SuperAPI AI Directory

codependix-configure

General & other

View repository

Updated 10-16 03:12

About this skill

Writing a codependix configuration

Codependix produces nothing until a configuration tells it what to export and where. Every field is optional, and a workspace with no configuration file resolves every graph type for every project to target: "none" — a successful run that writes nothing. The configuration is the whole control surface; there is no flag that selects a graph type.

import { type CodependixConfiguration } from "@codependix/configuration";

const codependixConfiguration: CodependixConfiguration = { defaults: { imports: { markdown: { anchor: "codependix-file-imports" }, target: "markdown" }, nestjs: { markdown: { anchor: "codependix-nestjs-modules" }, target: "markdown" }, nx: { markdown: { anchor: "codependix-nx-projects" }, target: "markdown" }, }, exclude: ["fixtures/"], include: [""], projects: { widgets: { nx: { json: { path: "output/nx-graph.json" }, markdown: { anchor: "codependix-nx-projects", path: "docs/architecture.md" }, target: "both", }, }, }, workspace: { nx: { markdown: { anchor: "codependix-workspace" }, target: "markdown" }, }, };

export default codependixConfiguration;


## Where the file is found

`codependix.config.{ts,mts,cts,js,mjs,cjs,json}` is searched for **upward**
from the run's directory, and the first file found wins outright — nothing from
a further ancestor is merged into it. A workspace carrying both a TypeScript
and a JSON file gets the TypeScript one.

Two paths behave differently on purpose:

- **A path given with `--config` must exist.** A typo fails the run rather than
  quietly resolving every graph to `"none"`. It is resolved against the working
  directory first, then against the workspace root, so a configuration kept in
  a `configuration/` folder can be named relative to either.
- **An absent configuration is legal.** A workspace that never wrote one is not
  told to; it simply exports nothing.

## `defaults`, never `default`

The field that applies to every project naming no override of its own is
spelled **`defaults`**. A configuration module's default export is unwrapped by
name while loading, so a field called `default` would collide with that
unwrapping. A configuration that writes `default:` is not rejected — unknown
keys are stripped rather than refused, so a newer configuration still loads
under an older codependix — it simply has no effect at all, which reads exactly
like a tool that ignored it.

## `target` is explicit

Each graph type's export target is named, never inferred from which
destinations are present:

| `target` | Writes |
| -------- | ------ |
| `"both"` | The JSON and the Markdown destination together |
| `"json"` | The JSON destination only |
| `"markdown"` | The Markdown destination only |
| `"none"` | Nothing — and the default for every unset target |

That explicitness is what lets a project keep a `json` destination it is not
ready to write yet: leave the destination in place and set `target: "markdown"`.

The configuration is validated, and two mismatches are rejected rather than
ignored:

- A `"both"` or `"json"` target with no `json` destination.
- A `"both"` or `"markdown"` target with no `markdown` destination.

