Dependency management

Every dependency is a long-term relationship entered in one line of a requirements file: its bugs become your bugs, its license binds your distribution, its abandonment becomes your maintenance burden - and the same is true of everything IT depends on. The discipline has three phases: vet before adopting, stay current while depending, and exit deliberately. An agent adding a dependency without the intake check below is skipping due diligence on the user's behalf; run it by default and report the verdicts.

Intake: vet before adopting

Check every axis, not just the one that motivated the search (rseng-software-reuse finds candidates; this is the gate they pass through):

1. Suitability: does it actually solve the need, at the right size? A 50MB framework for one function fails this axis; prefer the smallest dependency that does the job - or the standard library, which needs no vetting.
2. License compliance: identify the license AND the licenses of its own dependency tree, and check compatibility with the project (rseng-license-compliance owns the analysis; deps.dev shows tree-wide licenses in one view). A perfect library under an incompatible license is not a candidate.
3. Trust and vulnerabilities - first AND second degree: check the package against vulnerability databases (osv.dev covers the major ecosystems), and check its TRANSITIVE tree - the dependency's dependencies (deps.dev renders the full graph with known advisories and OpenSSF Scorecard results). A clean package pulling a vulnerable or unmaintained transitive is an inherited problem; also weigh supply-chain signals: release provenance, maintainer count, typosquatting-adjacent names (rseng-security).
4. Documentation quality: real docs (install, API reference, examples that run), not just a README stub - undocumented dependencies transfer their documentation debt to your users and your future self (rseng-documentation's standards as the yardstick).
5. Maintenance and currency: recent releases, responsive issue tracker, more than one maintainer; check the LATEST version and adopt that, never an old version copied from a tutorial or an LLM's memory - adopting outdated versions imports already-fixed bugs and already-patched vulnerabilities. Verify the current version against the registry, not recall - and when the technology or domain is unfamiliar, research the ecosystem's CURRENT tooling online (official docs, the community's guide) before choosing; the state of the art moves faster than training data, and requirements.txt-era habits are the canonical example of a stale default.
6. Community and continuity: bus factor, governance, whether the project has a succession story (rseng-maintenance-sustainability reads these signals) - for load-bearing dependencies, an abandoned upstream is a foreseeable crisis.

Record the verdict: a one-line entry per adopted dependency in the decision log (rseng-project-tracking) - what it is for, the axes checked, any accepted risks. Declined candidates with reasons are worth a line too (rseng-lessons-learned).

Staying current: the anti-rot regime

• Pin everything with lockfiles (rseng-reproducible-environments): reproducibility and auditability come from the same mechanism - you cannot vet what you cannot enumerate.
• Update on a cadence, not on crisis: automated update PRs (Dependabot/Renovate class) sized to the project's tier - security updates triaged promptly always; routine updates batched on a schedule the tests can absorb (rseng-ci-cd runs the suite against every bump; that is what makes updating cheap).
• Avoid outdatedness deliberately: track how far behind latest each dependency sits and treat growing lag as debt - the further behind, the more painful the eventual jump and the longer the exposure window to fixed bugs. End-of-life versions of runtimes and libraries (endoflife.date tracks them) are hard deadlines, not suggestions.
• Read the release notes on major bumps: breaking changes, deprecations and behavior changes belong in the update PR's description, with the project's affected call sites named (rseng-pair-programming's pre-review makes update PRs cheap to merge).
• Re-audit periodically: the milestone project review (rseng-code-review) re-runs the intake axes over the EXISTING tree - licenses change, maintainers leave, advisories land after adoption.

The transitive tree is your tree

• Know what you actually ship: the lockfile enumerates it; an SBOM makes it queryable (rseng-security). "We do not use X" is only true if nothing in the tree does.
• Minimize depth where choices exist: prefer dependencies with small, well-maintained trees; every transitive node is attack surface, license surface and rot surface.
• When a transitive is the problem (vulnerable, unmaintained, license-incompatible): options in order - update the direct dependency that pulls it, override/pin the transitive where the ecosystem allows, replace the direct dependency, or vendor-and-patch as the documented last resort.

Exit deliberately

Removal is part of management: when a dependency rots, duplicates another, or a few stdlib lines would do, remove it - each removal deletes risk. Migrations off a dependency get the parity-test treatment (rseng-legacy-code); keeping a known-bad dependency gets a dated, reasoned entry in the decision log with a revisit date, never silence.

Working with this skill

This skill is source-independent: its authority is the ecosystem services and tool documentation linked below. It gates what rseng-software-reuse discovers, applies rseng-license-compliance and rseng-security per axis, and hands the pinning to rseng-reproducible-environments.

Learn more (verified):

• https://deps.dev - Open Source Insights: transitive graphs, licenses, advisories, Scorecard
• https://osv.dev - OSV vulnerability database
• https://endoflife.date - end-of-life dates for runtimes and libraries
• https://docs.github.com/en/code-security/how-tos/secure-your-supply-chain - Dependabot automated updates
• https://docs.renovatebot.com - Renovate automated updates

Related skills

Check whether any of these applies before moving on:

• rseng-ci-cd - tests absorb automated update PRs
• rseng-license-compliance - license axis analysis
• rseng-maintenance-sustainability - upstream bus-factor and succession signals
• rseng-reproducible-environments - pinning and lockfiles
• rseng-security - vulnerabilities, SBOM, supply chain signals
• rseng-software-reuse - candidates entering the intake gate