read. Requires a constraint, an exclusion index, a materializing lock row, or Serializable. NEVER accept "we check before inserting" as protection.
• Unfenced distributed lock / leader election — a changed flow that takes a distributed lock or relies on leadership without a monotonic fencing token the storage layer rejects when stale, a stated lease duration, and defined behaviour when the lease expires mid-operation. A GC or VM pause makes a dead holder believe it is still the holder ⇒ split brain. Also flag consensus/quorum cluster SIZING: quorum is ⌊N/2⌋+1, so an even-sized cluster tolerates no more failures than the odd size below it (4 tolerates 1, same as 3) while paying an extra node and a larger quorum, and a 2-node cluster tolerates ZERO failures — strictly worse than 1 node. Flag the wasted/harmful sizing; NEVER claim an even-sized cluster cannot reach a majority or make progress. — why: an unfenced lock fails exactly once — under the pause, at peak, writing corrupted state — and it looks correct in every test.
• Wall-clock ordering or expiry (WARN, BLOCKED when correctness depends on it) — ordering distributed events, resolving conflicts, or computing lock/lease expiry from a wall clock (now(), request timestamps from another host). Use per-entity versions/sequences, logical/Lamport or vector clocks, an HLC, or a bounded-uncertainty clock; measure elapsed time with a MONOTONIC clock. Also flag a new multi-leader/multi-region write path with no named conflict-resolution rule. — why: clocks in a distributed system disagree without bound, so "latest timestamp wins" silently discards writes.

Staleness (WARN):

• A newly eventually-consistent read path with no declared, monitored staleness budget — unbounded, unmeasured lag IS the defect.
• A read-after-write path newly routed to a replica with no read-your-writes guarantee (sticky read / read-from-primary / version token). Most user-visible "consistency bugs" are this, not missing linearizability.

Tenant isolation (BLOCKED — treat as a security-adjacent defect):

• A tenant-scoped query, repository method, projection, background job, export or report with no tenant predicate, where the project's mechanism (row-level security / ORM global filter / mandatory repository base) does not automatically apply it. One missing WHERE tenant_id = ? is a cross-tenant breach that passes every functional test.
• tenant_id (or user/role/price) taken from a client-supplied field rather than the authenticated principal. Cross-reference security-review for authz depth — this category owns only the BOUNDARY placement.
• A cache, memo, or shared in-memory map keyed WITHOUT the tenant/identity/permission dimension. Cache-key omission is a recurring cross-tenant leak vector that no functional test detects.
• WARN when a tenant-isolation change lands with no test asserting a cross-tenant read returns ZERO rows (the fitness function from Category 0).

Migrations & schema (BLOCKED):

• A breaking schema change in one deploy — column/table dropped or renamed, type narrowed, or NOT NULL added without a default — while the deployment strategy is rolling/canary. Both versions run simultaneously; this breaks old pods and makes rollback impossible. The required shape is expand–contract: add nullable → dual write → backfill → switch reads → stop writing old → drop, in separate deploys.
• A migration that is not forward-only/idempotent, or that blocks writes on a large table with no online/batched strategy.

Data-access structure (WARN — route depth to performance-review):

• Unbounded query with no DB-side filter or pagination on a path whose result set grows with data (OOM and latency both scale with the table). Check row COUNT before row SIZE — pushing the filter to the DB beats projecting columns.
• Deep OFFSET pagination where keyset/cursor is available; a new query filter/FK/sort column with no index; analytics query newly added against the OLTP primary.
• Distributed cache treated as the source of truth (data unrecoverable after eviction/restart), or a new unbounded cache with no max size/eviction.
• A new denormalized copy with no named mechanism keeping it correct.

Lifecycle (WARN):

• New PII/PHI/PCI-class field with no classification, retention, residency or deletion story — retrofitting erasure into a denormalized or event-sourced store is brutally expensive, and erasure conflicts with immutable event stores by design.
• New public identifier exposing an internal sequential PK (enumerable, leaks volume).

