agentsclimarketplace

Dependency hygiene

Skill zakariaf/Flutter-Skills/skills/dependency-hygiene

Enforces pubspec/lock discipline — caret ranges in pubspec.yaml with a committed pubspec.lock as the only pin, a separately-recorded SDK version string, a version-pinned very_good_analysis include (whose missing file fails default analyze, or silently drops the ruleset where warnings are non-fatal), transitive-tree auditing before adding a package, a dependency gate that refuses network/telemetry/crash/ads/heavy-transitive deps by policy, and vendoring any bus-factor-1 native plugin behind an interface into third_party/. Use when running dart pub add/get/upgrade/outdated/deps, editing pubspec.yaml or pubspec.lock, bumping the Flutter/Dart SDK, choosing or rejecting a new dependency, removing a package, or auditing what a dependency drags in.From its SKILL.md

Install
npx -y skills add zakariaf/Flutter-Skills --skill dependency-hygiene

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

2 things to look at

  • 22 days oldThe repository was created 22 days ago. New is not bad, but a brand new repository carrying a familiar-sounding name is the shape a typosquat arrives in, and there has been no time for anyone else to find a problem with it.
  • 0 stars0 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

SKILL.md

11.9 KB, ~2.7k tokens by cl100k_base, as published. Nobody here has run it

Dependency hygiene

Every dependency is a permanent liability someone else may have to service. Optimise for resolves and builds years from now, not for latest. This skill governs pubspec/lock mechanics, the gate a new package must pass, and the escape hatch when a critical package rots.

Read the reference for the task at hand:

  • references/dependency-gate-and-audit.md — the refuse/accept gate, transitive auditing, licence recording, upgrade discipline.
  • references/sdk-pin-and-lint-include.md — recording the SDK version and the silent-lint-disable trap on any SDK bump.
  • references/vendoring-behind-an-interface.md — vendoring a bus-factor-1 native plugin into third_party/ without touching lib/.

Run scripts/audit-deps.sh before a PR that changes pubspec.yaml.

Non-negotiable rules

  1. Caret ranges in pubspec.yaml, exact pins only in pubspec.lock. drift: ^2.31.0, never drift: 2.31.0. The lock pins; ranges only keep resolution solvable. Exact pins in a pubspec manufacture unsolvable conflicts on the next SDK bump and buy nothing the lock is not already delivering.
  2. Commit pubspec.lock. This is an application, not a package — the asymmetry that flips the rule. The default Dart .gitignore template lists pubspec.lock; delete that line. The committed lock is the only thing that makes a stranger's git clone resolve the exact versions that were tested on a real device.
  3. A pubspec.yaml diff without its pubspec.lock delta is incomplete. Run flutter pub get and stage the lock in the same commit — add, upgrade, and remove alike.
  4. Record the SDK version separately from the tool. Keep environment: sdk: a real range so pub can solve; pin the exact tested Flutter version in a committed record CI reads (see reference). The record is what a stranger and CI read; do not confuse a version-manager tool with the file.
  5. The lint include filename is coupled to the resolved SDK. A version-pinned include: that names a file absent from the resolved linter package emits include_file_not_found — fatal to default dart analyze/flutter analyze, so standard CI catches it. It only goes silent-and-green where warnings are made non-fatal, and even then it is your ruleset's added/promoted rules that stop applying, not the analyzer's built-ins. Verify it after any SDK bump (see reference).
  6. Audit the transitive tree before adding, not the pubspec. The second hop is exactly where a banned SDK arrives. Run scripts/audit-deps.sh (or dart pub deps --json + the audit script) before committing a new dependency.
  7. Refuse by policy anything that opens a network path, reports crashes/usage, drags in a telemetry core, or collects device identifiers for an unshipped feature — directly or transitively. Green CI is not evidence a native capability works. See the gate reference.
  8. Grep lib/ after removing a dependency. A package deleted from pubspec.yaml but still transitively resolvable keeps compiling today and breaks on the clone that matters. The analyzer catches an unresolved import, not one that resolves by accident.
  9. Wrap a bus-factor-1 native plugin behind an interface now, vendor later. The interface is cheap insurance; pre-emptive vendoring is a maintenance burden against a break that has not happened. See the vendoring reference.

Ranges in pubspec, pins in the lock

DoNever
Caret range: drift: ^2.31.0Exact pin: drift: 2.31.0
environment: sdk: ^3.6.0 — a real rangeAn exact sdk: version
Commit pubspec.lockGitignore pubspec.lock

The lock is authoritative. Ranges keep the solver able to move on an SDK upgrade; the lock keeps everyone on the versions that were actually verified.

The gate every new dependency must pass

Refuse a package if it — or anything in its transitive tree — does any of:

  • Opens a network path you did not choose (http, dio, sockets, gRPC) when the product promise is no-network or offline-first.
  • Reports crashes or usage (Crashlytics, Sentry, any analytics or attribution SDK) against a no-telemetry policy.
  • Drags in a telemetry/analytics core transitively — a crash SDK's core can worsen the store privacy label even when you never call it.
  • Collects device identifiers for a feature you do not ship.
  • Requires an --enable-experiment flag — an abandoned repo that needs an experiment to build stops building.

Also weigh, and usually refuse: any bus-factor-1 package not behind an interface; anything whose function is a few hundred lines of first-party Dart; anything that only exists to save typing. Record the licence — a permissive one (MIT, BSD, Apache-2.0) is a precondition for the vendoring escape hatch; a copyleft dependency removes that option. Full checklist in references/dependency-gate-and-audit.md.

