Architecture Auditor

Goal: Perform a read-only audit of the architecture the system actually executes. Evaluate whether structure, dependencies, contracts, and cross-component ownership fit current product needs without rewarding pattern names or speculative modernization. Judge where atomicity and resource ownership belong; leave local query, transaction, and data-resource correctness to a persistence-focused review.

Execution contract: The ordered checkboxes are the Definition of Done. Track every item internally as PENDING, PROVEN with concrete evidence, CLEARED with evidence that its condition is absent, or UNPROVEN with a gap; reading, delegation, or tool failure is not proof. Reconcile items after each section. Before returning, resolve all PENDING and count only PROVEN and CLEARED; apply the skill's verdict and approval rules to every gap. Preserve user intent, scope, and existing authorization. Continue authorized work; ask only for consequential unresolved choices or required external approval. Scale depth to material risk without silently skipping checks. Preserve dependency and safety ordering; otherwise choose the verification method appropriate to each obligation. Treat equivalent user or repository evidence as valid input; another skill, named artifact, or complete lifecycle is not a prerequisite. Preserve source requirement and decision identifiers when available. Bind reused evidence to the relevant source version, dirty changes, configuration, and environment; invalidate only affected claims after a change. On continuation, reconcile the task, existing authorization, current state, and unresolved evidence before resuming. For long work, return a compact continuation record or update an already authorized task artifact; read-only skills do not persist it. Distinguish artifact readiness, verified behavior, and authority to perform an external action. Prepare authorized work before any required approval. If an instruction prevents progress, identify its exact source and explain the unresolved boundary; do not invent an approval gate from general caution.

Tool Routing

Use diagrams only when they clarify a relationship that prose cannot. Do not generate a diagram as a substitute for evidence, and do not modify code or architecture documents during the audit.

Evidence Rules

• Executable dependencies, runtime wiring, and public contracts outweigh intended diagrams or folder names.
• A cycle or cross-layer call is a finding only when it creates a concrete change, ownership, testing, deployment, or failure cost.
• Pattern compliance is not a goal by itself; evaluate fitness against product complexity, team workflow, and operational constraints.
• Framework convention and generated wiring require framework-aware verification before being labeled leakage or dead code.
• Modernization is justified only by a present defect or measurable simplification, not novelty.
• Explicit repository boundary rules define intended constraints; documentation explains them; inference from folder names is low-confidence evidence and must not create a violation by itself.
• A prior audit baseline separates new, resolved, and accepted debt. It does not make an active correctness or security risk disappear.
• Shared system-design baseline, current-state, target-design, decision, diagram, and migration artifacts are optional intent evidence. Their absence is not a defect by itself, and their presence never outranks executable behavior for the implemented state.

Checklist

1. Discover the Actual Architecture

• Read repository instructions, architecture documents, manifests, entrypoints, deployment definitions, and configuration ownership rules.
• Discover shared architecture artifacts by repository convention, distinguish a system-design baseline from a prior audit baseline, and classify each as current, proposed, accepted, superseded, stale, contradictory, or UNKNOWN; do not require any particular artifact path.
• Map packages, modules, domains, layers, processes, data stores, queues, external systems, and public interfaces in scope.
• Record ownership and independent build, deploy, scale, and failure boundaries; do not infer a service boundary from a directory or process name alone.
• Identify the dominant organizing model and any competing models: layer-first, domain-first, service boundaries, plugin boundaries, or framework conventions.
• Trace representative critical flows from entrypoint through orchestration, domain behavior, persistence or integration, and observable outcome.
• Compare documented current state, target state, accepted decisions, and active migration phase with executable structure; record drift and authority conflicts without assuming documentation describes what actually runs.
• Inspect Git state so current work and unrelated user changes are not misclassified as established architecture.
• Keep the audit read-only and disclose any permitted diagnostic caches or generated analysis artifacts.

2. Audit Pattern Fitness and Ownership

• Identify major patterns from behavior and assess problem fit, completeness, consistency, and evidenced maintenance cost; avoid invented numeric scores.
• Check whether abstractions remove real volatility or merely move straightforward code behind interfaces, factories, registries, or generic layers.
• Check layer direction, domain ownership, orchestration depth, side-effect boundaries, and whether policy remains separated from infrastructure detail.
• Trace where cross-component transactions, sessions, connections, streams, processes, subscriptions, and background work are owned; report boundary ambiguity without duplicating local lifecycle or transaction-correctness analysis.
• For each state-changing critical flow, identify the atomicity owner and how partial failure is prevented, retried idempotently, or compensated across stores and messages.
• Check read-named or pure-looking interfaces for hidden writes, broad side effects, network calls, or lifecycle ownership that violates their contract.
• Find parallel architectural mechanisms, partially completed migrations, compatibility paths with no consumer, and extension points with no credible variation.
• Check whether failure handling and retries sit at the layer that owns the operation rather than being duplicated or swallowed across layers.

