Backend Systems Guidance

Backend Systems Guidance

This is a composable overlay, not a standalone workflow. Use alongside the repo's implementation skill when the change touches backend code.

Choose activity independently from backend depth. A review request is read-only unless the user also asks for remediation; use the architecture and reliability rules as inspection criteria and report findings instead of changing code.

This is the canonical backend overlay in this repo. It extends the thin baseline backend-guidance overlay with stronger guidance for non-trivial service boundaries, repositories, reliability, and trust boundaries.

Prefer it over backend-guidance when the task includes new endpoints or consumers, multi-layer refactors, repository or transaction work, auth or trust-boundary logic, or backend review that needs explicit testing and reliability checks.

Use the bundled references only when needed:

• references/trigger-evals.md for lightweight prompt checks when revising the trigger or scope

When to use

• the repo has server-side networked code such as HTTP handlers, gRPC methods, webhooks, queue consumers, or message producers
• the task adds or reshapes routes, controllers, services, repositories, middleware, or request-processing boundaries
• the task changes auth, authorization, validation, idempotency, retries, external requests, caching, or observability
• the task needs backend review beyond basic handler thinness, especially for security, data access, or missing tests

Not for

• HTTP client code, CLI tooling, offline batch scripts, or data pipelines with no request or consumer boundary
• frontend-only work
• threat modeling or security audit work where security should be the primary workflow skill

Do not fold infrastructure deployment workflows, outbound-client-only guidance, or full security-audit checklists into this skill. Keep this overlay centered on backend request and consumer systems plus their immediate reliability and trust boundaries.

Core workflow

1. Read the touched backend files and map the request or consumer path end to end: boundary, service logic, data access, external calls, and state changes.
2. Pick the activity and depth before changing code:
   • implementation when the user asks to add, change, refactor, harden, or remediate backend behavior
   • review when the user asks for findings or assessment; inspect the remaining steps without editing or requiring findings to be fixed
   • baseline backend change when the work is mostly a thin handler or small service fix
   • service-boundary change when responsibilities, data access, or dependency direction may need to move
   • reliability-hardening or review mode when the main risk is missing tests, auth gaps, retries, observability, or unsafe failure handling
3. Keep the boundary thin: decode input, authenticate, validate transport shape, call the shared policy/service path, map transport errors, and serialize output. Authorization must cover every entrypoint; business decisions belong in service code that can run without the transport layer.
4. Place persistence and external integrations deliberately:
   • repositories or data adapters own query shape, batching, and transaction details when that improves clarity or testing
   • services coordinate business rules, idempotency, retries, and side-effect ordering
   • handlers and controllers do not reach directly into ORM or network clients unless the change is truly trivial and stays trivial
5. Place trust and validation concerns in their owning layers:
   • decode, authenticate, and validate the external transport shape at ingress
   • enforce authorization in the earliest shared policy layer traversed by every relevant entrypoint, before business actions
   • enforce business invariants in service or domain code and persistence constraints in the data layer
   • set timeouts, retry rules, and destination allowlists for outbound calls
   • use structured logging, correlation identifiers, and explicit error mapping for observable failure paths
6. Choose the smallest test set that proves the change:
   • unit tests for service logic and decision branches
   • integration tests for handlers, consumers, repositories, and transaction behavior
   • auth and permission tests for protected flows
   • contract or schema tests when the change alters external API or event shapes
   • load or concurrency tests only for changed hotspots, queue throughput, or latency-sensitive paths
7. Review the result for boundary leaks, unsafe defaults, data-access inefficiency, and missing verification before finishing.

Decision rules

• Start with backend-guidance for ordinary backend edits. Use this overlay when the task needs stronger design pressure, harder review, or explicit backend quality gates.
• Keep handlers thin in responsibility, not by literal line count. If a handler or consumer owns business decisions, retries, transaction branching, or query orchestration, extract inward.
• Keep business logic transport-free. If testing a rule requires booting HTTP, gRPC, or queue infrastructure, the logic is in the wrong place.
• Add a repository or data-access interface when it reduces duplication, isolates non-trivial queries, helps transaction composition, or makes tests materially simpler. Do not add one for single-call trivial CRUD.
• Validate transport shape once at ingress, then pass typed data inward. Keep authorization, domain invariants, and persistence constraints in their owning layers; avoid duplicate checks only when they protect the same contract and boundary.
• Treat retries as a design choice, not a default. Only retry idempotent or explicitly deduplicated work, and pair retries with deadlines or backoff.
• Use idempotency keys or duplicate-detection for retried creates, webhook handlers, and queue consumers that can be re-delivered.
• Every outbound request needs a timeout and failure policy. For user-controlled destinations, apply allowlists or equivalent SSRF protections.
• Keep error handling explicit: domain code returns or throws domain-level failures; boundary code maps them to HTTP, gRPC, queue, or job semantics.
• Measure before adding caching. Cache only stable read paths with clear invalidation or bounded staleness.

Validation

For implementation, a backend change is done when, in addition to the base implementation skill's validation:

• handlers or consumers stay as boundary glue and delegate business decisions to testable service code
• data access and external I/O live behind clear seams when the change is non-trivial
• external transport shape, authentication, and transport-specific error mapping stay at the edge; shared authorization and domain invariants cannot be bypassed through another entrypoint
• retries, idempotency, timeouts, and failure handling are explicit where the change can duplicate work or call remote systems
• tests cover the changed behavior at the correct level, including integration coverage for boundary behavior and permission or failure cases when relevant
• new high-risk paths emit enough evidence to debug production behavior

For review, completion means prioritized findings map the affected request or consumer path, cite evidence, explain likely impact, and identify unverified behavior. Unfixed findings do not make the review incomplete.