Violation format:

BLOCKED: {filePath}:{line} writes {store} then publishes {message} without outbox/CDC — dual write
BLOCKED: {filePath}:{line} tenant-scoped query has no tenant predicate and no enforced filter — cross-tenant read possible
BLOCKED: {filePath}:{line} cache key omits tenant/identity dimension ({key}) — cross-tenant leak
BLOCKED: {migrationFile}:{line} drops/renames {column} in one deploy under rolling release — use expand–contract
BLOCKED: {filePath}:{line} check-then-act at {isolationLevel} with no constraint/FOR UPDATE/Serializable/conditional write — write skew possible
BLOCKED: {filePath}:{line} distributed lock taken with no fencing token or lease-expiry handling — split brain possible
WARN: {filePath}:{line} orders/expires distributed events by wall clock — use versions/sequences or a logical clock
WARN: {filePath}:{line} new eventually-consistent read path with no declared staleness budget or SLI

MUST ATTENTION every Category-12 finding still obeys the evidence gate: read the actual query/handler/migration, grep 3+ existing examples of the project's tenant-filter and outbox primitives, and confirm the mechanism is NOT already applied automatically at a lower layer before flagging. — why: projects that enforce tenancy in a repository base or via row-level security will show no predicate at the call site and are CORRECT — flagging those is the highest-noise false positive available in this category.

---

Phase 4: Finalize — Architecture Compliance Report

Update report with final sections:

Verdict Scoring

Verdict	Condition
BLOCKED	1+ BLOCKED findings — must fix before merge
WARN	0 BLOCKED, 1+ WARN findings — review and decide
PASS	0 BLOCKED, 0 WARN — architecture compliant

Report Structure

# Architecture Review Report — {date}

## Scope

- Files reviewed: {count}
- Services affected: {list}
- Blast radius: {summary from Phase 2}

## Verdict: {PASS | WARN | BLOCKED}

## BLOCKED Findings (Must Fix)

### {Category}: {description}

- **File:** {path}:{line}
- **Rule:** {rule from project doc}
- **Evidence:** {what was found}
- **Fix:** {what to change}

## WARN Findings (Review)

### {Category}: {description}

- **File:** {path}:{line}
- **Rule:** {rule from project doc}
- **Evidence:** {what was found}
- **Recommendation:** {suggested action}

## PASS Categories

- {list of categories that passed with no findings}

## Architecture Health Summary

- Quality Tooling Baseline: {PASS/WARN/BLOCKED}
- Clean Architecture: {PASS/WARN/BLOCKED}
- Messaging Patterns: {PASS/WARN/BLOCKED}
- CQRS Compliance: {PASS/WARN/BLOCKED}
- Repository Patterns: {PASS/WARN/BLOCKED}
- Service Pattern Era: {PASS/WARN/BLOCKED}
- Entity Event Handlers: {PASS/WARN/BLOCKED}
- Service Boundaries: {PASS/WARN/BLOCKED}
- Frontend Architecture: {PASS/WARN/BLOCKED/N/A}
- ADR / Recorded-Decision Conformance: {PASS/WARN/BLOCKED/N/A}
- Spec-Loop Discipline (property-TC + dual-feedback): {PASS/WARN/BLOCKED}
- Scalability & Coupling Regression (4 coupling dimensions): {PASS/WARN/BLOCKED}
- Data, Consistency & Tenancy Boundaries: {PASS/WARN/BLOCKED/N/A}
- Technique applicability (advisory — INFO, does NOT alter verdict): {matrix summary or N/A-by-scale}

The "Technique applicability" line is advisory/INFO only — it reports the scale-tier technique matrix as guidance and NEVER changes any category severity or the Phase 4 verdict.

---

Architecture Boundary Check (Automated)

Per changed file:

1. Read docs/project-config.json → architectureRules.layerBoundaries
2. Determine layer — match file path against each rule's paths glob patterns
3. Scan imports — grep for configured language's import/include statements
4. Check violations — import path contains layer name in cannotImportFrom = violation
5. Exclude framework — skip files matching architectureRules.excludePatterns
6. BLOCK on violation: "BLOCKED: {layer} layer file {filePath} imports from {forbiddenLayer} layer ({importStatement})"

architectureRules absent from project-config.json → skip silently.

