node-deps-bumper — Claude Code skill: what it does, how to install
node-deps-bumper — a skill for Claude Code and Codex that structures package upgrades via Taze and guards against regressions. Runs whenever npm, pnpm, yarn, or bun dependencies need updates or dry-run checks.
MIT 106 installs ★ 107 updated 2026-10-24
npx skills add PaulRBerg/agent-skills --skill node-deps-bumper -a claude-codeWhat node-deps-bumper does
The skill runs Taze via a helper script run-taze.sh to classify candidate updates into categorized buckets: automatic ranged updates, major bumps requiring confirmation, manual reviews, and skipped pins. Passing --dry-run prints the upgrade counts and candidate table without modifying package manifests or lockfiles.
Before writing changes, the workflow captures a baseline run covering builds, lints, types, and test suites. Pre-existing unrelated test failures do not stop the run as long as dependency resolution remains stable. For projects utilizing Bun catalogs, the tool invokes update-bun-catalogs.py with uv to preview and apply transitions atomically.
Writing updates alongside Bun catalogs requires Taze >=21.3.0 <22 with the --direct-only flag to safeguard peer declarations. Manifest and lockfile modifications are synchronized using ni, followed by rerunning the recorded test suite to isolate and fix regressions introduced by new packages.
What it helps with
- Generate a structured dependency update plan across monorepos and packages
- Handle major version bumps in deliberate batches with release verification
- Safely modify Bun catalog entries via update-bun-catalogs.py
- Verify baseline builds and test suites before and after altering lockfiles
Example prompts for the agent
- Check outdated packages and build an update plan with --dry-run
- Bump minor and patch dependencies using node-deps-bumper
- Review and apply major version upgrades across the repository
How to install node-deps-bumper
1Skills CLI (Vercel)
npx skills add PaulRBerg/agent-skills --skill node-deps-bumper -a claude-code2Manually: put the files in place
~/.claude/skills/node-deps-bumper/SKILL.mdCopy the whole skill folder; for one project use .claude/skills/node-deps-bumper/
1Skills CLI (Vercel)
npx skills add PaulRBerg/agent-skills --skill node-deps-bumper -a codex2Manually: put the files in place
.agents/skills/node-deps-bumper/SKILL.mdFor all projects use ~/.agents/skills/
1Skills CLI (Vercel)
npx skills add PaulRBerg/agent-skills --skill node-deps-bumper -a cursor2Manually: put the files in place
.cursor/skills/node-deps-bumper/SKILL.mdCursor also reads .agents/skills/ and .claude/skills/
1GitHub CLI
gh skill install PaulRBerg/agent-skills skills/node-deps-bumper2Manually: put the files in place
.github/skills/node-deps-bumper/SKILL.md1Skills CLI (Vercel)
npx skills add PaulRBerg/agent-skills --skill node-deps-bumper -a gemini-cli2Platform CLI
gemini skills install https://github.com/PaulRBerg/agent-skills --path skills/node-deps-bumperManually: put the files in place
.devin/skills/node-deps-bumper/SKILL.mdThe legacy .windsurf/skills/ path still works
Skills CLI (Vercel)
npx skills add PaulRBerg/agent-skills --skill node-deps-bumper -a opencodeClaude.ai
Zip the folder and upload it in Settings → Capabilities → Skills
Step by step, with commands for each platform: How to install skills in Claude Code: five ways
Questions about node-deps-bumper
How to install node-deps-bumper in Claude Code?
Run npx skills add PaulRBerg/agent-skills --skill node-deps-bumper -a claude-code in your terminal.
What does node-deps-bumper do?
It scans projects using Taze, groups patch, minor, and major package bumps, checks the repository baseline, and updates lockfiles using ni.
What is required to run node-deps-bumper?
You need Bash to run the helper scripts, the ni package runner, uv for Bun catalog operations, and Taze >=21.3.0 <22 when using --direct-only.
