Coupling Map

Answers one operational question: which files should be refactored first to reduce systemic risk. A file qualifies when both conditions hold — it does too much, and enough of the system depends on it that breaking it breaks several other parts.

The output is a standalone HTML report. It is diagnostic, not a gate: it fails no build and blocks no commit.

References, loaded on demand, all of which travel with this skill: references/config-schema.md (every field of arch.config.json), references/report-contract.md (the shape of architecture.json), references/calibrating-cuts.md (deriving thresholds for a new codebase), references/rejected-metrics.md (read before proposing any metric or axis).

The full design argument lives in docs/Skill_Coupling_Map.md of the sdd-skills repository, which is not part of an installed copy. Everything needed to run is in the folder you are reading.

The point of the split is cost. The model inspects the tree exactly once, to write a config of about thirty lines. From then on it reads a single JSON file. On the codebase this was built against that is one file instead of 959, and the numbers are identical either way because the script computed them, not the model.

INPUT

Nothing required. Optionally a path to scope the run, --config / --out to override locations, and --open to launch the finished report in the default browser.

OUTPUT

• architecture-report/index.html — the report, opened straight from disk. The script prints it as an absolute file:// URL, which most terminals turn into a link and which pastes into a browser as-is. Hand that link to the reader; a relative path leaves them working out where the file went.
• architecture-report/architecture.json — the data, and the only file this skill reads on later runs
• A short spoken summary: the queue, what moved since the previous run, and what needs a human

PHASE 1 — Preflight

Three checks, in this order. All three passing means skip straight to Phase 3.

node -e "require.resolve('madge')" && echo "madge ok"
test -f scripts/arch-report.mjs && echo "script ok"
test -f arch.config.json && node -e "JSON.parse(require('fs').readFileSync('arch.config.json','utf8'))" && echo "config ok"

Never reinstall or regenerate what already answers. If only some pass, do only the missing part of Phase 2.

PHASE 2 — First run only

2.1 Ask permission before installing anything

Say what will be installed, where, and that nothing else changes. Wait for an answer.

2.2 Install madge at the repository ROOT

npm install --save-dev madge

At the root, not inside a sub-app. A monorepo app frequently carries peer conflicts of its own, and installing there drags the tool into somebody else's dependency argument.

Graphviz is not needed. It is required only by madge --image, which this never uses — the script reads obj() and draws the chart itself. That removes the one system binary most likely to fail to install.

2.3 Copy the script into the project

cp <skill>/scripts/arch-report.mjs scripts/
cp -r <skill>/scripts/coupling-map scripts/

Copy, never rewrite. See the rules.

2.4 Write arch.config.json by reading the project

This is the one moment the tree is inspected. Load references/config-schema.md and determine:

• the apps — every independently resolved source tree, one entry each
• extensions and exclude — and remember that migrations, fixtures, generated JSON and test folders are not application code
• tsConfig — for any TypeScript app. Omitting it is the most damaging single mistake available here
• importPrefixes — the non-relative prefixes the project resolves internally, read out of baseUrl and paths
• domainFrom / layerFrom / layers / layerOrder — how this project names things, which is rarely how the last one did
• leafLayers — layers meant to be called, never imported

Write it with a file tool, never a shell heredoc: the exclude patterns carry backslashes.

2.5 Calibrate the cuts

Placeholder thresholds first, then follow references/calibrating-cuts.md. The values from another project do not transfer. Aim for a queue somebody can work through — on a 959-file codebase, twelve.

2.6 Run it and check the run before believing it

node scripts/arch-report.mjs --open

--open launches the report in the default browser once it is written. Without it the script only prints the link, which is the better default when the run is not being watched.

Confirm, in order:

1. Coverage sits between the floor and the ceiling. Below, the graph lost edges — check tsConfig. Above, the denominator is blind — check importPrefixes. The script aborts on either and writes nothing.
2. Σ Ca* equals Σ Ce*. Printed on every run. They count the same pairs from opposite ends, so a mismatch is a traversal bug, never a property of the code.
3. The (sem domínio) count is small. A large bucket means the taxonomy does not fit this project.
4. Read the queue and sanity-check it. If it names small widely-used utilities, painLoc is too low.

PHASE 3 — Every run

node scripts/arch-report.mjs

Add --open when a person is at the screen and wants to look; leave it off otherwise, and give them the printed link instead.

Then read architecture-report/architecture.json and nothing else. Do not open source files. Every number the answer needs is already computed.

Report:

• the queue — the pain-zone files in order, with Ca*, LOC and any second detector
• what moved, if a previous architecture.json is available: files entering or leaving the queue, cycles opened or closed, coverage drift
• the counters — amplifiers, controllers used as dependencies, direction violations, cycles, orphans
• what the numbers cannot decide. The report ranks by risk; whether a given refactoring is worth doing now is a human call

When reporting, say Ca* and Ce* in words. The reader is not required to know the vocabulary, and the sidebar of the report spells both out.

RULES

Never regenerate the script. It is versioned in this skill and copied verbatim. A script written fresh per project produces as many divergent variants as there are projects, none reviewed, none fixable centrally, and all of them outside the repository the team receives skills from.

Never read the source tree in Phase 3. Reading hundreds of files to produce numbers the script already computed is exactly the cost this skill exists to avoid.

Never lower a coverage limit to make a run pass. Both limits mark a broken run, not a strict one. Fix the resolution or the denominator.

Cuts are absolute and frozen; never percentiles. A percentile always finds the same share of files, so a report that exists to show improvement becomes incapable of showing it.

High Ca alone is not a defect. A 43-line utility that 131 files import is correct. Say so when reporting, or someone will refactor the foundation of the system. That risk is the reason the chart has two axes.

Read references/rejected-metrics.md before proposing a metric. Abstractness, distance from the main sequence, LCOM, a weighted God-class score and a composite health rating were each implemented or measured against a real codebase and removed. Every one is the obvious next suggestion.

Leave the report where it can be reopened. It is a single self-contained file that makes no external request; it needs no server and no network, and it opens from file:// — there is nothing to serve and nothing to automate a browser for.

Never open the report without being asked. --open exists because the reader asked for it. A run that nobody is watching should not seize the screen.

Edge cases

A single-app project. One entry in apps. Nothing else changes.

No tsconfig.json. Fine for a JavaScript project — omit tsConfig. If the project is TypeScript and has none, say so rather than running without it; coverage will abort anyway.

Apps that genuinely import each other. The default assumes separately deployed trees with no edges between them. If they do share code, they belong in one app entry, or the graph will be cut where the imports really run.

A first run with no cycles and no violations. A real and good result. Report it as such rather than hunting for something to flag.

A previous run used different cuts. meta.cuts records them. Comparing across a cut change compares two different questions — say so instead of reporting movement that is an artefact.

--open does nothing, or opens the wrong application. The report is still written and the link still printed — the flag is a convenience over the operating system's default handler, and a headless machine or a container has none. Never treat a failed open as a failed run.

The project already has a scripts/ entry point of its own. Copy under a subfolder and pass --config and --out explicitly rather than overwriting anything.