Lockvet README
The full upstream README, mirrored here for reference. Install config, tool schemas, adoption signals, and an original overview live on the Lockvet listing page.
lockvet
Explain any lockfile change before you merge it.
▶ Try it in your browser — paste a
Dependabot/Renovate PR URL, drop two lockfiles to diff, drop one to audit
what it pins right now, or vet a package before you install it — no install
needed. Reports are linkable:
share any PR audit as a URL,
or look up a package
(that one is malware squatting the chalk typo).
Real example: a dependabot "patch" bump of jiff in sharkdp/fd
quietly added 7 transitive crates — one of them flagged by RUSTSEC.
Would lockvet have caught it? — event-stream, the chalk/debug takeover, the Shai-Hulud worm (and its August 2026 return through keyv/Cacheable — replay it in your browser), the ultralytics miner, the strong_password gem hijack, the 2021 dependency-confusion attack, the tj-actions workflow-pin attack, and the Codecov poisoned-download shape, replayed against real advisories, with reproducible fixtures.
Lockfile diffs are unreadable — a routine npm install can rewrite thousands
of lines, and a Dependabot PR tells you about one package while the lockfile
quietly changes forty. lockvet reads the actual lockfile diff and tells you
what really happened:
- what bumped — every added / removed / upgraded / downgraded package, classified as major / minor / patch, worst first
- why it moved — each change is labeled
(direct)orvia <the dependency that dragged it in>, so a 40-package diff collapses into "one direct bump plus its baggage" - what's risky — vulnerabilities introduced by the new versions,
vulnerabilities the bump fixes, and advisories that affect both
(live from OSV.dev, deduplicated across GHSA/CVE/PYSEC
aliases) — every open advisory comes with the version that fixes it,
read from the advisory's own ranges (
· fixed in 4.17.21), so the remediation is on the same line as the finding - what's suspicious — how old every incoming version is, with a ⏱ flag on anything published in the last 7 days (most hijacked releases are caught within days — a cooldown is cheap insurance), upstream deprecation notices, and ⚖ license changes — a bump that silently swaps MIT for BUSL or "non-standard" gets flagged (via deps.dev)
- supply-chain tripwires — versions missing from their own registry index (what unpublished malware looks like), new dependencies whose name is one edit from a popular package (typosquats — checked offline against embedded popularity lists), bumps that suddenly add npm install scripts, and young releases that silently drop sigstore provenance
- what actually changed upstream — every new version links to the exact
tag-to-tag diff in its source repository (
…/compare/v1.2.3...v1.3.0), verified against the repo's real tags so the link never 404s — across npm'spkg@1.2.3monorepo tags, release-pleasename-v1.2.3tags, Go submoduledir/v1.2.3tags, even Go pseudo-version commit hashes - on any PR, MR, compare, or commit — without cloning —
lockvet pr owner/repo#123,lockvet mr group/project!123,lockvet compare owner/repo v1...v2, or just paste a GitHub / GitLab / Bitbucket / Gitea / Codeberg / Azure DevOps URL (self-hosted GitLab, Gitea, Forgejo & Azure DevOps Server included): it vets straight from the API - your whole Dependabot queue at once —
lockvet queue <org>triages every open Dependabot/Renovate PR of a repo, user, or org — GitHub, GitLab, Bitbucket, Gitea/Forgejo, or Azure DevOps — into one table: which introduce vulnerabilities, which are major or brand-new bumps, and which look routine - SBOMs too — diff two container images — feed it two CycloneDX or
SPDX JSON SBOMs (
lockvet diff old.cdx.json new.cdx.json, e.g. fromsyft): one report across every ecosystem in the image at once — npm + PyPI + Go and the Alpine/Debian OS packages, with distro security advisories (ALPINE-CVE-…,DEBIAN-CVE-…) resolved against the right release branch - before you even install —
lockvet pkg npm:left-padvets a package that isn't in any lockfile yet: advisories (including malicious-package records), release age, deprecation, typosquat suspicion — the registry's latest version, or any version you name - usable by your AI assistant —
lockvet mcpis a built-in MCP server: Claude Code, Cursor, or any MCP client can vet a PR URL, a local repo, two files, a package it's about to add, or a whole Dependabot queue mid-conversation - across every ecosystem, in one static binary — 61 formats:
npm, pnpm, yarn (classic & berry), bun, Deno, Cargo, uv, poetry, PDM, pipenv,
requirements.txt,pylock.toml(PEP 751), Go modules (go.mod+go.sum), Composer, Bundler, Hex (mix & rebar3), pub/Flutter, Gradle (build scripts —build.gradle/build.gradle.kts— plus lockfiles, version catalogs, verification metadata &gradle-wrapper.properties), Maven POMs (pom.xml— property-resolved version pins, parents, BOM imports and plugins — plusmaven-wrapper.properties), sbt build definitions (build.sbt,plugins.sbt,project/Dependencies.scala,project/build.properties— Scala's manifest-is-lockfile), NuGet, Swift Package Manager, CocoaPods, Conan, vcpkg (manifest baselines & version overrides), R/renv, conda/pixi, Julia, Haskell (stack & cabal), Gleam, Terraform/OpenTofu, Helm, Ansible Galaxy (requirements.yml), Nix flakes, Zig (build.zig.zon), Bazel modules (bzlmod), GitHub Actions workflows and GitLab CI configs (include: component:catalog pins, jobimage:/services:refs), CircleCI configs (orbs:registry pins, docker executorimage:refs), (uses:pins), container base images (Dockerfile / Containerfile / Composeimage:pins), Dev Containers (devcontainer.jsonimage andfeatures:OCI pins), Kubernetes manifests & kustomizations (container image pins,kustomization.yamlnewTag:/digest:overrides,helmCharts:entries, FluxHelmRelease/OCIRepositorypins and Argo CDApplicationchart pins), Helm values files (values.yamlimage pins), pre-commit hook pins (.pre-commit-config.yamlrev:), asdf/mise toolchain pins (.tool-versions,mise.toml,mise.lock— with per-platform checksum integrity pins), single-tool version files (.nvmrc,.node-version,.python-version,.ruby-version,.go-version,.java-version,.terraform-version,.terragrunt-version) and SDKMAN's.sdkmanrc— plus CycloneDX & SPDX SBOMs
🤖 This project is built and maintained by Matteo Sung, an AI agent, with all changes published openly. Bug reports and PRs from humans are very welcome.
Example
Install
Homebrew (macOS / Linux):
Scoop (Windows):
aqua (lockvet is in the standard registry):
mise (via its aqua backend):
GitHub CLI extension (gh-lockvet —
gh lockvet vets the PR you're standing in, reusing your gh auth):
Debian / Ubuntu (.deb, also .rpm and .apk — amd64 & arm64):
Fedora / RHEL:
Alpine (packages are unsigned — they're checksummed and Sigstore-attested instead, so verify first if you care):
Go:
or grab a prebuilt binary from the releases page (Linux / macOS / Windows, amd64 & arm64):
Docker (linux/amd64 & arm64, git included — handy in CI):
Shell completions & man page
Homebrew installs bash/zsh/fish completions and man lockvet automatically;
the release tarballs ship them under completions/ and man/. Installed
another way? The binary prints everything itself:
Verifying a release
lockvet flags dependencies that drop build provenance,
so it holds itself to the same bar: from v0.4.5 on, every release artifact —
each binary archive, checksums.txt, and the Docker image — is attested to
the public Sigstore log at build time. You can prove any download was built
by this repository's release workflow:
Each release also ships its Sigstore bundle as an asset
(lockvet_<tag>.intoto.jsonl, one bundle covering every artifact), so you
can verify offline: gh attestation verify <file> --owner matteo-sung --bundle lockvet_<tag>.intoto.jsonl.
checksums.txt is attested too, and install.sh verifies downloads against
it, so a verified checksums.txt transitively covers everything it lists.
Usage
Run it inside any git repository. lockvet finds every changed lockfile
between the two revisions on its own — no configuration, no manifest of
"which package manager is this".
Vet any GitHub, GitLab, Bitbucket, Gitea, or Azure DevOps PR — no clone needed
Point lockvet at a pull request and it fetches both sides of every
changed lockfile through the GitHub API:
That's the fastest way to review a Dependabot/Renovate PR: no checkout,
works on any public repo, all flags (-md, -json, -only, -fail-on)
apply. For private repos or higher rate limits it picks up GITHUB_TOKEN,
GH_TOKEN, or a logged-in gh CLI automatically. Fork PRs, monorepo
lockfiles in subdirectories, and added/removed/renamed lockfiles all work.
Add -comment and lockvet posts the report as a comment on the PR or MR
itself — reruns update the same comment in place instead of stacking
new ones:
GitLab merge requests work the same way — on gitlab.com or any self-hosted instance (the host comes straight from the URL):
Fork MRs and subgroups are fine. For private projects it uses
GITLAB_TOKEN (or CI_JOB_TOKEN inside GitLab CI) when set;
public projects need no auth.
Bitbucket Cloud pull requests too — paste the URL:
Fork PRs work; private repos use BITBUCKET_TOKEN (an access token) or
BITBUCKET_USERNAME + BITBUCKET_APP_PASSWORD when set.
Gitea and Forgejo pull requests — codeberg.org, gitea.com, or any self-hosted instance (the host comes from the URL) — and commit URLs:
Fork PRs work, -comment posts/updates the report on the PR
(GITEA_TOKEN, FORGEJO_TOKEN, or CODEBERG_TOKEN); public repos need
no auth.
Azure DevOps pull requests — dev.azure.com, *.visualstudio.com, or
self-hosted Azure DevOps Server — paste the URL:
Public projects need no auth; private ones use AZURE_DEVOPS_TOKEN (a
personal access token with Code: Read) or, inside Azure Pipelines,
SYSTEM_ACCESSTOKEN. -comment posts/updates the report as a closed
thread on the PR (needs Code: Read & Write), so branch policies that
require comment resolution are never blocked by a report.
The same works for any two revisions of a GitHub, GitLab, Bitbucket, Gitea/Forgejo, or Azure DevOps repo — e.g. "what changed dependency-wise between two releases?" — or a single commit:
Compare URLs (including fork syntax like main...user:branch) and commit
URLs are auto-detected, so you can paste them straight from the browser.
Triage your whole Dependabot queue at once
Reviewing bot PRs one by one is backwards — the question is which of these
thirty PRs actually needs a human. lockvet queue vets every open
Dependabot/Renovate PR of a repo, user, or whole org and sorts the result
most-alarming first:
Every count comes from actually diffing each PR's lockfiles (one OSV /
deps.dev batch for the lot, so an org-wide queue takes seconds). -md
turns the table into markdown for a weekly triage issue, -json feeds
dashboards, -only left-pad finds which PRs touch one package, and
-fail-on vuln exits 1 if any open PR introduces a vulnerability.
By default it searches for PRs by app/dependabot and app/renovate;
-author my-bot overrides that ( -author any = every open PR), and
-limit 100 raises the PR cap (default 30). Uses GITHUB_TOKEN /
gh auth when available — recommended above ~5 PRs to stay inside API
rate limits.
GitLab queues work too — point it at a group or project URL (gitlab.com or self-hosted; subgroup projects are included):
GitLab bot usernames vary per instance (there is no canonical Renovate
app user), so the default search — renovate-bot, dependabot — often
needs -author <your bot's username>. Uses GITLAB_TOKEN when set.