## Markdown has two shapes

```ts
// Anchor mode — spliced into a named block inside an existing file.
{ markdown: { anchor: "codependix-nx-projects" }, target: "markdown" }

// Anchor mode, in a file other than the default README.md.
{ markdown: { anchor: "codependix-nx-projects", path: "docs/graphs.md" }, target: "markdown" }

// Standalone mode — the export is the whole contents of the file.
{ markdown: { path: "output/graph.md" }, target: "markdown" }
  • Naming an anchor splices the export between two comment markers inside the file at path, which defaults to README.md — the common case is a project's own readme, so it is worth not repeating.
  • Leaving anchor unset writes the export as the entire contents of a standalone file, and path is then required: there is no default worth guessing for a file codependix owns outright.
  • A markdown destination naming neither is rejected, since nothing would say where the export goes.

A missing anchor block is auto-created on --write, so anchors do not have to be placed by hand first. See the codependix-triage skill for exactly where a new block lands and what still fails.

Paths in both destinations resolve relative to the project's own root, so one defaults entry describes every project without naming any of them. The whole-workspace graph resolves relative to the workspace root instead.

include and exclude

Both are lists of globs matched against a project's Nx name and its workspace-relative root, so either spelling works:

{ exclude: ["examples/**", "*-fixtures"], include: ["**"] }

include defaults to nothing and exclude to empty, so leaving both out means no project participates and the run exports nothing. Participation is always declared: write include: ["**"] to cover the whole workspace. A project is included when at least one include glob claims its name or root and no exclude glob claims either.

--projects and --tags widen this from the command line — a project either list names participates even when no include glob claims it, and exclude still wins over both. See codependix-export.

A configuration naming defaults but no include is the one to watch for. It exports nothing while still exiting zero, and --check boundaries judges every project regardless of include — or the projects --projects and --tags name — so the gate stays green. An export run warns when it happens — 🕸️ Selected no project to export.

An excluded project resolves to target: "none" for every graph type regardless of what defaults or its own projects entry says — so excluding a project takes one line rather than rewriting each of its overrides to "none".

A project override replaces, it does not merge

When a projects entry names a graph type, that entry replaces the defaults entry for that graph type outright — it is not merged field by field. A project turning its Markdown export off by omitting markdown should not have the default's markdown destination resurface underneath it.

The practical consequence: an override has to restate every destination it still wants, including the target.

{
  defaults: { nx: { markdown: { anchor: "codependix-nx-projects" }, target: "markdown" } },
  projects: {
    // Loses the Markdown export entirely — this is the whole nx entry now.
    widgets: { nx: { json: { path: "graph.json" }, target: "json" } },
    // Keeps both, because both are restated.
    gadgets: {
      nx: {
        json: { path: "graph.json" },
        markdown: { anchor: "codependix-nx-projects" },
        target: "both",
      },
    },
  },
}

Graph types the override does not mention still fall back to defaults; the replacement is per graph type, not per project.

The whole-workspace graph

workspace is separate from defaults and projects, and accepts only nx:

{ workspace: { nx: { markdown: { anchor: "codependix-workspace" }, target: "markdown" } } }

There is no workspace-wide NestJS or import graph to configure. The workspace graph is exported once for the repository, carries no per-project override, and is unaffected by include and exclude — excluding a project from per-project exports does not remove it from the workspace graph.

--projects and --tags are the exception, and they reach it through its node set rather than through its destination: a run naming a selection draws the graph over the projects it named, and writes it wherever workspace.nx says. See codependix-export for both flags.

projectGraph — reading a graph rather than resolving one

{ projectGraph: "artifacts/graph.json" }

A path, relative to the workspace root, to the JSON nx graph --file=graph.json emits. Named, it is read instead of resolving the process working directory's own graph — the only way to graph a workspace the process is not standing in. Left out, nothing changes.

The file is trusted: only a file that is not a project graph at all is refused. See codependix-export for what that does and does not cover.

Configuring a graph type a project does not have costs nothing

A project that is not a NestJS project never appears in the NestJS pass; a project with no tsconfig.json never appears in the TypeScript import pass. Neither is a failure and neither needs an override, so a single defaults block naming all four graph types is the normal shape of a whole-workspace configuration rather than an over-broad one.

boundaries — the rules the graphs are judged against

defaults and projects say where an export is written. boundaries says what the graphs must look like, keyed by the same four graph levels, and is gated by --check boundaries rather than by --check reports.

const codependixConfiguration: CodependixConfiguration = {
  boundaries: {
    imports: [
      {