Source text
SKILL.md · MIT
---
argument-hint: "[--dry-run] [package ...] [repo-path ...]"
effort: medium
model: sonnet
name: node-deps-bumper
description: "Use for dependency updates: bump npm/pnpm/yarn/bun packages, check outdated, or run taze."
---
# Node Dependency Bumper
Use Taze to build one structured update plan, apply compatible ranged updates, and make major-version decisions as a
batch.
## Adding Dependencies
For a new package or CLI, use the repository's package manager and existing range convention instead of Taze. Installing
`package@latest` and retaining the package manager's resulting `^x.y.z` range is allowed. Do not replace a caret range
with an exact version merely because it can admit future releases.
When an effective package-manager minimum-release-age policy exists, it provides the freshness boundary for candidate
versions. In that case, read [references/conditional-workflows.md](references/conditional-workflows.md). The committed
lockfile and frozen deployment install provide reproducibility. Exact-pin only when the user or repository requires it,
a known compatibility constraint justifies it, or the package will run without a committed lockfile and frozen install.
## Multiple Repositories
When the user names two or more repositories or asks to sync or align dependencies across repositories, read
[references/multi-repo-alignment.md](references/multi-repo-alignment.md). It runs the Workflow below per repository and
adds shared-target selection, one cross-repository major batch, and per-repository commits.
## Workflow
1. Resolve the skill directory and save the helper plan from the target repository:
```sh
bash <skill-dir>/scripts/run-taze.sh --plan [--include package-a,package-b] > <taze-plan.json>
```
The JSON plan classifies every discovered update as `apply`, `review-major`, `review`, or `skip-fixed`. The helper
detects monorepos, includes locked versions during scans, and mirrors Bun minimum-release-age settings. If the
repository uses package-manager age gates or Bun catalogs, read
[references/conditional-workflows.md](references/conditional-workflows.md) for that active branch only.
2. If `--dry-run` was requested, present the plan and counts, then stop without changing manifests or lockfiles.
3. Select every ranged minor/patch update marked `apply`. Reuse explicit approval of package/version transitions. Never
auto-approve a major merely from its package name. Present unresolved `review-major` and unknown updates in one
decision batch with current version, target version, package role, and relevant migration/release notes. Apply only
authorized majors.
4. If nothing is selected, report the no-op and stop. If the root manifest uses Bun catalogs, preview the exact selected
catalog transitions from the accepted plan. Select catalogs used for installed dependencies with repeatable
`--catalog` arguments. Use `default` for the unnamed catalog. Preserve catalogs used for public peer ranges. Pass
only selected packages present in those catalogs. When that subset is empty, skip the helper:
```sh
uv run <skill-dir>/scripts/update-bun-catalogs.py \
--root <repo> --plan <taze-plan.json> --include package-a,package-b \
--catalog default --catalog dev
```
The preview is read-only. Unknown catalogs, missing entries, ambiguous targets, or stale current values fail before
writes. The helper matches each selected occurrence against the plan and preserves fixed holds. It does not select
upgrades.
5. Before the first manifest or lockfile write, select and run a baseline that can detect regressions from the chosen
updates. Honor repository-required checks. Otherwise, select dependency-resolution, build, test, typecheck, lint,
formatting, codegen, or invariant checks from the updated packages' actual consumers and risk. Use an aggregate suite
for shared toolchain or runtime changes, not merely because it exists. Use frozen or non-writing modes where
available, and record the exact commands and coverage for the post-bump rerun.
Attribute every failure before deciding whether it blocks. Proceed when an unrelated pre-existing failure is
reproducible, can be compared after the bump, and does not prevent dependency resolution or the checks needed to
detect regressions. Do not fix that failure as part of the bump. Dependency or peer-resolution conflicts, actionable
unsafe behavior, or a baseline that cannot provide trustworthy before/after signal block the bump. In that case, stop
with `### ⛔ Dependency bump blocked — baseline unusable` and report the exact prerequisite and diagnostics without
asking for redundant authorization. Informational notices such as unavoidable deprecations do not block.