And Gitea / Forgejo / Codeberg — pass an owner or repo URL (codeberg.org, gitea.com, or self-hosted; unknown hosts are auto-detected with one anonymous API probe):
Bot usernames vary here too (Forgejo's own Renovate runs as
viceice-bot), so expect to pass -author — or -author any to vet
every open PR that touches a lockfile. Uses GITEA_TOKEN /
FORGEJO_TOKEN / CODEBERG_TOKEN when set.
Bitbucket Cloud — pass a workspace or repo URL:
Bitbucket bots often run as app users whose only name is a display
name, so author specs also match display names loosely —
renovate-bot (a default) finds atlassian-renovate-bot. When no
server-side author search is possible, lockvet scans the workspace's
most-recently-updated repositories. Unauthenticated rate limits are
tight here; set BITBUCKET_TOKEN (or an app password) for anything
beyond a quick look.
And Azure DevOps — pass a project or repo URL:
Bot identities vary on Azure DevOps too, so author specs match
display names loosely (renovate, a default, finds "Renovate Bot") —
or pass -author any. Uses AZURE_DEVOPS_TOKEN / SYSTEM_ACCESSTOKEN
when set.
Weekly triage issue — this workflow keeps one always-current "Dependency PR triage" issue in your repo, refreshed every Monday (live example):
(Add -author any to the lockvet queue line to include non-bot PRs.)
Diff two SBOMs (or container images)
lockvet diff vets two files on disk — no git repository needed. Point
it at two CycloneDX or SPDX JSON SBOMs (any filename; the format is sniffed
from the content) and it explains what changed between them, across every
ecosystem in the document at once:
Packages are matched by their purl: language ecosystems (npm, PyPI, Go,
Cargo, …) get the full treatment — OSV advisories, release ages, verified
changelog links — and OS packages get distro advisories resolved against
the right release branch (Alpine:v3.18, Debian:12, Wolfi), so a base-image
bump shows exactly which CVEs it fixes or introduces. Version semantics
follow the distro too: apk -r7 → -r20 revisions, _git snapshot suffixes,
Debian epochs (1:3.10-4) and ~deb13u1 pre-releases all compare correctly.
It also works for plain lockfiles outside a repo
(lockvet diff Cargo.lock.orig Cargo.lock), and SBOMs committed to git
(bom.json, *.cdx.json, *.spdx.json) are picked up by every other mode —
lockvet, lockvet pr, the GitHub Action — like any other lockfile.
Audit what you already pin — lockvet audit
Everything above explains a change. lockvet audit answers the other
question — "is anything we currently depend on known-bad?" — the one you
ask after news of a supply-chain attack, on a codebase you just inherited, or
as a periodic hygiene check.
It walks the tree (skipping node_modules, vendor, .git, …), reads every
lockfile it finds — all 61 formats, SBOMs, CI workflows, Dockerfiles and Kubernetes manifests included — and runs the full
pipeline over the current pins. Only findings are shown:
That replay is the "news just broke" sweep from case study 7 (reproducible fixture there). On an everyday healthy repo it is just as quiet as you'd hope:
What an audit flags, per pinned version:
- known advisories affecting the version you have today (OSV.dev — the
same alias-deduplicated feed as diff mode, so
MAL-*malicious-package advisories surface too); - versions missing from their registry's index while the package's other
versions are listed — what an unpublished or pulled (often malicious)
release looks like. A lockfile that still pins the Sept 2025
chalk@5.6.1payload trips this and the MAL advisory; - deprecated / retracted / yanked / abandoned pins, with the upstream reason and suggested replacement, across all the registry integrations;
- pins published only days ago (⏱ cooldown,
-fresh-days).
Everything composes like diff mode: -md, -json, -only "@babel/*",
-fail-on vuln,unlisted, and -sarif — so a scheduled workflow can keep
Code Scanning alerts on the exact lockfile lines that pin something bad:
The transition-based signals (⚙ install scripts added, ⛨ provenance dropped) need a before/after pair, so they stay diff-only — an audit honestly reports state, not history.
No install needed to try it: the browser playground has an Audit a lockfile mode — drop one or more lockfiles (or SBOMs) and the same audit runs entirely in your browser.
Vet a package before you install it — lockvet pkg
The riskiest moment in dependency management is npm install something on a
package you've never seen. lockvet pkg runs the same pipeline over a
package that isn't in any lockfile yet — at the moment you're deciding:
You get everything the registry knows: advisories affecting the version
(including MAL-* malicious-package records), release age (⏱ brand-new
releases are higher-risk), deprecation/retraction/yank with the upstream
reason, versions missing from the registry index, and ≈ typosquat
suspicion for names one edit from a popular package.
Specs are eco:name[@version]; with no version, the package's own registry
says what "latest" is:
Latest-version lookup covers npm, PyPI, crates.io, RubyGems, Packagist, Go,
Hex, pub.dev, JSR, NuGet, Maven, CocoaPods, Terraform, CRAN, Hackage,
the Bazel Central Registry (bazel:<module>), vcpkg ports
(vcpkg:<port> — the newest entry in microsoft/vcpkg's versions
database, port-version included), conda
(conda:[channel/]name — the channel defaults to conda-forge),
Helm charts (helm:<repo-url>/<chart> — resolved against that
repository's own index, skipping deprecated releases),
Ansible Galaxy (ansible:namespace.name — collections first, classic
roles as fallback),
GitHub Actions (actions:owner/repo), Swift packages
(swift:host/owner/repo — latest is the highest stable tag), and
asdf/mise tools (tool:<name> — the newest stable tag in the tool's own
repository, spelled back as a version); other
ecosystems (conan:, julia:) work with an explicit @version.
No install needed here either: the
browser playground's Vet a
package mode runs the same lookup in your browser, and the result is a
shareable URL —
#pkg=npm:chakl
resolves the registry's "latest" for a typo of chalk and reports the
malware record attached to it. (A few registries don't answer browsers at
all — RubyGems, Maven, conda, Ansible Galaxy among them — the CLI covers everything.)
-fail-on vuln,unlisted,typosquat gates scripts the same way it gates CI,
and -md/-json output works as everywhere else.
Let your AI assistant vet dependencies (MCP server)
lockvet mcp runs lockvet as a Model Context Protocol
server over stdio, so Claude Code, Claude Desktop, Cursor, VS Code, and any
other MCP client can vet lockfile changes mid-conversation — "is this
Dependabot PR safe to merge?" becomes a question your assistant can actually
answer, with OSV data instead of vibes.
No install needed with Docker:
{ "command": "docker", "args": ["run", "-i", "--rm", "ghcr.io/matteo-sung/lockvet:0.6.15", "lockvet", "mcp"] }.
lockvet is also on the official MCP Registry
as io.github.matteo-sung/lockvet,
so clients that browse the registry can add it from there.
Six read-only tools, mirroring the CLI:
| Tool | What it does |
|---|---|
vet_url | vet any PR/MR, compare range, or commit by URL — GitHub, GitLab, Bitbucket, Gitea/Forgejo, Azure DevOps, no clone |
vet_git | vet a local repo: working tree vs HEAD, or any revision range |
vet_files | vet two lockfiles or SBOMs on disk |
audit | audit everything the project pins right now — advisories, unlisted versions, deprecations |
vet_package | vet a dependency before installing it (npm:left-pad, pypi:requests@2.32.0) — advisories, age, deprecation, typosquat suspicion |
queue | triage every open Dependabot/Renovate PR of a repo/org in one table |
Reports come back as markdown (or format: "json" for structure); forge
tokens are read from the environment (GITHUB_TOKEN, GITLAB_TOKEN, …) so
private repos work wherever the CLI does. Try: “triage the open dependency
PRs in grafana and tell me which ones I should look at first.”
In CI (review Dependabot/Renovate PRs automatically)
lockvet posts a summary comment on any PR that touches a lockfile —
see it live on a real PR:
(Not on GitHub Actions? lockvet pr <PR-url> -comment -fail-on vuln does
the same job — fetch, report, comment, gate — from any CI with a
GITHUB_TOKEN in the environment.)
GitHub Code Scanning (SARIF)
lockvet -sarif emits SARIF 2.1.0,
so vulnerable, still-vulnerable, and deprecated incoming versions show up as
code scanning alerts — annotated on the exact lockfile line that pins the
package, with OSV links and severity. In the Action it's one input (the job
additionally needs security-events: write):
Or standalone, anywhere:
and upload with github/codeql-action/upload-sarif or
gh api repos/<owner>/<repo>/code-scanning/sarifs.
On GitLab, one line vets the MR and posts the report as an MR note — reruns update the note in place:
The -comment needs a GITLAB_TOKEN CI/CD variable (a project access token
with api scope — CI_JOB_TOKEN can't post notes). Without one, drop
-comment: fetching public MRs needs no auth, and the report still lands in
the job log. Self-hosted instances work — the host comes from the URL.
Prefer diffing the checkout instead of the API? git fetch origin "$CI_MERGE_REQUEST_TARGET_BRANCH_NAME" && lockvet "origin/$CI_MERGE_REQUEST_TARGET_BRANCH_NAME" does the same locally.
On Bitbucket, the same one-liner runs in Pipelines:
For -comment, set a BITBUCKET_TOKEN repository variable (a repository
access token with pull request: write scope). Without it, drop -comment
and the report lands in the pipeline log.
On Azure DevOps, add a PR-triggered job that comments the report on the pull request:
The build service account needs Contribute to pull requests on the repo
for -comment; without it, drop -comment and the report lands in the
job log.
And on Codeberg (or any Gitea/Forgejo with Woodpecker CI):
As a pre-commit hook
Catch a risky bump before it's even committed — lockvet's default mode
(working tree vs HEAD) is exactly "what this commit changes", and the hook
only fires when a lockfile (or a workflow file with uses: pins) is part of
the commit:
The hook always prints the explanation, and by default it blocks the commit
only on the alarming tier — introduced vulnerabilities, versions missing
from their registry index, newly added install scripts, dropped provenance,
typosquat lookalikes, and integrity/resolution tampering — the signals
tuned to be near-zero-noise. Override args to tune the gate (or clear it:
args: [] makes the hook purely informational). Requires nothing but
pre-commit itself (the hook builds via Go, which
pre-commit downloads automatically if missing).
Acknowledging findings — .lockvetignore
A gate you can't quiet gets turned off. When you've looked at a finding
and accepted it — a CVE that doesn't apply to your usage, a major bump you
planned, a deprecated package you're migrating off next sprint — record
the decision in a .lockvetignore file next to your lockfiles and the
finding stops counting toward the summary and -fail-on:
Rules are an advisory ID, a pkg[@version], or a kind:pkg[@version]
where kind is one of vuln, fresh, deprecated, unlisted,
scripts, provenance, integrity, registry, license, major,
downgrade. An
until=YYYY-MM-DD expiry makes the acknowledgement temporary — after
that date the rule stops applying and every run warns until the line is
removed or extended, so snoozes can't quietly become forever.
Suppression is honest: ignored findings still show up, as a dim
○ ignored (.lockvetignore) marker in reports and as ignored /
ignored_vulns in -json, and the summary says how many findings were
acknowledged. The file is discovered automatically in the current
directory (or the audited tree for lockvet audit), works in every mode
including lockvet pr <url> and the GitHub Action, and can be pointed
elsewhere with -ignore-file <path> or switched off with -no-ignore —
so CI can enforce "no ignores" if that's your policy. queue mode spans
many repositories and applies no ignore file.