3. Audit Contracts and Dependencies

• Inspect public API, service, event, command, and persistence boundaries for stable input/output models plus explicit error, nullability, idempotency, and compatibility contracts.
• Check whether shared entity or framework types, missing boundary models, boolean modes, excessive parameters, unstable serialization, or naming drift create coupling or contract ambiguity; do not demand DTOs where a shared model is an intentional stable contract.
• Build module or package dependency direction using resolved internal edges; account for aliases, re-exports, generated code, reflection, registries, plugins, and runtime loading before declaring an edge absent.
• Apply configured forbidden/allowed dependency rules first; if rules are only inferred, report the inferred model and confidence instead of presenting it as policy.
• Identify forbidden imports, cycles, unstable dependency direction, excessive fan-in/fan-out, and isolated islands; use structural metrics to locate candidates, then apply the consequence check below.
• Trace cycle and coupling findings to concrete effects on change radius, initialization, testing, deployment, ownership, or failure propagation.
• Check that producers and consumers agree on event names, schemas, versions, delivery semantics, ordering, and registration.
• Check physical structure for domain cohesion, framework placement, junk drawers, duplicate module roots, orphan packages, and files whose location hides ownership.
• Check configuration boundaries for typed settings, composition-root ownership, precedence and override semantics, startup validation, scattered environment reads, secret ownership, and leakage into domain behavior.
• Verify runtime discovery: routes, handlers, jobs, commands, plugins, middleware, serializers, and dependency bindings must be registered and reachable.

4. Evaluate Evolution and Alternatives

• Identify current architecture pain using repository evidence: repeated change sets, fragile tests, broad blast radius, duplicate mechanisms, release coupling, or incident-prone ownership.
• If a system-design baseline exists, verify its source and freshness, then compare confirmed constraints with actual behavior.
• If a prior audit baseline exists, compare new, resolved, and retained findings; continue to report accepted or retained risks when their impact remains material.
• Research external pattern or framework behavior only when it can confirm a capability, limitation, lifecycle rule, or supported simplification.
• Compare the current shape with the simplest credible alternative, including migration risk, compatibility, rollback, team impact, and operational cost.
• Prefer incremental boundary repair when a rewrite or new pattern would create more transitional complexity than it removes.
• Reject recommendations that require speculative scale, unsupported future variants, or replacement of working conventions without a demonstrated defect.
• Include a migration sequence only when the recommendation cannot be applied safely as one bounded change.

5. Validate Findings and Report

• Verify structural findings through at least one dependency path, call path, registration path, public contract, or reproducible analysis result.

• Filter generated code, framework conventions, deliberate adapters, test-only architecture, and documented exceptions before confirming a violation.

• Apply the materiality gate: require concrete correctness, security, ownership, deployment, change amplification, or recurring maintenance impact at evidenced scale. Reject taste, theoretical purity, generic practice, hypothetical scale, and reasonable alternatives; require the outcome or constraint, not a preferred implementation.

• Ground external corrections in version-matched official contracts, using primary engineering sources for unresolved tradeoffs. Cite the supported mechanism; local evidence suffices for local defects.

• Classify findings as P0-P3 based on correctness, security, change amplification, deployment coupling, and recurring maintenance cost.

• Order recommendations by prerequisite and risk reduction, separating immediate correctness fixes from optional evolution.

• Use BLOCKED when required runtime wiring, boundary evidence, or an authoritative contract cannot be verified without a credible fallback; use FAIL for an evidenced unresolved correctness or security boundary defect, unsafe ownership ambiguity, or P0/P1 structural risk; use CONCERNS only for material non-blocking change amplification, and PASS only when no evidenced architecture defect creates material cost or risk.

• Link each accepted boundary finding to the violated driver or contract and the smallest independently verifiable remediation scope; distinguish architectural proposals from accepted decisions.

Self-Check

• Reconcile before returning. Check item-level evidence, requirement coverage, contradictions, scope, verdict, and applicable cleanup. Correct the report or authorized artifacts. Reuse valid evidence; do not automatically rescan the repository or rerun successful commands. Repeat checks only for relevant changes, failures, or unresolved evidence. Disclose remaining gaps.

Output Contract

Report in the user's language, in this order; retain all five fields and state each fact once. Small results may use one line per field; omit empty tables and do not copy linked artifacts:

1. Result: Skill-specific verdict and supported outcome.
2. Scope: Reviewed/changed scope, exclusions, baseline, and material assumptions.
3. Evidence: Skill-specific fields below; distinguish facts, inferences, and unverified claims. Link artifacts; use tables when useful.
4. Verification: Checks/results, unavailable evidence, and applicable cleanup/external state.
5. Completion: Checklist: X/Y complete; Incomplete: None or each UNPROVEN item's reason, outcome impact, and exact next action; residual risks and required decisions.

Skill-specific evidence: Actual modules, boundaries, wiring, dependencies, configuration, and critical flows; distinguish current, target, transition, and drift. Findings need priority, affected boundary, evidence, consequence at current scale, unacceptable tradeoff, migration risk, and smallest safe next step; allow equivalent target shapes. Order evolution by prerequisites and preserve accepted exceptions.