        from: { path: ["**/*.types.ts"] },
        kind: "forbid",
        message: "Types are the leaf of a module.",
        name: "types-files-do-not-reach-services",
        to: { path: ["**/*.service.ts"] },
      },
    ],
    nx: [
      {
        from: { tags: ["type:application"] },
        kind: "forbid",
        name: "applications-are-leaves",
        to: { tags: ["type:application"] },
      },
      { kind: "acyclic", name: "no-project-cycles" },
    ],
  },
};

Three kinds:

Kind Reports
forbid An edge whose source matches from and whose target matches to
allow An edge leaving from for anywhere to does not claim
acyclic A cycle among the nodes nodes selects, defaulting to every node

A selector's four fields are all lists of globs, and they narrow each other — every field a selector states must match, while within one field one glob matching is enough:

Field Matches Carried at
id The node's identifier — a project name, a file path, or a module class name every level
path A workspace-relative project root, or a project-relative file path nx, imports, pythonImports
project The Nx project a node belongs to nx, imports, pythonImports
tags The node's Nx tags; one tag matching is enough nx

An access rule may also narrow which edges it judges, with edges: { implicit: false } sees only edges backed by an import statement — exactly what an @nx/enforce-module-boundaries depConstraint sees — and { implicit: true } sees only the ones an implicitDependencies entry declares. Unset judges both, which is the stricter reading. Only the Nx level draws an implicit edge.

Four things to know before writing one:

  • A selector naming a field its level does not carry matches nothing, not everything. A path rule at the nestjs level selects no module, because the NestJS module graph is class names with no file paths at all.
  • A selector stating no field is refused by the schema. It reads exactly like a typo, and reading it as "every node" would silently widen a rule to the whole workspace.
  • message is appended to the generated sentence, not substituted for it. Write it as the why; the what — the rule name and both endpoints — is always generated.
  • A level with no rules is never built. Declaring nestjs rules means every container is booted in preview mode, and declaring imports rules means a ts.Program per project. Declare only the levels you actually gate.

Write rules that already hold. A rule that arrives red is a backlog rather than a gate, and a red pipeline nobody can act on teaches people to ignore it. Verify a candidate rule against the whole workspace before committing it.

Which project a finding fails

A finding is charged to projects, and a run fails only when one is charged to a project it judges. The rule kind decides who that is:

Finding Charged to
forbid match, or an edge no allow rule covers The project owning the edge's source, and no other
acyclic cycle Every project owning a node on the cycle, so a cycle across A and B fails both
Container that cannot boot The project whose container failed, named with the owner of the failing class when that is another project

A project that merely depends on a project with a cycle or a forbidden edge is not failed: it sees the finding as a note, "in dependency". So a rule is written with the source project in mind — it is the one that has to change. A container that cannot boot is the exception: a dependent whose own container fails because a dependency's class could not evaluate fails too, since its container really cannot boot, and its failure names that dependency as the owner.

The per-project gate

In an Nx workspace the @codependix/nx plugin infers a codependix-gate target for every project described by a project.json, except the workspace root. It runs --check boundaries --projects <project> over that project and its dependencies. A project Nx infers from a package.json alone gets no gate until it is given a project.json. Register the plugin once in nx.json:

{
  "plugins": [
    {
      "plugin": "@codependix/nx",
      "options": {
        "configurationPath": "configuration/codependix.config.ts",
        "gateTargetName": "codependix-gate"
      }
    }
  ]
}

configurationPath names the one configuration every gate reads; omit it and codependix.config.ts, then configuration/codependix.config.ts, are searched. Nothing else is configured per project: the rules in boundaries judge every project the same way. See codependix-export for running a gate.

The gate runs the command line under the @swc-node/register hooks the plugin registers itself, so the workspace installs nothing for them. The workspace root does need a tsconfig.json that emits decorator metadata, because the hooks read it from the working directory and NestJS constructor injection depends on it.

The gate is cached. A project's own codependix.config.* is one of its cache inputs, so editing it re-runs that gate. So is the codependix command line itself: its package.json and sources when it is a package of the workspace, or its installed version when it comes from a registry. So is the workspace root's tsconfig.json, which the loader reads. Test files are not inputs, so a test-only edit to the tool invalidates no gate beyond that package's own and its dependents', while any other edit re-runs them all. A package the plugin cannot resolve loses only its own inputs, and Nx's logger warns naming it.

A gate whose projects or tags select anything other than its own project — on the command line, in the target's options, or in a named Nx configuration — is never replayed from the cache, because it judges projects its inputs do not cover. A gate judging only its own project is cached as usual.