---

Systematic Review Protocol (10+ changed files)

1. Categorize — Group files by service/layer/concern.
2. Parallel Sub-Agents — Launch one architect sub-agent per category with architecture-specific checklist.
3. Synchronize — Collect findings, cross-reference service boundaries.
4. Consolidate — Single holistic report with per-category verdicts.

---

Phase 5: Why-Review Self-Validation Gate (MANDATORY when findings exist)

Purpose: Adversarial validation of own findings BEFORE handoff — catch over-flagged Highs, false positives, severity inflation at source, not downstream.

Trigger: Any finding produced (Critical, High, Medium, OR Low). Skip ONLY when report's verdict is unconditional PASS with literally zero findings.

Protocol:

1. Read own finalized report from plans/reports/{skill}-{date}-{slug}.md
2. Invoke /why-review skill with arg: validate findings in plans/reports/{skill}-{date}-{slug}.md — verify each finding has file:line proof, steel-man each rejected interpretation, and stress-test severity classifications
3. Read validation verdict path returned by why-review, expected as plans/reports/why-review-validate-{date}.md
4. why-review demotes/removes any finding → UPDATE own finalized report with revised severities, remove false positives, add ## Why-Review Validation Notes section citing what changed + why.
5. why-review confirms all findings → Append ## Why-Review Validation line to own report stating "All N findings re-validated against actual code; no severity changes."

Self-audit your OWN findings FIRST (before invoking why-review): run the 11 thinking red flags in .claude/docs/architecture-knowledge.md §20.3 against each draft finding and each recommendation. The four that fire most often in this skill: you cannot name what your recommendation SACRIFICES · you say "best practice" instead of naming the forces it balances · you flagged a scale problem you cannot evidence · you are treating a two-way door as irreversible. Any hit invalidates the REASONING — demote or delete the finding, do not reword it. — why: a finding that survives only because it sounds authoritative consumes the team's fix budget and trains them to ignore the report.

Skip conditions (record explicit reason if skipping):

• Verdict unconditional PASS with zero findings → log "Skipped — no findings to validate".
• Why-review skill itself is active context (avoid recursion).

Why this exists: AI sub-agent reports inherit confirmation bias — orchestrator absorbs severity claims as ground truth. The 2026-05-09 review incident produced 5 Highs; adversarial validation demoted 3. Codified as standard practice.

---

Next Steps

MANDATORY — NO EXCEPTIONS: After completing, use AskUserQuestion to present:

• "/code-simplifier" (Recommended) — Simplify and refine code
• "/code-review" — Deep code quality review
• "Skip, continue manually" — user decides

Combined audit: For a whole-project architecture + compliance + production-readiness audit in one pass, run /architecture-review-full (or /start-workflow workflow-architecture-audit) — it fans out this skill, architecture-scalability-review, and production-readiness-review as parallel sub-agents and synthesizes one consolidated report.

AI Agent Integrity Gate (NON-NEGOTIABLE)

Before reporting ANY work done:

1. Grep every removed name. Extraction/rename/delete → grep confirms 0 dangling refs across ALL file types.
2. Ask WHY before changing. Existing values intentional until proven otherwise — NEVER "fix" without traced rationale.
3. Verify ALL outputs. One build passing ≠ all builds passing — check every affected stack.
4. Evaluate pattern fit. Copying nearby code? Verify preconditions match — same scope, lifetime, base class, constraints.
5. New artifact = wired artifact. Created something? Prove registered, imported, reachable by all consumers.

[IMPORTANT] Use TaskCreate to break ALL work into small tasks BEFORE starting. Simple tasks: ask user whether to skip.

Fresh Context Re-Review — Eliminate orchestrator confirmation bias after fixes by restarting the full review with isolated sub-agents where applicable.

Why: The main agent knows what it (or /feature-implement) just fixed and rationalizes findings accordingly. A fresh sub-agent has ZERO memory, re-reads from scratch, and catches what the main agent dismissed. Sub-agent bias is mitigated by (1) fresh context, (2) verbatim protocol injection, (3) main agent not filtering the report.

When: ONLY after a validated-finding fix cycle, or when the user/workflow explicitly requests an independent high-risk architecture synthesis pass. A review pass that finds issues triggers validation first; it does NOT trigger a fresh-context pass over the same findings before validation/fix.