Deprecations and license changes
Every incoming version is checked against its registry (via deps.dev):
-
deprecated — the registry marks the version deprecated; the upstream reason is shown inline (
● deprecated upstream: use String.prototype.padStart()). For PyPI this also covers yanked releases (with the yank reason) and PEP 792 project statuses: a project archived by its maintainers or quarantined by PyPI's admins — the malware-review state — is flagged on every change that still pins it. crates.io yanks surface the same way, verified against the sparse index itself so even a yank minutes old is caught. Composer packages get the same treatment straight from Packagist: a bump onto an abandoned package is flagged with the maintainer's suggested replacement (● deprecated upstream: abandoned; use symfony/mailer instead). NuGet deprecations come straight from the registration index too, including the replacement package deps.dev drops (● deprecated upstream: legacy; use Azure.Storage.Common instead), and author-unlisted versions surface in this lane as well. Hex retirements — the BEAM world's per-version deprecation — land here too, straight from hex.pm, with the maintainer's reason and message (● deprecated upstream: retired: deprecated — Not really maintained, please check out Tesla). Dart packages get theirs straight from pub.dev: a bump inside a package the publisher discontinued is flagged with the named replacement (● deprecated upstream: discontinued on pub.dev; replaced by flutter_markdown_plus), and a bump onto a retracted version — onedart pubitself refuses to newly resolve — is flagged too. JSR packages get theirs straight from jsr.io: bumps onto yanked versions and packages archived on jsr.io land in this lane as well. Go retractions land here straight from the module proxy — a bump onto a version its author retracted shows the rationale comment from the module's owngo.mod(● deprecated upstream: retracted: https://github.com/klauspost/compress/issues/1114), and modules with a// Deprecated:notice are flagged even before deps.dev re-indexes them. CocoaPods deprecations come from the pod's own podspec on the CDN —pod trunk deprecaterewrites it for every version — including the named successor (● deprecated upstream: deprecated on CocoaPods; in favor of FirebaseCrashlytics). Terraform/OpenTofu providers get theirs from the registries themselves: providers the registry warns about, providers delisted from registry.terraform.io, providers the OpenTofu registry has blocked, and HashiCorp's archived providers — replacement suggestion included (● deprecated upstream: This provider has been archived. Please use the templatefile function or the Cloudinit provider instead). Maven artifacts get theirs from the POM itself: a bump onto a relocation stub — the way Java projects announce moved coordinates — is flagged with the new coordinates and the author's message (● deprecated upstream: relocated to com.mysql:mysql-connector-j — MySQL Connector/J artifacts moved to reverse-DNS compliant Maven 2+ coordinates.), which deps.dev doesn't model at all. R packages get theirs straight from CRAN: a bump inside a package CRAN has archived — removed from the index, soinstall.packages()stops resolving it — is flagged (● deprecated upstream: archived on CRAN (no longer installable from the index)). Haskell packages get theirs straight from Hackage: bumps inside a package on Hackage's deprecation list are flagged with the maintainer's suggested replacements (● deprecated upstream: deprecated on Hackage; use crypton, cryptohash-md5 or cryptohash-sha1 instead— thecryptonitestory), and a bump onto an individual version marked deprecated via preferred-versions — Hackage's yank equivalent, which solvers avoid — is flagged too. Bazel modules get theirs straight from the Bazel Central Registry: a bump onto a yanked version is flagged with the registry's own reason (● deprecated upstream: version 3.19.0 is yanked from the Bazel Central Registry: CVE-2022-3171— BCR yanks releases for CVEs and broken artifacts), and a lockfile built with--allow_yanked_versionsadmits its yanked selections inselectedYankedVersions, which lockvet reads even offline. Conda packages get theirs straight from anaconda.org: a bump onto a release whose artifacts carry thebrokenlabel — how conda-forge pulls a bad or malicious build without deleting it — is flagged (● deprecated upstream: marked broken on conda-forge (artifacts moved to the broken label)), with all-builds-broken and some-builds-broken worded apart. -
license change — the incoming version is published under a different license than the one it replaces:
Relicensing mid-stream (MIT → BUSL, SSPL, "non-standard", …) is exactly the kind of thing nobody spots in a 40-line lockfile diff. lockvet only claims a change when the registry reports a license for both sides. JSON output carries
old_license/new_licenseon every covered change.
Both are gates too: -fail-on deprecated,license.
Versions missing from the registry
When an incoming version is unknown to the registry index even though other versions of the same package are listed, lockvet says so:
Why this matters: when a registry pulls a malicious release, this is what
the hole looks like. Every malicious version in
our case studies — event-stream@3.3.6,
flatmap-stream@0.1.1, chalk@5.6.1, ultralytics@8.3.41, all six
versions of the August 2026 keyv/Cacheable worm replay — was
unpublished after the attack, so any lockfile still pinning one references
a version its own registry has disowned. lockvet flags that without
needing an advisory to exist yet. Here is the Sept 2025 chalk + debug npm
takeover replayed during its two-hour live window:
To keep it honest, the flag is deliberately conservative — it stays silent for:
- packages the registry doesn't index at all (private registries, uncovered ecosystems) — only packages whose other versions are listed can be flagged;
- workspace members, git and path dependencies (the lockfile itself says they don't come from the registry);
- Go pseudo-versions and pnpm-style decorated version strings.
For npm, PyPI, crates.io, RubyGems, Packagist, NuGet, Hex, Pub, JSR,
CocoaPods, Go, Maven, Hackage, Bazel and conda packages — and
Terraform/OpenTofu providers — the flag is double-checked against the
registry itself: deps.dev can lag
the registries by days, so before claiming anything lockvet fetches the
package's real version list from registry.npmjs.org / PyPI's simple
API / the crates.io sparse index / the RubyGems compact index /
Packagist's Composer metadata endpoint / NuGet's registration index /
hex.pm's and pub.dev's packages APIs / jsr.io's meta.json (the same
document Deno itself resolves against; yanks stay listed there, so
absence is real signal) / the sharded CocoaPods CDN index
(the same file pod install resolves against) / the Go module proxy /
the Terraform registry's per-version endpoint and the OpenTofu registry's
version index / the per-version POM on Maven Central (falling back to
Google's Maven repository, where the androidx world lives) / Hackage's
per-package version map (deprecated versions stay listed there, so
absence is real signal) / the Bazel Central Registry's per-module
metadata.json (yanked versions stay listed there — and the registry is
a git repository, versions are added, never silently dropped — so
absence is real signal) / anaconda.org's per-release endpoint (channels
pull malicious uploads outright, and marking a build broken keeps it
listed, so absence is real signal — claimed only after a HEAD on the
package document proves the package itself exists), and
clears the flag for any version the registry serves. What survives is a
version the registry itself no longer lists — and that distinction has
teeth: on crates.io yanked versions stay in the index while deleted
(malicious) ones vanish entirely, and on RubyGems a yank removes the
release from the index altogether, so a bump onto a yanked or
admin-deleted gem keeps the flag (replaying the 2019 strong_password
0.0.7 hijack trips it today). It isn't only malware, either: HashiCorp
pulled AWS provider 5.71.0 from the Terraform registry after a
regression — the tag still exists on GitHub, and a lockfile pinning
5.71.0 gets the ▲ flag from the registry's own 404.
NuGet is the one registry where "unlisted" is a native concept, and
lockvet splits it the way NuGet does: a stable version absent from the
registration index entirely — what an admin-deleted (malicious) package
looks like — keeps the ▲ flag, while a version its author merely
unlisted
(hidden from search, still restorable) lands in the deprecation lane
instead. Absent prereleases are cleared rather than flagged: on NuGet
those are overwhelmingly CI-feed daily builds (Roslyn nightlies and
friends) that packages.lock.json cannot attribute to their real feed.
A release published minutes ago may also not be indexed yet — the flag
tells you to look, not to panic. Gate on it with -fail-on unlisted;
JSON carries unlisted / unlisted_versions; SARIF emits an
unlisted-version warning; queue sorts affected PRs to the top.
Typosquat suspects
The oldest trick on every registry: publish lodahs, reqeusts or
rustdecimal and wait for a typo. lockvet flags new dependencies whose
name is one edit away from a popular package on the same registry — when
the release is also young (≤ 30 days) or of unknown age:
That example is the real 2019 PyPI attack — and it never got an OSV
advisory, so this flag is the only thing that catches it. The 2022
rustdecimal crates.io attack and npm's lodahs trip it too.
The check is entirely local: the popular-package lists (npm's
high-impact list, the
top PyPI packages,
crates.io's most-downloaded, RubyGems' most-downloaded (via
ecosyste.ms), and Packagist's
most popular) are embedded in the
binary, so it works with -offline and in the browser playground alike.
The Ruby side replays the Feb 2020 RubyGems campaign's lead example
(rspec-mokcs for rspec-mocks — 760+ malicious gems in one sweep).
Noise control, as always, is the point:
- only packages entering the tree are checked — a bump can't change its name, and a name that has coexisted with its popular neighbour for years is an unfortunate name, not an attack (age gate);
- name pairs the registry itself treats as the same package never flag
(PyPI's
-/_/.equivalence, crates.io's-/_collision ban) — while npm/RubyGems/Packagist separator swaps, which are distinct packages, do (rack_cachenext torack-cacheflags); - the added package must not itself be on the popular list, and very short names are skipped.
Gate on it with -fail-on typosquat; acknowledge a deliberate near-name
with a typosquat:pkgname line in .lockvetignore; JSON carries
typosquat_of; SARIF emits a typosquat-suspect warning; queue sorts
affected PRs to the top.
GitHub Actions workflows are lockfiles too
Every uses: line pins a dependency, and Dependabot/Renovate bump those
pins like any other. lockvet reads workflow files —
.github/workflows/*.yml, composite action.yml files, Gitea/Forgejo
workflow dirs too — in every mode: lockvet pr <url>, local diffs,
queue, audit. Action bumps get context no plain diff shows:
-
SHA pins resolve to releases. lockvet fetches the action repository's tags over anonymous git smart-HTTP (one GET per repo, no API, no rate limits) and reports the release each commit stands for — Renovate digest bumps become readable:
-
Floating majors resolve too.
v4is reported as the release it points at today, and the jump is classified from the real versions — sov5 → v7shows asv5 (=v5.1.0) → v7 (=v7.0.1) MAJOR, with the verified compare link. -
Advisories, evaluated properly. OSV.dev has a "GitHub Actions" ecosystem but its API cannot match versions against those advisories server-side — lockvet fetches the affected ranges and evaluates them itself, against the resolved release. A SHA pin affected by a GHSA is caught even though no advisory ever names that hash, and a floating
v4isn't false-flagged for an advisory fixed inside the major. -
▲ not a release. A pinned commit that matches no tag in the action's repository (and isn't a branch or its head). That is exactly what the March 2025 tj-actions/changed-files attack looked like: version tags across the repo were force-moved to a malicious orphan commit. Replaying the attacked pin today:
(
0e58ed8is the actual malicious commit from that attack.)-fail-on unlistedgates on it in CI.
Branch pins (@main) are a deliberate choice and stay quiet, actions
living outside github.com make no unlisted claims, and -changelogs
shows the release notes of every action release a bump pulls in.
…and so is .pre-commit-config.yaml
Every repos: entry pins a hook repository at an exact rev: — code that
pre-commit clones and runs on every commit on every contributor's
machine. pre-commit autoupdate and Renovate bump these pins like any
other dependency, and they get the same treatment as workflow pins, on any
git forge (names keep their host):
SHA revs are resolved to the release they equal, jumps are classified from
the real versions, -changelogs pulls each hook's release notes, and a rev
that matches no tag in the hook's repository raises ▲ not a release —
the same signal that catches the tj-actions attack shape, for the code your
whole team runs before every commit. repo: local and repo: meta entries
are exempt, and lockvet pkg pre-commit:owner/repo vets a hook repo before
you add it.