6. For Bun catalogs, apply the previewed transitions first: rerun `update-bun-catalogs.py` with the same plan and
catalog and package selections plus `--write`. It atomically updates only selected catalog occurrences and preserves
their prefixes. Then use `--direct-only` for the Taze write to protect all Bun catalogs and ordinary peer
declarations. This mode requires Taze `>=21.3.0 <22` and rejects unsupported project or ancestor configuration before
writes.
For repositories without Bun catalogs, omit `--direct-only`:
```sh
bash <skill-dir>/scripts/run-taze.sh --write --direct-only --include package-a,package-b
```
7. Run `ni` so the repository's package manager updates its lockfile.
8. Inspect the manifest and lockfile diff. Rerun the exact baseline commands, plus the narrowest checks that exercise
the updated dependencies and any required migrations. Treat every newly introduced error, type issue, check failure
(including tests, builds, lint, formatting, codegen, and repository invariants), dependency or peer-resolution
conflict, and actionable compatibility or safety warning as caused by the bump unless evidence shows otherwise.
9. Fix every issue caused by the bump, including required source or configuration migrations, while preserving intended
behavior. Do not suppress diagnostics, weaken validation, or change expected behavior merely to make checks pass.
After each fix, rerun checks whose inputs or behavior changed. At completion, require valid results for the entire
recorded suite against the final state, with no new failures against the accepted baseline. Reuse passing results
whose inputs remain unchanged. Only unrelated pre-existing failures meeting step 5's comparison requirements may
remain.
If no clear safe fix exists within the task's authority, stop with `### ⚠️ Dependency regression decision required`.
Present all such issues in one table with the evidence, affected locations, fix and revert options, and likely
effects. Do not report completion until the user chooses, the fix is applied or the offending update is reverted, the
lockfile is regenerated, and the complete suite meets that same baseline-comparison requirement.
## User-Facing Output
Present plans as `### 📦 Dependency plan` with counts and a compact table:
| ID | Plan value | Decision | Package | Current → target | Type | Notes |
| --- | ---------- | -------- | ------- | ---------------- | ---- | ----- |
Use the plan's exact `apply`, `review-major`, `review`, and `skip-fixed` values alongside plain-language decisions.
Assign stable IDs to rows needing a choice so the user can answer once. Use `### 🔎 Dry run — no files written` for a
preview and `### ✅ No selected updates` for a no-op.
Finish applied work with `### 🏁 Dependencies updated`, a tree of changed manifests/lockfiles, and
`### 🧪 Verification`. Use `### ⚠️ Remaining review` only for non-blocking informational matters, never for an
unresolved issue caused by the bump. Keep helper JSON, package/version strings, commands, and diagnostics exact and
undecorated.
## Invariants
- Fixed versions and non-semver protocols remain unchanged unless the user explicitly asks otherwise.
- Caret ranges are valid for new dependencies and CLIs. Do not describe them as unreproducible when a committed lockfile
and frozen install control resolution.
- Package arguments constrain both scan and write phases.
- The same maturity-period policy applies to scan and write.
- Bun catalog preview and write use the same accepted plan, catalog selection, and package set. Stale plans never write.
- Peer declarations describe supported consumer versions. They are not installed versions. Preserve public peer ranges
unless the user authorizes a contract change.
- Do not infer compatibility from SemVer alone when repository evidence, peer ranges, or release notes indicate
otherwise.
Completion requires an attributed, comparison-safe pre-write baseline, a reviewed plan, the retained selected updates, a
regenerated lockfile, no new failure in the recorded suite or dependency-specific checks, and no unresolved issue caused
by the bump. Dry-run completion requires the structured plan and zero writes.
© PaulRBerg · MIT · PaulRBerg/agent-skills