An Agent Skill for the safe, gated review and update of dependencies in an Android (Kotlin/Gradle) project that uses a Gradle version catalog. One
SKILL.mdfolder that works with Claude Code, OpenAI Codex, Gemini CLI, and any Agent-Skills-compatible tool.
Upgrading dependencies is easy to get wrong: automated bumpers jump every library to its latest version with zero analysis, silently breaking your build or your app. This skill does the opposite — it turns "update my dependencies" into a controlled, reviewable procedure that detects what's outdated, reasons about the risk, and never changes anything without your explicit confirmation.
It's designed for real Android projects: version catalogs (gradle/libs.versions.toml), the Android
Gradle Plugin, Compose/Firebase BOMs, coupled version blocks (Kotlin ↔ KSP ↔ Compose Compiler),
JitPack libraries, and convention plugins (build-logic/).
- Gated by design. It detects, groups, classifies by risk, and stops at a proposal for you to approve — all / safe-only / a specific subset. Nothing is edited or committed until you say go.
- Sees the whole picture, not just version numbers. It groups coupled blocks and BOM-governed
artifacts into single items, ignores transitive noise, and flags hidden requirements (a
"safe-looking" minor that actually needs a higher
compileSdk/AGP/Kotlin). - Covers the tool's blind spot. The ben-manes
gradle-versions-plugincan't see JitPack (com.github.*) libraries — the skill checks those by hand. - Verifies before it trusts. Applied bumps are built with
:app:assembleDebug; if something breaks, it isolates the culprit, reverts it, and returns to the proposal — never leaving your tree broken. - Clean git hygiene. Commits locally on a feature branch (never the default branch, no push), one thematic commit per theme, with a separate commit for any code adaptation to new APIs.
- Speaks your language. User-facing prose mirrors the language you write in (English or Spanish).
| Concern | Behavior |
|---|---|
| Version catalog | Edits gradle/libs.versions.toml version.refs — the single source of truth |
| Detection | ./gradlew dependencyUpdates --no-parallel (ben-manes), aggregated across all modules |
JitPack (com.github.*) |
Checked manually via JitPack metadata / GitHub releases (plugin blind spot) |
| BOMs (Compose, Firebase, …) | Bumps only the BOM; ignores the governed child artifacts in the report |
| Coupled blocks | Kotlin ↔ KSP ↔ Compose Compiler, AGP ↔ Gradle wrapper, Retrofit, OkHttp, Room, Hilt… treated as single items |
| Risk | Semver magnitude + hidden compileSdk/AGP/Kotlin requirements + known vulnerabilities + license changes |
| Verification | :app:assembleDebug (plus tests/linters when relevant); isolates any culprit bump |
| Code adaptation | Post-bump deprecation/migration pass, in a separate refactor(deps) commit |
A fixed, repeatable procedure (the skill stops at step 5 for your approval):
- Discover the project shape — catalog, detection tool, JitPack libs, coupled blocks, SDK/JDK/wrapper.
- Detect updates with the ben-manes plugin (+ a manual JitPack pass).
- Aggregate & dedupe the per-module reports (helper script included).
- Filter noise — drop BOM-governed children and transitive satellites; group coupled blocks.
- Classify by risk and propose a table (safe vs. handle-with-care) — ⛔ GATE: waits for you.
- Apply only what you confirmed.
- Verify with a build; isolate and revert any culprit.
- Commit locally on a branch, one thematic
chore(deps)commit, no push. - Adapt the code to new APIs if needed — separate
refactor(deps)commit.
Agent Skills are an open format — the same SKILL.md folder
works across compatible agents. Install it wherever your agent looks for skills.
Codex, Gemini CLI, VS Code and others scan a shared skills directory. Clone the repo there:
# user-level (all your projects)
git clone https://github.com/alvarose/android-update-deps.git ~/.agents/skills/android-update-deps
# or project-scoped
git clone https://github.com/alvarose/android-update-deps.git .agents/skills/android-update-depsgemini skills install https://github.com/alvarose/android-update-deps --consentPlace the folder in ~/.agents/skills/android-update-deps/ (or a repo .agents/skills/), then
mention it in chat with $skills.
Install as a plugin (one command), then update with /plugin update android-update-deps@alvarose:
/plugin marketplace add alvarose/android-update-deps
/plugin install android-update-deps@alvarose
…or clone into a directory Claude scans for skills:
git clone https://github.com/alvarose/android-update-deps.git ~/.claude/skills/android-update-depsOr download the packaged android-update-deps.skill from the
Releases page.
Once installed, just ask your agent in an Android project — the skill triggers on phrases like:
- "review / update / bump the dependencies of my Android app"
- "what's outdated in my
libs.versions.toml?" - "actualiza las dependencias, solo las seguras"
- or invoke it by name (e.g.
/android-update-depsin Claude Code,$skillsin Codex,/skillsin Gemini CLI)
The agent will detect, analyze, and show you a risk-grouped proposal. You pick what to apply; it verifies with a build and commits on a branch. It won't touch anything without your OK.
- An Android (Kotlin/Gradle) project, ideally with a version catalog (
gradle/libs.versions.toml). - The ben-manes
gradle-versions-pluginapplied (the skill offers to add it temporarily if it's missing). On Gradle 9+ use0.52.0+. - A working Android SDK (
sdk.dirin the repo'slocal.properties).
android-update-deps/
├── SKILL.md # the skill: discovery, the 9-step gated procedure
├── references/
│ └── reference.md # coupled-versions table, JitPack procedure, report.json format
├── scripts/
│ └── aggregate-updates.py # dedupe report.json across modules; list JitPack libs
├── .claude-plugin/ # Claude-specific plugin + marketplace manifests (for /plugin install)
├── evals/ # evaluation prompts + assertions used to test the skill
├── CHANGELOG.md
├── LICENSE
└── README.md
Issues and PRs welcome. The skill is intentionally concise and gated — proposals that add
scope should preserve those principles (see the "handle with care" philosophy in SKILL.md).
MIT © Alvaro Serrano