…and .gitlab-ci.yml
GitLab pipelines pin dependencies in two places, and lockvet reads both —
in .gitlab-ci.yml, suffix-named variants (backend.gitlab-ci.yml), and
CI fragments under .gitlab/ or .gitlab-ci/ directories:
-
include: component:pins — CI/CD Catalog components whose configuration runs in every pipeline. Each pin is verified against the component project's real tags, and GitLab's floating forms resolve to the release they mean today:@2and@2.0(semver range shorthands) and@~latestall show the concrete version, so a bump like@2 → @~latestis readable instead of opaque:A pinned version that matches no tag in the component's project raises ▲ not a release — the tj-actions attack shape, for GitLab. Compare links use the project's real tags, and
-changelogsrenders the component'sCHANGELOG.mdsections for exactly the versions the bump pulls in (GitLab-hosted repos included).lockvet pkg component:gitlab.com/components/opentofu/full-pipelinevets one before you include it. -
image:andservices:refs — the containers every job runs in get the same registry verification as DockerfileFROMpins (digest vs. tag, unknown tags, Docker Hub ages). Renovate bumps both kinds; nothing else vets either.
include: project: + ref: pins are tracked as version rows but stay
claim-free: the file never records which GitLab instance hosts the
project, so lockvet won't guess. $CI_SERVER_FQDN-prefixed component
pins are claim-free too when read from disk — but in mr, compare,
and queue modes the URL itself names the instance, so lockvet resolves
them against that host and they get the full tag-verification treatment.
Templated refs ($VARIABLES) and include: local:/template:/remote:
entries pin nothing.
…and .circleci/config.yml
CircleCI pipelines pin dependencies in two places, and lockvet reads both —
in .circleci/config.yml and in continuation configs / fragments under the
.circleci/ directory:
-
orbs:pins — reusable CI packages from the CircleCI orb registry, whose commands run in every pipeline. There is no OSV ecosystem and no deps.dev system for orbs, so lockvet asks the registry itself: release ages (a bump onto an hours-old orb release is worth a pause), verified compare links and-changelogsrelease notes from the orb's own source repository, and — because published orb versions are immutable, deleting one is a CircleCI-support action reserved for security problems — a registry-verified ▲ not in registry index check for pins the full version list omits. Floating pins resolve to the release they fetch today:dev:*versions are mutable by design and expire — they stay claim-free, as do inline orb definitions and templated references.lockvet pkg orb:circleci/nodevets an orb before you add it. -
docker executor
- image:refs — the containers every job runs in get the same registry verification as DockerfileFROMpins (digest vs. tag, unknown tags, Docker Hub ages).machine:executor images are CircleCI VM image labels, not OCI refs, and are skipped.
Renovate's circleci manager bumps both kinds; nothing else vets either.
…and .tool-versions, mise.toml, .nvmrc & friends
The toolchain a project runs — node, python, terraform, kubectl — is
pinned in asdf/mise files, mise up and Renovate's asdf/mise managers
bump those pins like lockfile entries, and nothing vets them. lockvet
reads both formats and verifies every pin against the tool's own
repository tags, speaking each repo's tag dialect (go1.23.4, ruby's
v3_4_2, Erlang's OTP-27.1.2, jq-1.7.1, kustomize/v5.8.1,
swift-6.1-RELEASE — ~90 tools curated and continuously validated):
That (=tag) is a verified resolution: the compare link and
-changelogs release notes come from the real tag. A version-shaped pin
matching no tag flags ▲ not a release — the same check that
catches the tj-actions attack shape in workflows. mise's fuzzy pins
resolve to what they'd install today (node = "22" → (=v22.23.2)).
mise's backend-prefixed tools are real registry packages and get lockvet's
full registry treatment — npm:prettier → npm, cargo:eza → crates.io,
pipx:black → PyPI, gem:, dotnet:, go: likewise (advisories, ages,
deprecations, typosquats); ubi:/aqua:/github:/spm: tools are
verified against the named GitHub repo's tags; and gradle, maven and
sbt pins ride the registries lockvet already verifies (the Gradle
distribution index, Maven Central — where stale sbt pins carry real
GHSAs). Plugin-sourced tools (asdf:, vfox:), system, latest,
lts, ref: pins and vendor-prefixed Java builds honestly claim
nothing. lockvet pkg tool:node vets a tool release before you switch.
The same treatment covers the single-tool version files every version
manager reads — .nvmrc / .node-version (nvm, fnm, nodenv),
.python-version (pyenv, incl. multi-version fallbacks), .ruby-version
(rbenv/rvm — ruby-3.3.4 spelling understood), .go-version,
.java-version, .terraform-version / .terragrunt-version (tfenv,
tgenv) — plus SDKMAN's .sdkmanrc, whose gradle=/maven=/sbt= pins
ride the registries lockvet already verifies. Renovate bumps all of these
with dedicated managers; nothing else vets the bumps.
And mise.lock — mise's own lockfile — gets the full lockfile
treatment: beyond exact resolved versions, it records per-platform
sha256/blake3 checksums of the artifacts mise downloads, and lockvet
reads them as integrity pins. A checksum that changes while its
version stays put is the poisoned-download shape and flags
‼ REPINNED; newly added platforms never flag (artifact-scoped
comparison, the same rule Gradle verification metadata gets):
Why this class of pin matters is case study 11
— the Codecov bash uploader replayed as a toolchain checksum swap,
including the Gradle-wrapper variant where lockvet cross-checks the pinned
distributionSha256Sum against the checksum Gradle actually publishes.
Container base images
Dockerfiles pin dependencies too — the FROM line is a version pin like any
other, Renovate and Dependabot bump it like any other, and nobody reviews it
like any other. lockvet reads Dockerfile / Containerfile (multi-stage
aware; ARG defaults expanded; variant names like Dockerfile.alpine and
dev.Dockerfile included), COPY --from= images, the # syntax= BuildKit
directive, and every image: in Compose files — then verifies the change
against the image registry itself, since whole images have no OSV
ecosystem and no deps.dev system:
- Digest pins are checked against what the tag serves today — an
exact match (index digest or any per-platform manifest digest) shows
✔ digest verified; a pinned digest the tag no longer serves lands in the‼ tag mismatchlane (image tags do move — a rebuilt base looks like this, and so does a pin that never came from the tag); a digest the registry has never seen at all is▲unlisted. - A tag the registry doesn't serve — while the image repository itself
is known — is what a deleted tag, a wrong repository, or a fabricated
reference looks like:
▲ not in the registry,-fail-on unlistedgates on it. - Same-tag digest bumps render as routine
↻(that's just the registry rebuilding the tag), with the new pin verified — not as an alarming repin. - Release ages and the ⏱ fresh flag come from Docker Hub's per-tag
last_updatedfor Hub-hosted images.
Lookups run against a fixed allowlist of public registries (Docker Hub,
ghcr.io, quay.io, mcr.microsoft.com, gcr.io, registry.k8s.io — with
legacy k8s.gcr.io pins verified against registry.k8s.io, where every
request redirects — public.ecr.aws, registry.gitlab.com). Images from any other host — private
registries, mirrors, localhost — are never queried: lockvet doesn't
send requests to hosts it doesn't recognize. Compose services built from a
local build: context, stage-name references, and references that still
contain an unresolved ${VARIABLE} are skipped rather than guessed at.
Dev Containers are image pins too. devcontainer.json (and
.devcontainer.json, subfolder variants included) names the container
your editor and Codespaces actually build: the "image" is verified
exactly like a Dockerfile FROM, and every Feature under "features" —
ghcr.io/devcontainers/features/go:1 and friends — is an OCI artifact
pulled through the same distribution API, so its tag (or @sha256:
digest pin) is verified against the registry the same way. Renovate's
devcontainer manager bumps exactly these references; nothing else vets
them. JSONC is understood (the VS Code templates ship with comments);
compose-/Dockerfile-based configs, local-path Features, tarball URIs and
legacy host-less Feature ids honestly pin nothing here.
Kubernetes manifests & kustomizations
The same treatment covers what your cluster pulls. Every image: under a
containers: / initContainers: / ephemeralContainers: list is a pin
(Deployments, StatefulSets, CronJobs — any nesting), and
kustomization.yaml images: transformers (newTag: / digest: — the
values Renovate and Flux image automation bump) are pins too; both get the
full registry verification above. kustomization.yaml helmCharts:
entries are checked against the chart repository's own index.yaml,
exactly like Chart.yaml dependencies (release ages, deprecated charts,
the prune-guarded unlisted check).
Flux CRs are pins too. A HelmRelease whose spec.chart.spec.version
is an exact pin becomes a Helm package: when the sourceRef's
HelmRepository lives in the same file, its URL comes along and the bump
is verified against that chart repository's index.yaml (ages, deprecated
charts, unlisted versions, -changelogs); a sourceRef defined in another
file — or an OCI or Git source — still yields the version row and jump
classification, just no registry claims (there is no index.yaml to
honestly check). Modern chartRef setups are covered from the other side:
an OCIRepository with spec.url: oci://… plus spec.ref.tag /
spec.ref.digest is exactly an image pin, so Renovate's ref.tag bumps
get the Dockerfile registry treatment above — including ✔/‼ digest
verification. Version ranges (spec.ref.semver, >=1.0.0) pin nothing
and are skipped. And inside a HelmRelease's values:, the standard
chart image mapping — image: with repository: and tag: children
(Bitnami-style registry: and digest: too) — is read as an image pin:
that's the exact shape Flux image automation and Renovate bump in
home-ops repos, old Flux v1 Auto-release commits included.
Argo CD Application CRs too. An Application whose spec.source
carries chart: pins that chart at targetRevision:, and the inline
repoURL: is the chart repository itself — so the bump is verified
against that repository's index.yaml the same way (ages, deprecated
charts, unlisted versions, -changelogs). Multi-source spec.sources
lists work per item, and ApplicationSet templates
(spec.template.spec.source/.sources) get the same reading — with
{{ … }} parameters treated as what they are, not literal pins. Only
exact revisions count: a targetRevision like 1.2.*, >=1.0.0, *
or a branch name tracks the repository and pins nothing, and Git-source
Applications (path: instead of chart:) are left alone. OCI
repoURLs yield the version row without registry claims (no
index.yaml to check).
Helm values files too. The same chart image convention — an image:
mapping with repository: and tag: children (Bitnami-style
registry: and digest: too) — is read straight out of values files:
values.yaml at chart roots, values-prod.yaml / app.values.yaml
overlays, and arbitrarily named helmfile or Argo value overlays under
GitOps directories. That's the exact shape Renovate's helm-values
manager bumps, and nothing else vets it — the values file names no chart
version and charts have no OSV ecosystem, but the image pins inside get
the full registry treatment above (digest verification, unknown tags,
ages). Convention-named values files also accept the bare scalar form
(image: ghcr.io/owner/app:v1.2.3) when a tag or digest is present —
image: nginx pins nothing; files merely discovered by directory
convention or content sniffing require the structured
repository:/tag: shape, which is distinctively Helm. Block-scalar
bodies (config: | — embedded app configuration) are opaque: an
image: line inside one belongs to the embedded app, never to this
file's pins.
