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 contracts, 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 diff-scoped cleanup pass (tidy), or structural review of a local diff (pr-reviewer).

Contents

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

Modes

Pick by the problem, not by the artifact.

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 the two look equally right, prefer Deepen. A user reporting “agents keep using the old pattern” is describing a Harden symptom, but if the cause is one concept living in two places, quarantine only freezes the duplicate in place: Deepen deletes it and Harden holds the line until that lands. Harden alone is the right answer when the old thing genuinely has to stay.

When you cannot write to the repo (no checkout, read-only request, or the user asked a question rather than for 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 for simplicity: 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, not a simplification. Where a corner is genuinely cut on purpose, mark it with its ceiling and upgrade path rather than leaving the next reader to guess whether it is finished.

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 craftsmanship.md’s surface-area budget: 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 you made 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 contracts, each naming its enforcement (lint boundary rule 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(); never thread a ctx parameter through business functions. Implementation in references/api-design.md.
   • Require explicit auth policy per RPC method at registration; a method without one fails registration, never defaults to open.
   • Keep auth, logging, errors, and context in shared middleware, not per-handler code.
5. Define frontend boundaries:
   • Default to Server Components; add "use client" only at leaf components needing interactivity.
   • Server state in TanStack/Connect Query; client state in component state; MobX only for cross-cutting client state that fits neither.
   • Each piece of data has exactly one owner. Never mirror server data into useState or sync two stores with useEffect; both are the red flag that ownership is unclear.
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; do not defer them to a later polish pass.
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.
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.

Harden mode (make it stick)

Two halves: guardrails stop the wrong thing landing, wayfinding makes the right thing cheap to find. Both exist because agents arrive by grep, not by reading docs, so the warning has to live where they land and the rule has to be an exit code rather than a sentence someone might recall.

Steps 1 to 3 always run. Steps 4 to 6 run only when their condition holds, and a request to add one check stops at step 3. Running all six for every request loads most of the bundle and is the failure this mode is most prone to.

1. Survey what exists. Package scripts, CI steps, hook config, lint config, the instruction file, the docs index. Find two things: checks that run locally but do not gate the merge, and dormant config nobody invokes. Wire the first, delete the second (references/contagion-markers.md).
2. Choose checks by the failure they prevent, never by tool popularity. Categories and tools in references/guardrail-tooling.md. Pick the two or three failures this repo actually exhibits; installing the full set at once forces the weakest enforcement rung on all of them.
3. Install each check: pick an enforcement rung for the violations that already exist (references/enforcement-ladder.md), ship it green, then prove it bites (run it, break it on purpose, watch it fail with a message naming the fix, revert). Wire it into both a pre-commit hook and CI.
4. Wayfinding per references/wayfinding.md: naming and locality, the add-a-new-X recipe file, the trust-labeled docs index, one canonical instruction file.
5. Contagion markers per references/contagion-markers.md: anything an agent must not copy or must not edit gets a greppable marker at the code site naming what to use instead.
6. Runtime ergonomics: verification tiers in references/verification-tiers.md; session hooks, permission allowlists, and review gating in references/agent-runtime.md.

Validation loop

Run the items matching the modes you ran, and record results in the output. Each needs evidence; “looks consistent” is not a pass. An item that cannot execute yet, because nothing is installed or the repo is not writable, is recorded N/A with that reason. Silently passing it is how an unenforced contract ships looking verified.

1. Consistency (Design, Deepen): naming, module contracts, and middleware rules read the same across every service. Evidence: a contradiction scan with zero findings.
2. Enforceability (all): every contract names its lint rule, type check, or test. Evidence: an enforcement note per contract, and for any check actually installed, the pass, then fail on a deliberate violation, then pass after revert.
3. Operability (Design): observability, health checks, and a rollback path per deployable surface. Evidence: the rollout section names each.
4. Quality gates (whenever code changed): the repo’s lint, type-check, and targeted tests (npm run lint, npm run check-types, npm run test --workspace=<pkg> or equivalents). Evidence: passing output.
5. CI and local agree (Harden): the CI step invokes the same umbrella command a developer runs, or the difference is deliberate and stated.
6. No dangling pointers, and a recipe works cold (Harden): a grep proving every path named in the docs index and instruction file exists, plus a fresh-context agent following one add-a-new-X recipe end to end with no further guidance.
7. Net simplicity (all): the result leaves a reader less to hold, not more. Evidence: the net change in files, surfaces, and exported names, with every increase named and paid for by what it removed elsewhere; plus, for each layer, port, or indirection introduced, the second caller or implementation that made it real. An architecture pass that only adds has failed this check even when every other item passes.