How:

1. Start a NEW full review invocation/task breakdown; when that protocol calls for agents, spawn a NEW Agent tool call — use architect subagent_type for architecture reviews (see Sub-Agent Type Override above)
2. Inject ALL required review protocols VERBATIM into the prompt — see SYNC:review-protocol-injection for the full list and template. Never reference protocols by file path; AI compliance drops behind file-read indirection (see SYNC:shared-protocol-duplication-policy)
3. Sub-agent re-reads ALL target files from scratch via its own tool calls — never pass file contents inline in the prompt
4. Sub-agent writes structured report to plans/reports/{review-type}-round{N}-{date}.md
5. Main agent reads the report, integrates findings into its own report, DOES NOT override or filter

Rules:

• NEVER reuse a sub-agent across rounds — every iteration that uses sub-agents spawns a NEW Agent call
• NEVER fix unvalidated findings; validate first using the caller's validation gate
• NEVER skip the full review restart after a validated fix cycle — every fix invalidates the prior verdict
• Continue until a complete full review pass has zero findings; if the same blocker repeats across 3 full invocations with no progress, escalate via AskUserQuestion
• Track iteration count in conversation context (session-scoped, no persistent files)

Sub-Agent Type Override

MANDATORY: Architecture reviews spawn architect sub-agent, NOT code-reviewer. Keep subagent_type: "architect" from canonical template below; NEVER revert to code-reviewer. Rationale: architect carries cross-service impact analysis, ADR creation, multi-service security/performance context that code-reviewer lacks for architecture-level decisions.

Review Protocol Injection — Every fresh sub-agent review prompt MUST embed 11 protocol blocks VERBATIM. The template below has ALL 11 bodies already expanded inline. Copy the template wholesale into the Agent call's prompt field at runtime, replacing only the {placeholders} in Task / Round / Reference Docs / Target Files / Output sections with context-specific values. Do NOT touch the embedded protocol sections.

Why inline expansion: Placeholder markers would force file-read indirection at runtime. AI compliance drops significantly behind indirection (see SYNC:shared-protocol-duplication-policy). Therefore the template carries all 11 protocol bodies pre-embedded.

Subagent Type Selection

• architect — ALWAYS for architecture reviews (cross-service, ADR, security/performance at system level)
• code-reviewer — for code quality reviews only (NOT architecture)

Canonical Agent Call Template (Copy Verbatim)

Agent({
  description: "Fresh Round {N} review",
  subagent_type: "architect",
  prompt: `
## Task
{review-specific task — e.g., "Review all uncommitted changes for code quality" | "Review plan files under {plan-dir}" | "Review integration tests in {path}"}

## Round
Round {N}. You have ZERO memory of prior rounds. Re-read all target files from scratch via your own tool calls. Do NOT trust anything from the main agent beyond this prompt.

## Protocols (follow VERBATIM — these are non-negotiable)

### Spec ↔ Tests ↔ Code Triangulation
DO THIS FIRST — before any per-protocol check below. The review target is the WHOLE PACKAGE, not the diff alone: load the behavior's spec (§3 ACs / §4 BRs / §8 TCs), its tests, and the changed code TOGETHER, and reason about their mutual consistency BEFORE judging any one in isolation.
1. Locate all three faces: the Feature Spec section(s) governing the changed behavior, the tests that guard it, and the production code that implements it. A missing face is itself a finding (SPEC-GAP / TEST-GAP / DEAD-SPEC).
2. Triangulate pairwise — every disagreement is a finding; classify which face is wrong:
   - code vs spec: behavior the code does that no §3/§4/§8 rule describes → CODE-EXTRA or SPEC-STALE; a [HARD] §4 rule or §5 invariant with no enforcing code path → CODE-WRONG.
   - tests vs spec: a §8 TC with no test, or a test asserting behavior no TC/rule names → TEST-GAP or SPEC-SILENT.
   - tests vs code: a changed code path with no covering test → TEST-GAP; a test that still passes against a deliberately broken invariant → WEAK-TEST (apply the mutation thinking in Bug Detection).
3. Hidden-rule capture: any invariant the code enforces but the spec never states (SPEC-SILENT) MUST be surfaced as a finding to add into §3/§4/§8 AND guarded with a test — the enrichment loop, never a silent pass.
4. Only after the three faces agree — or every disagreement is logged as a finding — proceed to the per-protocol checks below; when enrichment adds spec/test content, re-review the package against the enriched spec.
NEVER mark review PASS while any spec/test/code face disagrees without a logged finding. The diff is the entry point; the package is the unit of judgment.