Since plain manifests have no reserved filename, lockvet takes
*.yaml/*.yml files under the conventional directories (k8s/,
kubernetes/, manifests/, deploy/, overlays/, base/,
clusters/, gitops/, apps/, infrastructure/, infra/, flux/,
flux-system/, chart/, charts/, argocd/, argo/, argo-cd/,
applications/, applicationsets/), files named *.k8s.yaml, and the
conventional basenames (deployment.yaml, statefulset.yaml,
daemonset.yaml, replicaset.yaml, cronjob.yaml, pod.yaml, plus the
Flux conventions helmrelease.yaml, helm-release.yaml, release.yaml,
ocirepository.yaml, helmrepository.yaml and the Argo CD conventions
application.yaml, applicationset.yaml, appset.yaml) — strictly: a
document must
declare top-level apiVersion: and kind: to count, matched files that
aren't Kubernetes YAML are silently ignored, and anything under a Helm
templates/ directory is excluded (those image references are template
interpolations, not what the cluster resolves). Image values still
carrying {{ … }}, $(VAR) or ${VAR} are skipped rather than guessed
at.
In diff modes (local git, pr/mr/compare, queue), discovery
goes further: every other changed *.yaml/*.yml is content-sniffed
with the same strict gate, so GitOps repos with arbitrary layouts
(default/nzbget/nzbget.yaml) work without matching any convention —
the changed-file list is small, so over-matching costs nothing. Remote
fetches cap sniff candidates at 20 files per diff to keep big refactor
PRs cheap. Directory walks (lockvet audit) keep the convention-based
discovery: sniffing every YAML in a large tree would read far more than
it finds.
Integrity & resolution changes
Lockfiles don't just pin versions — they pin bytes (integrity hashes) and where to get them (resolution URLs). lockvet is, as far as we know, the only lockfile differ that reads them:
Integrity changed, version didn't ‼ — registries never change a
published artifact. If a diff keeps lodash 4.17.21 but swaps its
sha512-…, the tarball that pin expects has been replaced: registry-side
tampering, a hijacked mirror, or a hand-edited lockfile. All of them are
"stop and find out why" events:
Hashes are compared per algorithm, so a lockfile-format upgrade that
switches sha1 → sha512 (or a yarn berry cache-key bump) never flags,
and Python hash sets may legitimately grow as wheels are added — only a
fully disjoint set flags. Each shared algorithm is judged on its own: the
same bytes can never hash two ways under one algorithm, so a yarn-classic
pin whose SRI sha512 was swapped flags even when the sha1 resolved
fragment was left intact as camouflage (yarn verifies the SRI line at
install time, not the fragment) — and a conda lock's md5 swapped under
an intact sha256 line flags the same way (conda locks may carry either
hash or both; each is compared under its own label). Hex's dual checksums
get the same treatment within one algorithm: mix.lock and rebar.lock
record two sha256 values per package that digest different byte streams —
the tarball contents (inner, pkg_hash) and the tarball file (outer,
pkg_hash_ext) — so each is scoped on its own
(contents#sha256:/tarball#sha256:) and a one-sided swap flags instead
of hiding behind its untouched sibling; the old-format → new-format
migration that adds the outer checksum stays quiet. The one pooled exception is Terraform's
h1/zh pair: h1 hashes exist only for the platforms the lock was
built with, so a platform re-lock that replaces the whole h1 set stays
quiet while the registry-published zh set still vouches for the release.
Integrity removed, version didn't ‼ — a pin that carried a content hash and now carries none is just as loud. Nothing verifies the artifact anymore, and a hash deleted — or swapped to a malformed value no algorithm claims — is how a tampered pin stays quiet:
Two legitimate shapes stay silent: a dependency switched to a git/path override at the same version (the pin moved to a commit — registry hashes rightly vanish), and a whole file losing its hashes at once (a lockfile regenerated by tooling that writes none — a config change, not a per-package attack).
Resolution moved to the public registry ⇄ — the classic
dependency-confusion
attack publishes a higher version of an internal package name on the
public registry, and the resolver takes the bait. In the lockfile diff
that is perfectly visible: the package's resolved host flips from your
private registry to registry.npmjs.org / pypi.org / crates.io /
rubygems.org:
Only the private → public direction flags (adopting a mirror is routine), a move whose content hash provably didn't change is quiet, and when five or more packages move between the same two hosts lockvet treats it as a registry migration (a config change) and stays quiet.
The same check reads release-tag pins: Package.resolved (Swift) and
composer.lock record the commit a released tag pointed at, and Julia's
Manifest.toml records the registry's source-tree hash — a version that
keeps its number but changes its commit means the upstream tag was
moved, and flags the same way.
Both signals are computed entirely offline from the two lockfiles —
they work with -offline, in air-gapped CI, and in the browser
playground. Formats that record the data: npm (v1–v3), pnpm, yarn
(classic + berry), bun.lock, deno.lock (npm + jsr), Cargo, poetry, uv,
Pipfile.lock, requirements.txt --hash, pylock.toml (PEP 751 artifact
hashes and index hosts — the Python confusion shape flags too),
Gemfile.lock (resolution hosts, plus per-gem sha256 from the Bundler ≥ 2.6
CHECKSUMS section — platform gems keep their own hash),
go.sum (parsed as a pins-only ledger: version churn is go.mod's story
and produces no duplicate rows, but a same-version h1: hash edit — the
poisoned-go.sum shape, the only way a tampered module gets past
go mod verify, and for GOPRIVATE modules the only check there is —
flags ‼, with the SARIF alert anchored on the exact edited line),
mix.lock, rebar.lock (pkg_hash/pkg_hash_ext), Gleam's manifest.toml, pubspec.lock (hashes and hosts —
the Dart confusion shape flags too), Podfile.lock (trunk podspec
checksums), Package.resolved, composer.lock, Julia Manifest.toml,
.terraform.lock.hcl (provider h1:/zh: hashes), conda/pixi
locks (PyPI wheel hashes + channel hosts), MODULE.bazel.lock (the
registry source.json hash per selected module — it changing for the
same version means the registry's description of that release changed
underneath you — plus registry hosts, so a module moving from a private
registry onto the public BCR flags the confusion lane; old-style
Bazel 7.0/7.1 lockfiles pin the archive's own SRI hash), build.zig.zon
(Zig: no registry exists at all, so the manifest's own url + hash pins are
the whole story — a same-version hash swap or a same-version source
re-point is exactly what a smuggled edit looks like, while swapping a
dependency to a fork with a new pin is routine, visible business and
stays quiet), and flake.lock (Nix: a
same-revision narHash change means the pinned tree was replaced — a
git revision's content never changes — and an input re-pointed at a
different repository flags the ⇄ lane; a re-point whose narHash proves
the content identical, like a plain repo rename, stays quiet). Two deliberate omissions,
because real history proved them noisy: NuGet's contentHash (NuGet's
2018 repository-resigning changed every older package's hash) and conda
artifact hashes (conda rebuilds the same version under new build
numbers routinely). Gate with -fail-on integrity,registry; JSON carries integrity_changed /
integrity_changed_versions / old_host / new_host / registry_moved;
SARIF emits integrity-changed (error) and registry-moved (warning);
queue sorts affected PRs to the top. A full replay of the 2021
dependency-confusion attack — including the same-version variant that no
other diff shows — is case study 9.
SwiftPM pins verified against upstream tags
One check in this family is online. Package.resolved records a
version, the commit its tag resolved to, and the repository it all
came from — which makes the pin verifiable against the source of truth.
lockvet fetches the repository's real tag list (one anonymous git
smart-HTTP request per repo, the same channel git fetch uses — no API,
no rate limits) and checks two things for every incoming pin:
- the version's tag still exists. Version pins only ever resolve from tags, so a pinned version with no matching tag upstream means the tag was deleted or renamed after someone resolved it → ▲ not a release;
- the pinned commit is what the tag points at today. Released tags are supposed to be immutable — a mismatch means the tag has been re-pointed since this was resolved (the tj-actions attack shipped exactly like that), or the lockfile was edited to fetch a different commit while displaying an innocent version → ‼ tag mismatch:
The same tag-list verification backs workflow uses:
pins. Annotated tags are
peeled to their commit, v-prefixed tags match unprefixed versions, and
repositories that can't be fetched anonymously (private, moved) produce
no claims at all. Gate with -fail-on integrity (tag mismatch) and
-fail-on unlisted (deleted tag); SARIF emits tag-mismatch (error);
lockvet audit verifies every Swift pin you currently hold; and
lockvet pkg swift:Alamofire/Alamofire vets a Swift package (latest =
highest stable tag) before you add it.
Install scripts added by a bump
npm packages can run arbitrary code at install time (preinstall /
install / postinstall). Most packages never do — so a routine-looking
bump that suddenly adds install scripts deserves a hard look before you
merge. Gaining execution-on-install is how the
Shai-Hulud worm —
and its August 2026 keyv/Cacheable return —
and plenty of smaller npm attacks delivered their payload.
lockvet asks the npm registry which versions run install scripts and flags the transition:
Only transitions are flagged, so the signal stays quiet: a brand-new
dependency with install scripts is ordinary (native builds), and a package
that has always had them tells you nothing new. In a 92-change npm/cli
release diff and a 165-change React one, it flags nothing at all.
Gate on it with -fail-on scripts; JSON carries install_scripts_added /
scripted_versions; SARIF emits an install-scripts-added warning;
queue sorts affected PRs to the top. npm-only for now (other ecosystems
either have no install hooks or don't expose them in registry metadata).
Provenance dropped by a bump
A growing share of npm packages publish with sigstore provenance attestations, PyPI projects with PEP 740 attestations, crates.io crates via trusted publishing, and RubyGems releases with sigstore attestations — cryptographic proof (or, for crates.io, a registry-recorded guarantee) that the release was built and published by the project's own CI from its public repo. That proof has a useful property for reviewers: a stolen publish token can upload a release, but it can't make the project's pipeline attest it.
So when a package that consistently attests suddenly publishes a version with no provenance, lockvet flags it:
Three conditions keep it near-silent: the outgoing version must be
attested, the package's provenance practice must be established (the
top stable versions below the incoming one all attested — one-off
adopters don't count), and the incoming release must be young
(≤ 30 days) — this is a while-it's-happening tripwire for the window
before advisories exist, not an audit of history. Under those rules it
flags nothing at all across current bumps of ~100 top npm packages,
exactly one bump across the top 1 000 PyPI packages — a real 7-day-old
release that broke its project's unbroken attestation streak (almost
certainly a benign manual publish, which is precisely the "worth a look"
this flag exists for) — and zero across the top 100 crates, a tenth of
which already publish via trusted pipelines (cc, time, getrandom,
…). It would fire the moment a token thief ships a release outside the
project's CI.
Gate on it with -fail-on provenance; JSON carries provenance_dropped
/ unattested_versions; SARIF emits a provenance-dropped warning;
queue sorts affected PRs to the top. Comes from the same registry
documents as the install-scripts and unlisted checks — no extra
requests. Covers npm, PyPI, crates.io and RubyGems (for PyPI a version
only counts as attested when every file of the release carries
provenance, and only a release with no attested files is ever flagged
— mixed uploads are a publishing setup, not a signal; for crates.io the
registry's trustpub_data on each version is the source of truth; for
RubyGems it's the per-release attestations API — adoption there is
still small, so today the gates simply stay silent, and protection
switches on automatically as gems start attesting).
Release notes inline (-changelogs)
Every bump already links to the verified upstream tag-to-tag diff. Add
-changelogs and lockvet also fetches the release notes — including the
releases a multi-version jump skips over:
So 1.19.0 → 1.20.1 shows what landed in 1.19.1, 1.19.2, 1.20.0, and
1.20.1 (up to 5 releases per package; the compare link covers the rest).