On failure: fix the brief, the conventions, or the wiring, then re-run the loop.

Output template

Design mode produces this brief. Deepen mode’s ranked-opportunity template is in references/deepening-existing.md. Harden mode’s output is the wiring itself plus the loop’s evidence, not a document.

# Architecture brief
 
## Context and constraints
## Repo shape
## Backend module contracts
## Request context and middleware policy
## Frontend boundaries
## Testing strategy
## Quality bar and surface-area budget
## Rollout and rollback plan
## Open risks and follow-ups

Excuses

Each rebuttal redirects to the step being skipped; none of them repeat a gotcha below.

Gotchas

Design and Deepen

• Don’t default to microservices for teams under 5: each service adds a deploy pipeline, contract versioning, and on-call surface. Start with a modular monorepo; split when a boundary is proven by team or scale pressure.
• Don’t put app-level deps in a monorepo’s root package.json: hoisting hides missing declarations, so an app builds locally but breaks deploying alone. Each app owns its deps.
• Don’t define module contracts (handler/service/dao) without enforcement: an unenforced contract decays at the first deadline. Add an import-boundary lint rule (e.g. dao may not import handler) the day you write it.
• Don’t thread a ctx parameter through business functions instead of AsyncLocalStorage: every signature grows, and adding one field later touches hundreds of call sites.
• Don’t place "use client" at page or layout level: it converts the whole subtree to client rendering and forfeits streaming and direct server data access. Push it to leaves.
• Don’t propose a big-bang rewrite in Deepen mode: migrate one vertical slice, verify it, then generalize.
• Don’t extract to packages/ early: wait until 3+ apps need the same code; a premature shared package couples release cycles for nothing. The exception is the contract two surfaces already share (generated types, the RPC schema, branded IDs): that is not speculative reuse, it is the interface between them, and it belongs in a package at two apps.
• Don’t finalize a brief without a rollback plan per change: an irreversible decision needs a documented fallback before it ships.
• Don’t dual-write to a database and a queue/webhook without an outbox (or CDC): one side commits, the other fails, and you silently lose or fabricate a notification. See references/distributed-correctness.md.
• Don’t enforce an externally-forceable invariant by construction (unsigned type, hard CHECK): when the outside world forces the state, the system crashes or clamps instead of recording it. Represent it, detect it post-factum, recover explicitly.
• Don’t scan a whole codebase for deepening opportunities: without git hot-spot scoping the list fills with modules nobody touches, and every entry on it is speculative by definition.
• Don’t write a glossary entry for a general programming concept: a glossary earns its keep on the terms this product argues about, and padding it with “timeout” and “retry” trains everyone to skim it.

Harden

• Don’t hand-roll a shrink-only baseline (*-ratchet.mjs plus *.baseline.json): it was built once and deleted, because knip, eslint, madge, and jscpd all ship native ignore, allowlist, and warn mechanisms that do the same job with no code to maintain.
• Don’t land a rule red: agents and humans both learn the check is noise, and the next person adds --no-verify instead of a fix.
• Don’t leave dormant config or an unused devDep in place: an agent reads it as live convention and extends it, and one pointing at a missing file yields a confident wrong answer instead of an error.
• Don’t index a doc that does not exist: agents cite confidently, so a dangling pointer is worse than a missing one.
• Don’t write add-a-new-X recipes from memory: a recipe naming a moved file sends the agent somewhere wrong, and it will not doubt the doc.
• Don’t rely on pre-commit hooks alone: they are not guaranteed installed on a fresh clone or in a worktree, which is exactly where agents run.
• Don’t put a marker only in docs/legacy.md: an agent that arrived by grep never opens it. The marker goes in the frozen file.
• Don’t generate the instruction file wholesale: generated files mostly restate documentation the agent can already read, cost 20 to 23% more per task, and measurably lose to hand-written ones. Hand-curate it and update it in the PR that changes the convention.