### Evidence-Based Reasoning
Speculation is FORBIDDEN. Every claim needs proof.
1. Cite file:line, grep results, or framework docs for EVERY claim
2. Declare confidence: >80% act freely, 60-80% verify first, <60% DO NOT recommend
3. Cross-service validation required for architectural changes
4. "I don't have enough evidence" is valid and expected output
BLOCKED until: Evidence file path (file:line) provided; Grep search performed; 3+ similar patterns found; Confidence level stated.
Forbidden without proof: "obviously", "I think", "should be", "probably", "this is because".
If incomplete → output: "Insufficient evidence. Verified: [...]. Not verified: [...]."

### Bug Detection
MUST check categories 1-4 for EVERY review. Never skip.
1. Null Safety: Can params/returns be null? Are they guarded? Optional chaining gaps? .find() returns checked?
2. Boundary Conditions: Off-by-one (< vs <=)? Empty collections handled? Zero/negative values? Max limits?
3. Error Handling: Try-catch scope correct? Silent swallowed exceptions? Error types specific? Cleanup in finally?
4. Resource Management: Connections/streams closed? Subscriptions unsubscribed on destroy? Timers cleared? Memory bounded?
5. Concurrency (if async): Missing await? Race conditions on shared state? Stale closures? Retry storms?
6. Stack-Specific: Check the configured language/runtime pitfalls and framework-specific failure modes discovered from local code.
Classify: CRITICAL (crash/corrupt) → FAIL | HIGH (incorrect behavior) → FAIL | MEDIUM (edge case) → WARN | LOW (defensive) → INFO.

### Design Patterns Quality
Priority checks for every code change:
1. DRY via OOP: Same-suffix classes (*Entity, *Dto, *Service) MUST share base class. 3+ similar patterns → extract to shared abstraction.
2. Right Responsibility: Logic in LOWEST layer (Entity > Domain Service > Application Service > Controller). Never business logic in controllers.
3. SOLID: Single responsibility (one reason to change). Open-closed (extend, don't modify). Liskov (subtypes substitutable). Interface segregation (small interfaces). Dependency inversion (depend on abstractions).
4. After extraction/move/rename: Grep ENTIRE scope for dangling references. Zero tolerance.
5. YAGNI gate: NEVER recommend patterns unless 3+ occurrences exist. Don't extract for hypothetical future use.
Anti-patterns to flag: God Object, Copy-Paste inheritance, Circular Dependency, Leaky Abstraction.

### Logic & Intention Review
Verify WHAT code does matches WHY it was changed.
1. Change Intention Check: Every changed file MUST serve the stated purpose. Flag unrelated changes as scope creep.
2. Happy Path Trace: Walk through one complete success scenario through changed code.
3. Error Path Trace: Walk through one failure/edge case scenario through changed code.
4. Acceptance Mapping: If plan context available, map every acceptance criterion to a code change.
5. Tests Verify Intent: For test/spec changes, verify tests name the protected business rule or invariant and would fail if that intent breaks.
6. Migration Test Exclusion: Do not write tests for migration code. Schema/data migrations are one-time execution paths, not core application logic.
NEVER mark review PASS without completing both traces (happy + error path).

### Test Spec Verification
Map changed code to test specifications.
1. Identify the project's test/spec format from existing docs, test-case files, BDD feature files, or spec folders.
2. Every changed code path MUST map to a corresponding test case/spec (or flag as "needs test case").
3. New functions/endpoints/handlers → flag for test spec creation.
4. Migration files are excluded from test/spec creation; schema/data migrations are one-time execution paths, not core application logic.
5. If spec evidence fields exist, verify they point to actual code (file:line, not stale references).
6. Verify each meaningful test case names the business intent/invariant; flag behavior-only cases that only mirror implementation details.
7. Auth/data changes → verify corresponding authorization and data-state test cases exist.
8. If no specs exist for a changed path → log the gap and recommend the project's test-spec workflow.
NEVER skip test mapping. Untested code paths are the #1 source of production bugs.

