Typescript CLI Architecture

TypeScript CLI / Agent Project Architecture & Testing

Overview

Group code by feature, not by technical layer. Give each subsystem one central registry. Co-locate everything a feature needs — including its tests. Distilled from the Claude Code CLI source (<workspace>\claude-code-main\src) plus 2026 industry practice (vertical-slice / modular-monolith, Vitest testing honeycomb).

Core principle: adding a feature should mean adding a folder, not editing files across 5 directories. Deleting a feature should mean deleting one folder.

Language-agnostic parent: baseline-dev-architecture (the cross-language invariants this skill instantiates for TypeScript/Bun).

When to Use

• Starting a new TS/Bun/Node CLI, agent runtime, or backend service.
• Deciding where a new tool, command, route, or module belongs.
• Setting up or fixing the test harness, smoke tests, or CI gates.

When NOT to use: tiny scripts (<10 files) — a flat src/ is fine. Rust projects → use rust-workspace-architecture instead.

Canonical Folder Layout

src/
  main.ts(x)              # CLI entry: arg parsing + routing ONLY (thin)
  app.ts                  # builds/configures the program (testable, no side effects)
  <Subsystem>.ts          # the INTERFACE/contract (e.g. Tool.ts)
  <subsystems>.ts         # the REGISTRY (e.g. tools.ts → getAllTools())
  config/
    env.ts                # Zod-validated env, fails fast at startup
  tools/                  # one folder per feature (vertical slice)
    BashTool/
      BashTool.ts         # the feature's main impl (matches folder name)
      bashPermissions.ts  # concerns as sibling files, prefixed by feature
      bashSecurity.ts
      BashTool.test.ts    # unit test, CO-LOCATED
      index.ts            # public API — the only legal import surface
  commands/  screens/  services/   # same registry+folder pattern per subsystem
  shared/ (or utils/, lib/)        # cross-feature helpers ONLY
  types/                  # shared/branded types

• Registry pattern: tools.ts imports each tool and exports a getter (getAllTools()); feature-gating/conditional includes live ONLY in the registry, never scattered.
• Interface contract: every feature of a kind implements one typed contract (Tool<Input, Output> with a Zod input schema + call() + permission/render hooks). No partial implementations.
• Co-located concerns, no deep nesting: all of a feature's files are direct children of its folder, prefixed by feature name (bashPermissions.ts). No inner lib/.
• Split entry from config: main.ts only parses args + routes; app.ts builds the configured program so it's testable without launching.

Naming & Boundaries (quick reference)

Boundary rules (enforce with eslint-plugin-boundaries):

• A feature NEVER imports another feature directly. The index.ts barrel governs reuse within a subsystem; any cross-feature or cross-subsystem need (a command needing a tool, shared auth/session state) goes through shared/ — extract the shared piece there and call it explicitly.
• UI/screens import services; services never import UI.
• Validate at the boundary (Zod), translate errors centrally, keep one typed config object.

Testing Strategy

Framework: Vitest (2026 default — native ESM/TS, ~200ms cold start, Jest-compatible API).

On Bun: bun test also works (same *.test.ts naming; bun build --compile makes a single-file binary) but its API is not 100% Vitest/Jest-compatible — pick one runner project-wide. Vitest runs fine under Bun and is the safer default for Jest-API parity.

Shape = testing honeycomb, not a naive pyramid: integration tests give the best bug-catch ROI. Mock only at boundaries with no interesting runtime behavior; use real infra (DB/queue/cache via Testcontainers) where implementation details bite.

• Separate runs via Vitest projects: unit (watch locally) vs integration (CI, single-thread, long timeout) — vitest run --project unit|integration.
• Coverage gate (provider: 'v8'): fail CI below ~lines/functions 80, branches 75. Don't chase 100%.
• AAA structure; reset mocks (vi.clearAllMocks); fake time with vi.setSystemTime.

Smoke tests (run BEFORE the full suite / before expensive jobs)

A smoke test proves the thing boots and the happy path works — fast, no deep assertions:

1. Build/bundle succeeds (bun build / tsc --noEmit / vite build) — the cheapest real gate.
2. CLI boots: invoke the built binary with --help and --version; assert exit 0.
3. One end-to-end happy path: simplest real command produces expected output.

Wire as npm run smoke and run it first in CI; only proceed to unit→integration→e2e if smoke is green.

// package.json scripts
"smoke":            "bun build src/main.ts --outdir dist && node dist/main.js --version",
"test":             "vitest run --project unit",
"test:integration": "vitest run --project integration",
"test:e2e":         "playwright test"

CI order

type-check → smoke → unit → integration → e2e, unit and integration as parallel jobs so signal arrives fast.

Common Mistakes

Reference

Structural patterns observed in <workspace>\claude-code-main\src (feature folders, Tool.ts contract, tools.ts/commands.ts registries, co-located concerns, ESM .js imports). That source is a bundled architectural reference with no test suite — the testing section here is sourced from 2026 Vitest/honeycomb practice, not from it.