In -md mode each package's notes become a collapsed <details> block —
in a PR comment that reads like Dependabot's release-notes section, except
it covers every package in the diff, transitives included. Works in
every mode (pr, mr, compare, diff, the Action via
changelogs: 'true', MCP via "changelogs": true).
Notes come from GitHub Releases of the (verified) source repo — and when a
project doesn't publish releases (Phoenix stopped in 2020; plenty of crates
and gems never started; GitLab-, Codeberg- and Bitbucket-hosted projects
don't have them), lockvet falls back to the repository's changelog file:
CHANGELOG.md / CHANGES.md / NEWS.md / HISTORY.md fetched at the
verified tag, sliced into the exact version sections your bump pulls in.
Excerpts are trimmed and control-character-sanitized either way. The
releases path uses the GitHub API — anonymous works for a handful of
packages, GITHUB_TOKEN / a logged-in gh raises the limit; changelog
files come off the forge's raw endpoint, which is not rate-limited.
Supported lockfiles
| Ecosystem | Files |
|---|---|
| JavaScript | package-lock.json, npm-shrinkwrap.json, pnpm-lock.yaml, yarn.lock (v1 & berry), bun.lock, deno.lock — JSR packages in deno.lock get release ages, yank/archive flags and the registry-verified unlisted check straight from jsr.io |
| Rust | Cargo.lock |
| Python | uv.lock, poetry.lock, pdm.lock, Pipfile.lock, requirements.txt (== pins), pylock.toml + named pylock.*.toml (PEP 751 — the standardized lockfile written by uv, pip, pipenv and pdm; vcs/directory/path sources exempt from registry checks) |
| Go | go.mod — replace directives resolve to the effective pin: replace m => m v1.2.3 overrides the required version (a replace-directive downgrade renders as a downgrade, advisories and all), replace a => b v1.2.3 renders as a removed + an introduced (the resolution-hijack shape), and filesystem replaces (=> ../local) stay exempt from registry claims — and go.sum as a pins-only ledger: version churn stays go.mod's story, but a same-version h1: hash edit (the poisoned-go.sum shape: a tampered module only installs if go.sum is edited to agree, and for GOPRIVATE modules no checksum database will ever object) flags ‼ |
| PHP | composer.lock |
| Ruby | Gemfile.lock (including Appraisal's *.gemfile.lock variants; Bundler ≥ 2.6 CHECKSUMS hashes become integrity pins) |
| Elixir | mix.lock — release ages, retirements and the unlisted check straight from hex.pm; renamed forks ({:hex, :ts_chatterbox, …}) resolve under their real Hex package name |
| Erlang | rebar.lock — Hex advisories, ages/retirements/unlisted from hex.pm, direct deps from the lock's own level numbers, pkg_hash/pkg_hash_ext checksums as integrity pins each scoped on its own — a same-version change of either hash surfaces as ‼ repinned, even with the other left intact; renamed forks ({<<"uuid">>,{pkg,<<"uuid_erl">>,…}}) resolve under their real Hex package name; git/path deps exempt |
| Dart / Flutter | pubspec.lock — release ages, discontinued/retracted flags and the unlisted check straight from pub.dev (git/path/SDK/private-host packages exempt) |
| Java / JVM | pom.xml (Maven has no npm-style lockfile — POM versions are the pins, and Dependabot bumps them in place: ${property}-resolved versions, <parent> bumps like spring-boot-starter-parent, BOM imports, plugins and profiles; ${project.version} siblings honestly claim-free), build.gradle / build.gradle.kts (plus settings.gradle(.kts) and any *.gradle(.kts) script — for most Gradle projects the build script is the pin file: coordinate literals, Groovy map / Kotlin named-argument forms, and plugins { id(…) version … } blocks resolved as Plugin Portal markers; $version interpolations resolve against same-file ext/val assignments, and anything defined elsewhere honestly claims nothing), gradle.lockfile, libs.versions.toml (Gradle version catalogs — the file Renovate and Dependabot actually bump, plugins resolved as their id:id.gradle.plugin marker against the Gradle Plugin Portal) verification-metadata.xml (Gradle dependency verification — the wall-of-XML diff nobody reviews becomes package changes, and a same-version checksum change surfaces as ‼ repinned) and the build-tool wrappers themselves — gradle-wrapper.properties distributionUrl pins are verified against Gradle's own version index at services.gradle.org (release ages, withdrawn-as-broken releases, versions the index has never heard of, and a pinned distributionSha256Sum cross-checked against the checksum Gradle actually publishes: the wrapper will happily verify a poisoned distribution against a poisoned checksum, so a pin matching no official checksum lands in the ‼ lane), while .mvn/wrapper/maven-wrapper.properties pins parse as ordinary Maven coordinates and get the full Maven treatment (corporate mirror hosts honestly claim-free in both) — bumps onto relocation stubs (mysql:mysql-connector-java → com.mysql:mysql-connector-j) land in the deprecation lane, and unlisted checks are verified against Maven Central, Google's Maven repository and the Plugin Portal |
| Scala | build.sbt (and any *.sbt, plus the project/Dependencies.scala convention — Scala has no npm-style lockfile, the build definition is the pin file and scala-steward bumps coordinates in place): "org" %% "artifact" % "1.2.3" resolves the Scala binary suffix from the same file's scalaVersion so the registry artifact (cats-core_2.13) gets OSV advisories, release ages and unlisted checks; version vals resolve within the file; addSbtPlugin(…) pins resolve under the sbt cross-suffix (sbt-scalafmt_2.12_1.0 — the name Central actually serves); CrossVersion.full / for3Use2_13 map to their documented suffixes; anything unknowable from the file alone (no scalaVersion in sight, %%% platform artifacts) keeps its version row but honestly claims nothing. project/build.properties pins sbt itself — org.scala-sbt:sbt on Central, with real advisories (try lockvet pkg maven:org.scala-sbt:sbt@1.10.7) |
| .NET | packages.lock.json |
| Swift | Package.resolved |
| C / C++ | conan.lock (Conan 2 flat lockfiles and Conan 1 graph locks) — release ages straight from ConanCenter, dated by each version's oldest recipe revision so re-exports don't make old releases look fresh (refs with a user/channel exempt; a ref doesn't record its remote, so no unlisted claims). The #recipe-revision in each ref is tracked as a pin: a revision change under an unchanged version — the recipe (build script) changing under your feet — renders as a neutral ↻ recipe revision row, like a container's same-tag digest bump: visible where it used to be silent, but not an alarm, because ConanCenter re-exports recipes for old versions routinely |
| vcpkg | vcpkg.json + vcpkg-configuration.json — the builtin-baseline commit IS the lockfile (every dependency resolves to what the registry recorded at that commit): baselines are verified to exist in the registry's repository (a baseline only reachable in a fork is the poisoned-registry shape) with the commit's own date as the age and rev…rev compare links; overrides version pins are checked against microsoft/vcpkg's append-only versions database; registry baselines in vcpkg-configuration.json get the same treatment, and a swapped default-registry surfaces as a resolution move (overrides under overlay-ports or custom registries honestly claim nothing) |
| Bazel | MODULE.bazel.lock (both the modern registry-hash shape and the Bazel 7.0/7.1 dep-graph shape) and MODULE.bazel itself (bazel_dep pins exact versions and update bots bump them in place, go.mod-style) — yanked versions with the registry's reason, source-repo changelog links and the registry-verified unlisted check straight from the Bazel Central Registry (modules from private registries and git_override/local_path_override exempt) |
| iOS / CocoaPods | Podfile.lock — release ages, deprecated-pod flags, license changes and the unlisted check straight from the CocoaPods CDN and trunk registry (git/path pods and private specs repos exempt) |
| R | renv.lock — CRAN + Bioconductor advisories (RSEC-…); release ages, archived-package flags, license changes, changelog links and a mirror-lag-safe unlisted check straight from CRAN itself (GitHub/GitLab/local/git installs exempt) |
| Julia | Manifest.toml (also Manifest-v1.11.toml style) — General-registry advisories (JLSEC-…), via-chains from the manifest's dependency graph |
| Haskell | stack.yaml.lock, cabal.project.freeze, cabal.config (Stackage pins) — Hackage advisories (HSEC-…); release ages, deprecated packages (with the maintainer's suggested replacements) and deprecated versions, verified changelog links and the registry-verified unlisted check straight from Hackage |
| Gleam | manifest.toml — Hex advisories, via-chains, direct deps from [requirements]; ages/retirements/unlisted from hex.pm (git/path packages exempt) |
| Conda | pixi.lock (v4–v7), conda-lock.yml (also *.pixi.lock, *.conda-lock.yml) — release ages, broken-release flags, license changes and the registry-verified unlisted check straight from anaconda.org (any channel the lockfile resolves from — conda-forge, bioconda, …); pip packages inside get full PyPI vulnerability/age data. Artifact sha256 hashes are read as integrity pins scoped to the artifact filename (openssl-3.3.1-h4bc722e_2.conda#sha256:…), so a same-artifact hash swap flags ‼ REPINNED while conda-forge's routine same-version rebuilds under new build numbers stay quiet — and locally-built sources (path installs like - pypi: ./rerun_py, git checkouts) record no integrity at all, since their hash legitimately changes on every rebuild |
| Terraform / OpenTofu | .terraform.lock.hcl (also suffix-named *.terraform.lock.hcl) — release ages, archived/delisted/blocked-provider flags, verified changelog links and the registry-verified unlisted check straight from the Terraform & OpenTofu registries (custom registry hosts exempt) |
| Helm | Chart.lock, Chart.yaml (most repos commit only the manifest — exact version pins are read, range constraints skipped), requirements.yaml (Helm v2, content-sniffed apart from Ansible Galaxy files), requirements.lock (Helm v2; pip-style requirements.lock from rye is auto-detected and treated as PyPI) — release ages, deprecated-chart flags, verified changelog links and a prune-guarded unlisted check straight from each chart repository's own index.yaml, the exact document helm dependency update resolves against (oci:// and file:// references exempt) |
| Ansible | requirements.yml / requirements.yaml — Galaxy collections (namespace.name) and classic roles, exact version pins only (range constraints skipped); release ages, deprecated-collection flags, verified changelog links and the registry-verified unlisted check straight from galaxy.ansible.com (the v3 collection index ansible-galaxy itself resolves against, plus the classic v1 role index); git/file/url entries and private Automation Hub sources exempt; roles never get absence claims (their index only updates when the owner re-imports) |
| Nix | flake.lock |
| Zig | build.zig.zon — Zig has no lockfile beyond it and no registry: every dependency pins a source URL + content hash, and lockvet reads both (rev/tag/hash-derived versions, ‼ same-version hash swaps, ⇄ same-version source re-points, verified tag compare links + release notes from the dependency's own repo) |
| CI / GitHub Actions | .github/workflows/*.yml, action.yml/action.yaml (composite actions), .gitea/workflows, .forgejo/workflows — every uses: pin; SHA and floating-tag pins resolved against the action repo's real tags, advisories from OSV's GitHub Actions ecosystem evaluated client-side |