### Behavioral Delta Matrix
MANDATORY for any bugfix review. Produce input-state × pre-fix × post-fix × delta table BEFORE writing verdict.
- Minimum 3 rows; include at least one row OUTSIDE the original bug report.
- Any "REGRESSION" delta → review returns FAIL until a preservation test is added.
- Narrative descriptions do NOT substitute for the matrix.
Example rows (external-record sync fix):
| Input                 | Pre-fix | Post-fix                  | Delta      |
| --------------------- | ------- | ------------------------- | ---------- |
| Record exists (valid) | Reused  | Always recreated → orphan | REGRESSION |
| Record missing (404)  | Error   | Recreated                 | Fixed      |

### Fix-Layer Accountability
NEVER fix at the crash site. Trace the full flow, fix at the owning layer. The crash site is a SYMPTOM, not the cause.
MANDATORY before ANY fix:
1. Trace full data flow — Map the complete path from data origin to crash site across ALL layers (storage → backend → API → frontend → UI). Identify where bad state ENTERS, not where it CRASHES.
2. Identify the invariant owner — Which layer's contract guarantees this value is valid? Fix at the LOWEST layer that owns the invariant, not the highest layer that consumes it.
3. One fix, maximum protection — If fix requires touching 3+ files with defensive checks, you are at the wrong layer — go lower.
4. Verify no bypass paths — Confirm all data flows through the fix point. Check for direct construction skipping factories, clone/spread without re-validation, raw data not wrapped in domain models, mutations outside the model layer.
BLOCKED until: Full data flow traced (origin → crash); Invariant owner identified with file:line evidence; All access sites audited (grep count); Fix layer justified (lowest layer that protects most consumers).
Anti-patterns (REJECT): "Fix it where it crashes" (crash site ≠ cause site, trace upstream); "Add defensive checks at every consumer" (scattered defense = wrong layer); "Both fix is safer" (pick ONE authoritative layer).

### Rationalization Prevention
AI skips steps via these evasions. Recognize and reject:
- "Too simple for a plan" → Simple + wrong assumptions = wasted time. Plan anyway.
- "I'll test after" → RED before GREEN. Write/verify test first.
- "Already searched" → Show grep evidence with file:line. No proof = no search.
- "Just do it" → Still need TaskCreate. Skip depth, never skip tracking.
- "Just a small fix" → Small fix in wrong location cascades. Verify file:line first.
- "Code is self-explanatory" → Future readers need evidence trail. Document anyway.
- "Combine steps to save time" → Combined steps dilute focus. Each step has distinct purpose.

### Graph-Assisted Investigation
MANDATORY when .code-graph/graph.db exists.
HARD-GATE: MUST run at least ONE graph command on key files before concluding any investigation.
Pattern: Grep finds files → trace --direction both reveals full system flow → Grep verifies details.
- Investigation/Scout: trace --direction both on 2-3 entry files
- Fix/Debug: callers_of on buggy function + tests_for
- Feature/Enhancement: connections on files to be modified
- Code Review: tests_for on changed functions
- Blast Radius: trace --direction downstream
CLI: python .claude/scripts/code_graph {command} --json. Use --node-mode file first (10-30x less noise), then --node-mode function for detail.

### Understand Code First
HARD-GATE: Do NOT write, plan, or fix until you READ existing code.
1. Search 3+ similar patterns (grep/glob) — cite file:line evidence.
2. Read existing files in target area — understand structure, base classes, conventions.
3. Run python .claude/scripts/code_graph trace <file> --direction both --json when .code-graph/graph.db exists.
4. Map dependencies via connections or callers_of — know what depends on your target.
5. Write investigation to .ai/workspace/analysis/ for non-trivial tasks (3+ files).
6. Re-read analysis file before implementing — never work from memory alone.
7. NEVER invent new patterns when existing ones work — match exactly or document deviation.
BLOCKED until: Read target files; Grep 3+ patterns; Graph trace (if graph.db exists); Assumptions verified with evidence.

## Reference Docs (READ before reviewing)
- docs/project-reference/code-review-rules.md