dependency-hygiene

Dependency Hygiene

Architecture is mostly the shape of the dependency graph. This subskill keeps that graph legible: dependencies point one way, modules expose less than they contain, every third-party package pays rent.

Use This Skill When

• Auditing a codebase's import/module structure ("untangle these dependencies")
• A change adds an import that feels like it points the wrong way
• Deciding whether to admit a new third-party dependency or upgrade one
• Carving a module boundary: what is public, what is internal
• Cycles, god-modules, or "everything imports utils" complaints

Do Not Use This Skill For

• Within-function clarity — code-readability
• Retired-interface leftovers — cruft-cleanup
• Repo-wide topology documentation (lay-and-land pillar) — /th-projects

Core Rule

Dependencies point from less stable to more stable, never back. Policy depends on mechanism, callers depend on interfaces, leaves depend on core. Any edge pointing the other way — and any cycle, which is both directions at once — is a finding regardless of whether the code currently works.

The Bar

Reviewable expectations — cite the one violated, with module/import evidence:

1. One direction per layer boundary — Name the layers the project implies (e.g. domain ← application ← interface/infra); verify every import crosses boundaries in the sanctioned direction. Invert upward imports via an interface owned by the lower layer, never wave them through.
2. No cycles, at any granularity — Module- or package-level cycles make units untestable in isolation and changes non-local. Break them by extracting the shared piece downward or inverting one edge; never by deferring imports inside functions to hide the cycle from tooling.
3. Public surface is deliberate and minimal — A module exports what its contract promises, not whatever happens to be defined. Reaching into another module's internals (deep imports, private-by-convention access) couples to accidents. Shrinking surface beats documenting it.
4. No god-modules — A utils/common/helpers that everything imports becomes a cycle hub and change magnet. Split by domain; a helper used by one consumer moves to that consumer.
5. Third-party dependencies pay rent — Admitting a package is an ownership decision: you adopt its security surface, upgrade cadence, transitive tree. Before adding one, state what it saves over ~50 lines of owned code, check maintenance health, prefer the standard library. Wrap volatile or heavy dependencies behind an owned interface at the boundary so they stay replaceable; don't wrap stable, idiomatic ones. For seam placement by dependency category, the deletion test, and the test strategy per category, read references/seams-and-dependencies.md.
6. Versions are pinned, upgrades are deliberate — Commit lockfiles; make each upgrade its own reviewed change with the changelog read, not a side effect of unrelated work. Record new dependencies wherever the host project documents them (for example, the dotfiles repo that ships this skill uses a DEPENDENCIES.md).
7. Test/dev dependencies stay out of production chains — Fixtures, fakes, and tooling must not be importable from shipped code paths.

Workflow

1. Map before judging: extract the actual import graph (language tooling, or rg/grep over import statements) for the scope under review. For audits, render the module graph — ../excalidraw-diagram/SKILL.md if available, otherwise Mermaid — so direction violations are visible, not asserted.
2. Establish intended direction: from the project's stated layering (lay-and-land docs, README) or, absent that, from stability — what changes most often must depend on what changes least.
3. Walk the edges: flag wrong-direction imports (1), cycles (2), deep imports (3), god-module fan-in (4). For each, propose the concrete inversion or extraction, smallest first. Judge suspect wrappers and layers with the deletion test (references/seams-and-dependencies.md).
4. For each third-party package in scope, apply the rent test (5); check pinning/recording (6) and test/prod separation (7).
5. Apply small in-scope fixes (move a helper, narrow an export, pin a version) directly; structural inversions get a sequenced proposal with the cycle-breaking step first.

Trigger Sanity Check

• Should trigger: "core imports the HTTP layer and I can't test it alone — untangle this."
• Should trigger: "should we pull in this package or write the ~40 lines ourselves?"
• Should trigger: "everything imports utils and every change touches it — is that fine?"
• Should not trigger: "Rename these variables" or "update the lockfile after CI told us to."