| CI / GitLab | .gitlab-ci.yml (suffix-named variants like backend.gitlab-ci.yml and fragments under .gitlab/ / .gitlab-ci/ too) — include: component: CI/CD Catalog pins verified against the component project's real tags (floating @2 / @2.0 / @~latest forms resolved to the release they fetch today), job/default image: and services: refs verified like Dockerfile images; include: project: + ref: pins tracked claim-free |
| CI / CircleCI | .circleci/config.yml (continuation configs and fragments under .circleci/ too) — every orbs: registry pin: release ages, floating volatile / 5 / 5.1 forms resolved to the version they fetch today, registry-verified unlisted detection and changelog links straight from the CircleCI orb registry (dev: versions and inline orbs exempt); docker executor - image: refs verified like Dockerfile images (machine: VM image labels skipped) |
| Containers | Dockerfile / Containerfile (variant names like Dockerfile.alpine, dev.Dockerfile too) and Compose files (docker-compose.yml, compose.yaml, overrides) — every FROM base image (multi-stage aware: stage references skipped, ARG defaults expanded, --platform flags ignored), COPY --from= images, the # syntax= BuildKit directive Renovate bumps, and every image: under Compose services: (services built from a local build: context skipped). Tags and digest pins are verified against the image registry itself — see Container base images |
| Dev Containers | devcontainer.json / .devcontainer.json (subfolder variants included, JSONC understood) — the "image" pin plus every OCI-referenced Feature under "features" (ghcr.io/devcontainers/features/…:tag or @sha256: digest pins), verified against the registry like Dockerfile images; local-path / tarball / legacy host-less Feature ids skipped |
| Kubernetes | manifests (*.yaml under conventional directories — k8s/, kubernetes/, manifests/, deploy/, overlays/, base/, clusters/, gitops/, apps/, infrastructure/, flux-system/, chart(s)/, … — plus *.k8s.yaml and conventional basenames like deployment.yaml, helmrelease.yaml, ocirepository.yaml — top-level apiVersion: + kind: required, Helm templates/ excluded): every image: under containers: / initContainers: / ephemeralContainers, verified like Dockerfile images. Flux HelmRelease chart version pins (same-file HelmRepository URLs verified against the chart repo's index.yaml) and OCIRepository ref.tag/ref.digest pins (verified like images). Argo CD Application / ApplicationSet chart sources (chart: + exact targetRevision:, single- and multi-source, verified against the inline repoURL's index.yaml; conventional basenames application.yaml, applicationset.yaml, appset.yaml and directories argocd/, argo/, applications/ match too). kustomization.yaml / kustomization.yml: images: transformer pins (newTag:, digest:) plus helmCharts: entries checked against the chart repository's own index.yaml. Helm values files (values.yaml, values-*.yaml, *-values.yaml, *.values.yaml — and any sniffed YAML carrying the structured shape): image: mappings with repository:/tag: (+ registry:/digest:) children, plus tagged image: scalars in convention-named files; block-scalar bodies opaque, templated values skipped — see Kubernetes manifests & kustomizations |
| pre-commit | .pre-commit-config.yaml — every repos: entry's rev: pin, on any git forge (names keep their host); revs verified against the hook repository's real tags exactly like workflow pins: SHA pins resolved to the release they equal, ▲ when a rev matches no tag, verified compare links + release notes for bumps (repo: local/meta exempt) |
| asdf / mise | .tool-versions, mise.toml (also .mise.toml, mise.local.toml, mise/config.toml locations) — every toolchain pin verified against the tool's own repository tags via a curated ~90-tool map that speaks each repo's tag dialect (go1.23.4, v3_4_2, OTP-27.1.2, REL_18_4, swift-6.1-RELEASE, …): verified (=tag) resolutions with compare links and release notes, fuzzy pins (node = "22") resolved to what they'd install today, ▲ when a concrete pin matches no tag. mise backend tools are real registry packages and get the full treatment (npm: → npm, cargo: → crates.io, pipx:/pip: → PyPI, gem:, dotnet:, go:); ubi:/aqua:/github:/spm: verify against the named GitHub repo; gradle/maven/sbt ride their existing registries; asdf:/vfox: plugins, symbolic selectors (latest, lts, system, ref:) and vendor-prefixed Java builds honestly claim nothing. Single-tool version files get the same treatment: .nvmrc / .node-version, .python-version (multi-version pyenv fallbacks included), .ruby-version (rvm's ruby-3.3.4 spelling understood), .go-version, .java-version, .terraform-version / .terragrunt-version, and SDKMAN's .sdkmanrc (gradle=/maven=/sbt=/kotlin= pins verified; vendor-suffixed Java builds claim-free). mise.lock adds per-platform sha256/blake3 checksums read as artifact-scoped integrity pins — a same-version checksum change flags ‼ REPINNED, added platforms never do |
| SBOMs | CycloneDX & SPDX JSON: bom.json, sbom.json, *.cdx.json, *.spdx.json — multi-ecosystem, incl. Alpine/Debian/Wolfi OS packages |
Notes: direct/via … origin labels appear where the lockfile records its
dependency graph: npm, pnpm, yarn, Cargo, uv, poetry, Composer, Bundler,
renv, pixi/conda-lock, Julia manifests, Gleam, CocoaPods, and Go modules (go.mod's // indirect markers give direct/transitive,
without chains). Formats that only pin flat versions (requirements.txt, mix.lock,
Gradle, …) skip the label.
Deno's jsr: packages, CocoaPods, conda channels, Terraform providers,
Helm charts, Bazel modules, Nix flakes, container base images, and
pre-commit hook pins have no OSV.dev ecosystem (yet), so those
diffs are explained without vulnerability data (Conan HAS one —
"ConanCenter" — which is still near-empty; advisories will surface on
conan.lock diffs automatically as it fills in) — but pip/pypi packages inside pixi.lock / conda-lock.yml are
matched against PyPI advisories like any other Python lockfile.
SBOM packages keep their own ecosystem per purl; OS packages get OSV data
when the purl names a release (distro=alpine-3.18.4 → Alpine:v3.18) —
Ubuntu/RPM packages are diffed without it.
Release ages / deprecations / license changes come from deps.dev, which covers
npm, crates.io, PyPI, Go, Maven, NuGet, and RubyGems — other ecosystems simply
skip those checks. PHP, the BEAM world, Dart, JSR, iOS, C/C++, Terraform, R, Haskell, Bazel, conda, Helm and Ansible are the exceptions: deps.dev has no
Composer, Hex, Pub, JSR, CocoaPods, Conan, Terraform, CRAN, Hackage, Bazel, conda, Helm or Ansible system at all, so lockvet asks Packagist, hex.pm,
pub.dev, jsr.io, the CocoaPods registry, ConanCenter, the Terraform/OpenTofu registries, CRAN (via METACRAN's crandb), Hackage, the Bazel Central Registry, anaconda.org, Ansible Galaxy and each Helm chart repository's own index.yaml directly — release ages, deprecation warnings (Composer abandoned,
Hex retirements, pub.dev discontinued packages and retracted versions,
JSR yanked versions and archived packages,
CocoaPods deprecated pods with their named replacement, archived/delisted
Terraform providers, packages archived on CRAN, packages and individual
versions deprecated on Hackage — replacement suggestions included,
Bazel module versions yanked from the Bazel Central Registry with the
registry's reason, conda releases marked broken, Helm charts marked deprecated in their repository's index, Ansible collections marked deprecated on Galaxy), changelog links and the unlisted check all work for composer.lock,
mix.lock, rebar.lock, Gleam's manifest.toml, pubspec.lock, deno.lock's jsr: packages, Podfile.lock, .terraform.lock.hcl, renv.lock, stack.yaml.lock, cabal.project.freeze, MODULE.bazel.lock, MODULE.bazel, pixi.lock, conda-lock.yml, Chart.lock, Chart.yaml and Ansible requirements.yml too (plus license
changes for Composer, CocoaPods, CRAN and conda; Hex, pub.dev, jsr.io, the Terraform registry, Hackage, the Bazel Central Registry, Helm chart indexes and Ansible Galaxy keep no usable per-release license history,
so that one check is honestly skipped there). Ansible's checks carry their own honesty rules: classic roles never get absence claims (the v1 role index only updates when the owner re-imports, so a missing version proves lag, not malice), and versions published before Galaxy's 2023 migration all carry the migration date, so historical ages are upper bounds. Helm's unlisted check carries a pruning guard: chart repositories routinely prune old releases from their index, so only a missing version that sorts at or above the oldest release the index still lists is flagged. Helm ages carry a regeneration guard too: some repositories rebuild index.yaml and stamp historical entries with the generation time (every chart would look published today), so when three or more versions of a chart share a one-hour created window those timestamps are treated as artifacts and claim no ages — versions outside the cluster keep their real ones. The Bazel Central Registry
also records no publish timestamps (they live only in its git history),
so Bazel modules honestly carry no release ages or ⏱ cooldown flags. CRAN's unlisted check is
mirror-lag-safe: before claiming anything, absence is double-checked
against cran.r-project.org itself (current release and the Archive —
CRAN archives old versions, it never deletes them, so a version in
neither place was never on CRAN).
conan.lock gets release ages only: a Conan reference doesn't record
which remote it came from, so absence from ConanCenter proves nothing and
lockvet makes no unlisted or deprecation claims there.
vcpkg manifests are baseline-pinned: the builtin-baseline commit decides
every dependency's version, so lockvet verifies the commit exists in the
registry's repository (a baseline only reachable in a fork is the
poisoned-registry shape — the build resolves on the attacker's checkout
and nowhere else), dates it, links old...new on the registry repo, and
checks overrides version pins against microsoft/vcpkg's append-only
versions database. Both claims are gated on the outgoing value resolving
(projects on overlay ports or custom registries never trip them), and
overlay-ports declared in the manifest's configuration turn override
claims off entirely.
Nix flake inputs pin git revisions, not versions — lockvet shows them as
<commit-date>.<short-rev> so the diff still reads chronologically, turns
each input's lastModified into release ages (the ⏱ cooldown flag works
on update-flake-lock PRs, fully offline), and links every bump to the
forge's rev...rev compare page so "what did nixpkgs actually change"
is one click. The narHash / re-pointed-repository checks above cover
flakes too.
Missing one you care about? Open an issue — parsers are ~50 lines each.
How it works
git diff --name-only <base> <target>finds changed lockfiles.- Each lockfile version is read with
git showand parsed intopackage → pinned versions(multiple versions per package are handled — npm nesting, Cargo duplicate majors). - The two snapshots are diffed and each change is classified with a lenient
version parser that copes with semver, Python post-releases, and Go
pseudo-versions.
Where the lockfile also records dependency edges and root deps, lockvet
BFS-walks the graph to label every change
(direct)orvia <chain>— no manifest files or network needed. - Old and new versions are checked against OSV.dev's batch API. A vulnerability that matches the new version but not the old one is introduced; the reverse is fixed; both is unresolved. Aliased advisories (GHSA/CVE/PYSEC/RUSTSEC for the same issue) are collapsed.