Auditing the transitive tree

dart pub add --dry-run <package>     # see what would resolve, resolve nothing
dart pub deps --json > /tmp/deps.json
scripts/audit-deps.sh                 # walks the resolved set, flags banned patterns

The audit walks the full resolved graph, matches each name against a policy pattern list, marks every hit direct or TRANSITIVE, and separates APK-shipping deps from build/test-only ones (a banned package reachable only from dev_dependencies never reaches the binary and is not a shipping defect). Exit 1 means refuse the dependency or find one that does not pull those in. To learn who introduced a package: dart pub deps | grep -B4 <name>.

Edit the BANNED/ALLOW lists at the top of scripts/audit_deps.py to match your policy; every ALLOW entry needs a written justification beside it.

Upgrading

Run dart pub outdated on your own schedule; there is no bot. Dependabot has never supported pub; Renovate raises PRs a solo repo has no workflow for. Keep their one useful idea by hand: never blind-upgrade a native plugin your app's core capability depends on.

Green CI is not evidence of a native capability. No CI job on a headless emulator can prove the app plays audio, reads a sensor, or talks to a platform channel end-to-end — integration_test can assert a channel call was issued, nothing more. After any bump to a capability-critical native plugin or the SDK itself, run a manual pass on a real device. After any upgrade touching a codegen package (drift, freezed, json_serializable), regenerate and commit — stale generated code is caught by CI, not the analyzer (see run-codegen).

Vendoring a bus-factor-1 plugin

A single-maintainer native plugin whose failure means the app cannot build is the one dependency worth pre-planning for. The mitigation is a thin interface, authored on day one, so vendoring never touches lib/:

// A port the app depends on; the plugin lives behind it. The `Gateway` suffix
// marks a thin wrapper over a SPECIFIC plugin (here flutter_secure_storage) —
// `naming-conventions` owns that Service-vs-Gateway distinction.
abstract interface class SecureStorageGateway {
  Future<void> write(String key, String value);
  Future<String?> read(String key);
}

Only when the plugin actually breaks — stops building against a Flutter release, or ships a regression upstream will not fix — clone it at the last-good tag into third_party/<plugin>/, point the pubspec at path:, record the SHA and every changed line in a VENDORED.md, confirm the licence permits redistribution, and patch, don't refactor. Full procedure and triggers in references/vendoring-behind-an-interface.md. See examples/vendored_plugin_behind_interface.dart.

Removing a dependency

Delete the pubspec entry, run flutter pub get, commit the pubspec.lock delta in the same commit — then grep -r "package:<name>" lib/. A transitively-resolvable import compiles today and breaks the clone that matters.

When multi-package (workspace)

In a Dart pub workspace / monorepo, each package keeps its own pubspec.yaml with ranges, but the workspace resolves one shared lock at the root — audit and commit that root lock. Run the audit against the whole workspace's resolved tree, not one package's. A single package's pubspec cannot show what a sibling drags into the shared resolution. See codegen-and-toolchain for workspace toolchain mechanics.

Anti-patterns

  • Exact pin in pubspec.yaml. Turns the next SDK upgrade into an unsolvable-conflict debugging session; the lock already pins.
  • Gitignoring pubspec.lock in an app. A fresh clone resolves whatever pub feels like today, not what was tested.
  • Editing pubspec without committing the lock delta. The next machine resolves a different graph.
  • Reading pubspec.yaml to decide if a dep is safe. The banned SDK is on the second hop, invisible there.
  • Trusting green CI for a native capability. The emulator has no voice engine / real sensor; green proves the Dart compiled, not that the feature works.
  • Pre-emptively vendoring a healthy plugin. You take on the maintenance burden before any break; the interface is the cheap insurance, not the fork.
  • Vendoring by refactoring. Every line you touch in third_party/ is a line you own forever — patch the break and stop.
  • Adding a bus-factor-1 package with no interface. When it rots you have no seam to vendor behind.

Definition of done

  • New/changed deps use caret ranges; no exact pins in any pubspec.yaml.
  • pubspec.lock delta is staged in the same commit as the pubspec.yaml change.
  • scripts/audit-deps.sh exits 0 (or every hit has a justified ALLOW entry).
  • On an SDK bump: the version record is updated and the pinned lint include: file is confirmed present in the resolved package (references/sdk-pin-and-lint-include.md).
  • Licence recorded for any new dependency; permissive if it is a vendoring candidate.
  • Capability-critical native plugin sits behind an interface; a real-device pass was run after its bump.
  • After a removal: grep -r "package:<name>" lib/ is empty.

Related skills

  • codegen-and-toolchain — workspace linking, SDK pinning at the toolchain level, generated-code commit-vs-gitignore.
  • run-codegen — the deterministic build_runner pass to run after upgrading a codegen package.
  • lint-and-style-config — the strict analysis_options.yaml the version-pinned include: feeds.
  • service-boundary-and-native — the injectable-interface-per-side-effect pattern the vendoring seam relies on.
  • naming-conventions — owns the Service-vs-Gateway boundary-suffix rule; a plugin wrapper like SecureStorageGateway takes the Gateway suffix.
  • ci-pipeline-and-gates — wiring the audit and lock-freshness checks into CI.

References

What ships with it: 6 files

25.2 KB alongside SKILL.md, 2 of them executable

scripts/

Keep looking

Skills are one crate of 326,629. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.