Codebase Architecture

Codebase Architecture

Decide a TypeScript codebase's structure, improve it where change has become expensive, and make it hold. The target is a codebase a reader can hold in their head: few surfaces, one canonical way to do each job, and behaviour where you would first look for it.

• IS: folder structures, module interfaces, request context and middleware pipelines, frontend/backend boundaries; architecture briefs; domain language and decision records; domain-informed deepening; guardrail tooling, CI gates, and agent wayfinding.
• IS NOT: scaffolding a new repo (scaffold-nextjs for a Next.js turborepo, scaffold-cli for a TypeScript CLI), multi-tenant domain/isolation/routing (multi-tenant-architecture), the content of AGENTS.md itself (agents-md), a plan for one feature (planning), a diff-scoped cleanup pass (tidy), structural review of a local diff (tidy), or a product's own verify CLI, doctor command, and feature map (app-verification, call it by name).

Contents

• Modes
• References
• Design mode (new codebase)
• Deepen mode (existing codebase)
• Harden mode (make it stick)
• Validation loop
• Output template
• Gotchas
• Related skills

Modes

Pick by the problem, not by the artifact, and say which you picked.

Modes compose, and running more than one is normal. Design ends in Harden, because a contract with no check is a suggestion. Deepen ends in Harden, so the new seam cannot decay back. Harden runs alone when the structure is already right and only the enforcement is missing, which is the common case in a repo that agents work in.

When two look equally right, prefer Deepen. "Agents keep using the old pattern" sounds like Harden, but if the cause is one concept living in two places, quarantine only freezes the duplicate: Deepen deletes it and Harden holds the line until that lands. Harden alone is right when the old thing genuinely has to stay.

When you cannot write to the repo (no checkout, read-only request, or a question rather than a change), each mode's output degrades to its plan: the brief, the ranked opportunities, or the named checks with their rungs. Say which checks remain unproven, since none of them are wired.

As simple as possible, no simpler. Every mode cuts: surfaces in Design, concepts in Deepen, dual paths and dormant config in Harden. The floor does not get cut: validation at trust boundaries, error handling that prevents data loss, security, accessibility, observability on anything deployed, and whatever was explicitly asked for. A simplification that reaches one of those is a bug. Where a corner is cut on purpose, mark it with its ceiling and upgrade path rather than leaving the next reader to guess whether it is finished.

Copy this to track progress, and delete the lines for modes you are not running:

Codebase architecture progress:
- [ ] Modes chosen and stated (Design / Deepen / Harden)
- [ ] Design: assumptions stated, repo shape, every contract names its check, brief written
- [ ] Deepen: git hot spots, glossary, ranked opportunities with paths, one slice migrated
- [ ] Harden: existing checks surveyed, checks picked by failure, each landed green and proven to bite
- [ ] Validation loop run for the modes used; evidence recorded, N/A items named

References

Load only when the condition applies.

Design mode (new codebase)

Before any of this, ask whether each surface needs to exist. A module, service, app, or entrypoint that could be a folder in something that already ships is the cheapest architecture decision available, and the only one that stays cheap. Every surface you do accept pays the relationship cost in the surface-area budget (brief-conventions.md): name its owner, tests, observability, and deletion path before it goes in the brief.

1. Constraints first: product scope, team size, compliance/security, expected scale, deploy targets, required integrations, and quality bar. A one-line request supplies none of these, so assume the common case, state every assumption in the brief's first section, and invite correction. Ask outright only where a wrong guess would restructure the brief rather than extend it, which in practice is multi-tenancy and whether the API is public.
2. Choose repo shape:
   • apps/ for deployable surfaces (api, web, admin).
   • packages/ for shared libraries (shared, ui, icons, auth, proto).
3. Define backend module interfaces, each naming its enforcement (import-boundary lint or type check):
   • handler: transport only.
   • service: business orchestration.
   • dao: database access only.
   • mapper: DB/proto/domain transformations.
   • constants and types: module-local contracts.
4. Define request context and middleware:
   • Carry tenantId, userId, and traceId in an AsyncLocalStorage-backed RequestContext, initialized in every entrypoint (RPC, HTTP, jobs, CLI) and read via getContext(). A threaded ctx parameter grows every signature, and adding one field later touches every call site. Implementation in references/api-design.md.
   • Require an explicit auth policy per RPC method at registration; a method without one fails registration rather than defaulting to open.
   • Keep auth, logging, errors, and context in shared middleware, not per-handler code.
5. Define frontend boundaries (Next.js App Router default):
   • app/ holds routing files only (page, layout, loading, error, route). Domain code lives in src/modules/<name>/ behind its root files; UI private to one route goes in a _components/ folder beside its page. A page that grows logic moves it to a module, not to a sibling file in app/.
   • Server Components by default; "use client" at the interactive leaves. Where a client wrapper needs server-rendered content, pass it in as children.
   • Server state in TanStack or Connect Query; client state in component state; MobX only for cross-cutting client state that fits neither. Each piece of data has one owner: server data mirrored into useState, or two stores synced with useEffect, is the sign that ownership is unclear.
   • proxy.ts (Next 16's name for middleware.ts) handles redirects, rewrites, and headers. Authorization is decided inside each route handler and Server Function, because a matcher-excluded path skips the proxy and Server Functions post to their page's route.
6. Testing and release:
   • Unit tests stay DB-free; integration/E2E run in parallel with dynamically generated IDs so runs never collide on fixtures.
   • Release in small, complete, reversible vertical slices with a rollback plan per change.
   • A slice is complete only when reliability, error paths, observability, and user-facing states are covered; deferring them to a polish pass is how they never ship.
7. Every contract in the brief names the lint rule, type check, or test that catches its violation, then continue into Harden mode to wire them.

Deepen mode (existing codebase)

Goal: domain-informed deepening, not a rewrite. Load references/deepening-existing.md for the analysis method, opportunity patterns, and output template.

1. Map the domain language and decisions. Read CONTEXT.md, docs/adr/, or local equivalents if present, then read the code for entities, actions, and contexts as the team names them. Note divergence (one concept, three names; or one name, three concepts). Format and ADR rules in references/domain-language.md.
2. Scope the scan by where change lands. Deepening pays off on code that keeps changing, so git log --oneline over a good stretch of history first and weight the files that keep coming up. An unscoped scan drifts into speculative cleanup.
3. Find deepening opportunities. Look for anemic concepts, shallow modules, leaking seams, naming divergence, duplicated concepts, primitive obsession, misplaced logic, and tests forced past the public interface. Record each with file paths, never a vague smell. Check deletion first on every candidate: a concept with no live caller, a flag whose branch never runs, a layer with one implementation. Deleting it is the deepening, and it is the only move that cannot make the codebase harder to read.
4. Rank by leverage. Prefer opportunities that pass the deletion test, localize named future changes, have low churn, meet a current requirement, and have a viable testing seam. Rank candidates before designing target interfaces; drop speculative cleanups. For the top candidate, design the interface twice before picking one.
5. Migrate one vertical slice first. Prove the highest-leverage move end to end through one slice before generalizing.
6. Enforce the new seam with lint, type, or test checks so it cannot decay, then roll out module by module. Continue into Harden mode for the enforcement rung and the check-bites test.