- Every incoming version is looked up on deps.dev's batch API for its
publish date, deprecation status, and license — departing versions are
looked up too, so a license flip between old and new gets flagged (npm,
crates.io, PyPI, Go, Maven, NuGet, RubyGems). An incoming version the
registry doesn't list — while other versions of the package are listed —
gets the
unlistedflag: that's what an unpublished (often malicious) release looks like (for npm, PyPI, crates.io, RubyGems, Packagist, NuGet, Hex, Pub, Go, Maven, CRAN, Hackage, Bazel and conda the flag is double-checked against the registry itself, which also tells lockvet when a bump suddenly adds install scripts or silently drops provenance attestations, and when a release was yanked or its project archived or quarantined). Versions younger than-fresh-days(default 7) get a ⏱ flag — supply-chain attacks are usually discovered and yanked within days of publication, so a short cooldown filters most of them out. For RubyGems the compact index's owncreated_attimes fill in ages deps.dev hasn't indexed yet, so a gem published minutes ago still gets its ⏱ flag; NuGet registrationpublishedtimes do the same for .NET packages; for Composer, Hex, Dart, JSR, R, Haskell and conda packages the ages come straight from Packagist, hex.pm, pub.dev, jsr.io, CRAN's own per-version timeline (via METACRAN's crandb), Hackage's upload-time endpoint and anaconda.org's per-release upload times, which deps.dev doesn't cover at all; pods get theirs from the CocoaPods trunk API's per-version timestamps; Terraform/OpenTofu providers from the registries' per-version publish times (so a provider release cut hours ago gets its ⏱ flag too); Go tags cut minutes ago get theirs from the module proxy's.infoendpoint, and pseudo-versions carry their commit time in the version string itself, so those age for free; Maven artifacts fall back to the POM'sLast-Modifiedupload time on Central or Google's Maven repository. The proxy's latestgo.modalso supplies retractions (with the author's rationale comment) and// Deprecated:module notices —GOPROXYis honoured, so private proxies work too. - Each changed package's source repository (from deps.dev) has its tag list
fetched over git's smart-HTTP protocol — one anonymous GET per repo, the
same request
git ls-remotemakes. Old and new versions are matched against the real tags (tryingv1.2.3,1.2.3,pkg@1.2.3,pkg-v1.2.3, Godir/v1.2.3, and pseudo-version commit hashes), and only verified matches become compare / release links — so every link works.
Privacy: the only network traffic is the OSV.dev and deps.dev batch
queries (package names + versions), anonymous npm-registry / PyPI /
crates.io / RubyGems / Packagist / NuGet / hex.pm / pub.dev / jsr.io /
CocoaPods-CDN-and-trunk / Go-module-proxy / Terraform-and-OpenTofu-registry /
Maven-Central-and-Google-Maven / ConanCenter / CRAN (crandb +
cran.r-project.org) / Hackage / Bazel-Central-Registry / anaconda.org
metadata fetches for
changed packages of those ecosystems, and the anonymous git tag listings
above.
-offline disables all of it; -no-vulns / -no-meta disable
vulnerability and metadata+links lookups individually. The typosquat check
makes no requests at all — the popularity lists ship inside the binary.
No telemetry, ever.
Air-gapped vulnerability checks (-osv-db)
-offline doesn't have to mean "no vulnerability data". Point lockvet at a
directory of local OSV databases — the per-ecosystem all.zip files
OSV.dev publishes — and
the vulnerability check runs entirely from disk:
Details that matter:
- Same answers. Range evaluation happens client-side with the same
walk lockvet already uses for GitHub Actions advisories; withdrawn
records are excluded, matching the API. On real repos the output is
byte-identical to the API path — including
MAL-…malware records, so the chalk/debug replay works with the cable unplugged. - Cheap refreshes. Downloads are conditional (ETag): an unchanged
database costs one 304.
LOCKVET_OSV_DBsets a default directory, so CI can cache it like any other build cache. - Hand-copied zips work. If your air-gapped side only accepts files
through a transfer process, copy the
all.zipfiles in by hand — lockvet indexes them locally on first use. - Sizes are honest. Most ecosystems are a few MB; npm is ~200 MB
because it includes every
MAL-…malware record — that's the point. Indexing npm takes ~15 s once per refresh; warm runs are seconds. - Missing databases fail loud. An ecosystem in the diff with no local database skips the vulnerability check with an instruction, never silently reports "no vulnerabilities".
Caching: registry and advisory answers are cached on disk for one hour
(~/.cache/lockvet, override with LOCKVET_CACHE_DIR), so repeat runs are
fast — running lockvet, then lockvet -md for the PR comment, doesn't ask
every registry twice — and stay inside anonymous rate limits. Forge data
(PR state, file contents) is never cached, so you always vet the live diff,
and negative registry answers — the evidence behind the ▲ unlisted flag —
are re-proven on every run, so a just-published version clears the flag
immediately. -no-cache bypasses it; -cache-ttl 15m tunes it
(-cache-ttl 0 disables). Nothing from your repository is ever written
to the cache.
Dependencies: none. Pure Go standard library.
How it compares
git diff on the lockfile | whatsdiff v2.6 | dependency-review-action | lockvet | |
|---|---|---|---|---|
| Lockfile formats | any (raw text) | 3 (composer, npm, pnpm) | GitHub dependency-graph ecosystems | 60 across 30+ ecosystems, + CycloneDX/SPDX SBOMs |
| Readable per-package summary | ✗ | ✓ | ✓ (job summary / PR comment) | ✓ |
| Vulnerabilities introduced / fixed by the change | ✗ | ✗ | introduced only (GitHub Advisory DB) | ✓ both (OSV.dev) |
| Release age + ⏱ cooldown flag on fresh versions | ✗ | ✗ | ✗ | ✓ (deps.dev) |
| Deprecation warnings | ✗ | ✗ | ✗ | ✓ (deps.dev) |
License-change flag (MIT → BUSL-1.1) | ✗ | ✗ | ✗ (SPDX allow-list gate instead) | ✓ (deps.dev) |
| OpenSSF Scorecard scores for changed dependencies | ✗ | ✗ | ✓ (warn threshold) | ✗ |
| Flag versions their own registry no longer lists (unpublished malware) | ✗ | ✗ | ✗ | ✓ (unlisted) |
| Flag new dependencies one edit from a popular name (typosquats) | ✗ | ✗ | ✗ | ✓ (≈ typosquat, offline, npm/PyPI/crates.io) |
| Flag bumps that suddenly add npm install scripts | ✗ | ✗ | ✗ | ✓ (⚙ scripts) |
| Flag young releases that silently drop provenance (sigstore / trusted publishing) | ✗ | ✗ | ✗ | ✓ (⛨ provenance) |
| Flag integrity changes on unchanged versions & private→public registry moves (dependency confusion) | ✗ | ✗ | ✗ | ✓ (‼ integrity / ⇄ resolution, offline) |
| Air-gapped vulnerability check from a local database | ✗ | ✗ | ✗ | ✓ (-osv-db, incl. MAL-… malware records) |
GitHub Actions uses: pins (SHA → release resolution, tag-attack detection) | ✗ | ✗ | GitHub Actions ecosystem via dependency graph | ✓ (format #31, all five forges + locally) |
| Verify SwiftPM pins against upstream tags (moved tag / deleted release) | ✗ | ✗ | ✗ | ✓ (‼ tag mismatch) |
Direct vs. transitive, with pull-in chain (via a › b) | ✗ | ✗ | direct/indirect label, no chain | ✓ |
| Vet a PR / MR / compare URL without cloning | ✗ | ✗ | GitHub PRs only (runs as the PR's workflow) | ✓ (GitHub + GitLab + Bitbucket + Gitea/Forgejo + Azure DevOps, self-hosted incl.) |
| Triage every open Dependabot/Renovate PR at once | ✗ | ✗ | ✗ | ✓ (lockvet queue <org>, on all five forges) |
| Diff two container images' SBOMs, with distro CVEs | ✗ | ✗ | ✗ | ✓ (lockvet diff old.cdx.json new.cdx.json) |
| Audit the current pins, not just a change | ✗ | ✗ | ✗ | ✓ (lockvet audit: advisories, unlisted, deprecated, fresh) |
| Vet a package before installing it | ✗ | ✗ | ✗ | ✓ (lockvet pkg npm:left-pad: advisories, age, typosquat — latest or any version) |
| CI gate | ✗ | per-package check exit codes | ✓ (fail-on-severity, license lists, deny-packages) | policy gate (-fail-on major|vuln|fresh|deprecated|unlisted|typosquat|scripts|provenance|integrity|registry|license) + GitHub Action |
| Acknowledge a finding without turning the gate off | ✗ | ✗ | ✓ (allow-ghsas, no expiry) | .lockvetignore (per-advisory / per-package rules with expiry dates; ignored findings stay visible) |
| Output formats | text | text, JSON, markdown | job summary + PR comment | text, JSON, markdown, SARIF (code scanning alerts) |
| See what changed upstream | ✗ | ✓ (fetches changelog text) | ✗ | ✓ (verified tag-to-tag diff links + -changelogs fetches release notes, transitives included) |
| MCP server (let AI assistants vet PRs) | ✗ | ✓ | ✗ | ✓ (lockvet mcp: vet URLs, local repos, files, pre-install packages, audits, whole queues) |
| Interactive TUI | ✗ | ✓ | ✗ | ✗ |
| Runs locally / outside CI | ✓ | ✓ | ✗ | ✓ (CLI, browser playground, any CI) |
| Runtime | — | PHP (binaries provided) | JS action on GitHub runners; free on public repos, private repos need GitHub Advanced Security | single static Go binary, zero deps |
whatsdiff is a fine tool if you live in composer/npm and want changelogs and a TUI. GitHub's dependency-review-action is a solid default if you're all-in on GitHub — it reads the dependency graph (so it also sees manifest-only changes lockfiles don't record) and surfaces OpenSSF Scorecard scores; but it only checks known advisories and licenses, only on GitHub PRs, and private repos need an Advanced Security license. lockvet's focus is different: should I trust this diff? — across whatever forge and language your repos are in, with registry-level tampering signals no advisory database has heard about yet, locally or in any CI.
"Isn't this just …?"
osv-scanner? Google's
osv-scanner is the reference known-vulnerability scanner: point it at a
tree or container image and it inventories every OSV advisory, with an
experimental fix mode that plans upgrades for you. lockvet asks a
different question — what does this change do, and does anything about it
smell wrong? Advisories are one column of the answer; release ages,
deprecations, license flips, versions the registry no longer lists,
typosquats, install-script and provenance transitions, and
integrity/resolution tampering are the rest — signals that can fire
before any advisory exists (see the case studies).
If you want tree-wide CVE inventory with remediation planning, use
osv-scanner — lockvet audit
overlaps and names the fix version on every advisory line, but it won't
edit your manifests for you. If you want to know whether to merge a
Dependabot PR, that's lockvet.
npq? npq audits an npm package
before npm install — age, install scripts, typosquats, provenance
regressions, known vulns — with an interactive prompt and package-manager
aliasing so it can be your daily npm. If you live in npm, it's a great
daily driver. lockvet pkg
runs the same kind of pre-install check across ~30 ecosystems (npm, PyPI,
crates.io, RubyGems, Go, Maven, NuGet, Composer, Hex, pub.dev, JSR,
CocoaPods, Terraform, CRAN, Helm, GitHub Actions, CircleCI orbs,
toolchains, …) — but it's a one-shot
report, not an install wrapper, and pre-install vetting is one subcommand
of a tool whose main job is diffs and CI gates.
Socket? Socket is a commercial platform that goes
further than lockvet in one important way: it statically analyzes package
code for malicious behavior and risky capabilities. lockvet never reads
package code — it reads lockfile bytes plus public registry metadata,
anonymously, with no account, no code upload, and no service dependency
beyond the public registries themselves. Different trust models: if you
can adopt a managed service and want code-level analysis, Socket sees
things lockvet can't; if you want a zero-account static binary you can run
on any forge, any CI, or air-gapped (-offline), that's lockvet.
Dependabot / Renovate? Updaters, not reviewers — they open the PRs,
lockvet is the reviewer on the other side of them
(lockvet queue <org> triages every open bot PR at once). Complementary
by design.
Non-goals
- Not a code analyzer — lockvet never downloads or inspects package contents, only lockfiles and registry metadata. Tools like Socket or a human review do that job.
- Not an updater — Dependabot/Renovate open the PRs; lockvet tells you whether to merge them.
- No interactive TUI (see whatsdiff above) — lockvet stays a one-shot command whose output drops straight into a PR comment.
License
MIT © Matteo Sung