Why CLI, Not Only Linters

Short answer

ESLint, Biome, and Oxlint look at one file (or a string-import heuristic). Cycles, inbound predicates, and promotion are properties of the whole graph. That needs a separate analyzer — dma check.

Linters in the editor remain useful. But CI without the CLI is a hole in architecture.

What the linter sees

DMA file-scoped rules:

import in the wrong layer direction (layer-direction)
feature → feature (feature-to-feature)
import bypassing public/ (public-api)
barrel index with re-exports (no-barrel)

ESLint does this precisely — via @derived-modular/boundaries and path resolution. Biome — by heuristics on string patterns.

What the linter does not see

Rule	Why the graph is needed
`no-cycle`	A cycle is a path A→B→C→A across multiple files
`feature-has-inbound`	Inbound is counted across all modules, not one import in one file
`service-no-inbound`	A service without consumers is a subgraph property
`dma doctor`	Evolution signals across the whole project

Example: a cycle services/billing → services/cart → services/inventory → services/billing — in each file one import looks legal, while no-cycle only sees the full graph. Or catalog imports checkout directly — ESLint catches feature-to-feature in the file, but the inbound predicate feature-has-inbound and chains through several modules — only check.

DMA's hybrid model

Editor     →  linter (fast, file-scoped)
CI         →  dma check (full graph)
Locally    →  dma doctor (soft signals)

One analyzer (@derived-modular/cli) — source of truth. Linters are thin adapters on top of @derived-modular/boundaries, not a second source of truth.

"We configured ESLint — architecture is under control" is an illusion until you run dma check in CI.

Why not one mega-ESLint rule

In theory ESLint could analyze the whole project. In practice:

linters are optimized for per-file work and per-file cache;
a full graph pass on every save is expensive;
cycles and inbound are a separate class of algorithms, not the AST of one module.

DMA deliberately moved graph audit into the CLI. Linters are for the IDE feedback loop.

Coverage matrix

Concern	ESLint	Biome	dma check
`layer-direction`	✓	heuristic	✓
`feature-to-feature`	✓	heuristic	✓
`public-api`	✓	heuristic	✓
`no-barrel`	✓	heuristic	✓
`no-cycle`	—	—	✓
inbound predicates	—	—	✓
doctor	—	—	✓

Practical recommendation

Wire up ESLint (or Biome / Oxlint) in the editor
Add npx @derived-modular/cli check . to CI
Once per sprint — npx @derived-modular/cli doctor . for growth signals

Why CLI, Not Only Linters

On this page