* fix(release): record dependency advisories without blocking releases Release dependency evidence now blocks only on known malware. Vulnerability advisories of every severity are recorded in the evidence summary and surfaced as GitHub warning annotations, and CI dispatched by Full Release Validation or release publication reports a failing production audit as a warning. The per-release risk-acceptance table existed only to accept advisory blockers and is removed. The release skills also record that main CI health never gates a release and that every failed test gets an explicit real-blocker-or-flake decision. * fix(ci): keep release audit relaxation within the workflow size budget ci.yml sits at the 480000-byte guard, so the release-dispatch check moves into a trusted harness script that security-fast already checks out.
7.9 KiB
| summary | read_when | title | |||
|---|---|---|---|---|---|
| How OpenClaw reviews dependency changes and packages plugin runtime dependencies |
|
Dependency locking |
Read this page as a dependency reviewer or a release engineer: it covers reviewing dependency changes for supply-chain risk, and validating root and plugin npm packages before publishing.
OpenClaw uses pnpm-lock.yaml as its committed product dependency review boundary. It records the resolved dependency graph used by source checkouts and CI, so transitive changes remain visible in code review.
OpenClaw does not commit npm-format locks for product packages or publish them in package tarballs. npm 12 removed shrinkwrap support, including the npm shrinkwrap command and loading npm-shrinkwrap.json from package roots or dependency tarballs.
The trusted ClawHub and Vercel release toolchains are separate exceptions: .github/release/clawhub-cli/package-lock.json and .github/release/vercel-cli/package-lock.json are committed npm project locks used by release automation. Neither is shipped in an OpenClaw package.
These projects do not inherit root pnpm overrides. The Vercel project uses approved, version-scoped npm overrides for vulnerable upstream pins; remove each override when its owning dependency accepts a fixed version. Lock audits cover resolved package identities, not code already bundled into upstream CLI artifacts.
Check dependency advisories
The production audit pre-commit hook and ordinary CI's security-fast job remain zero-install, npm-only checks of the product production graph. They query npm bulk advisory data, not upstream repository advisories. A passing result is limited to that source and graph; it does not establish that dependencies are unaffected by all known vulnerabilities. CI runs dispatched by Full Release Validation or release publication record a failing audit as a warning instead of failing: advisories never block a release.
pnpm deps:vuln:gate, used by release dependency evidence, audits the target's product pnpm lock plus each release-tool lock whose package is present in that target. It checks npm advisory data and adds published security advisories from verified public GitHub repositories. Repository mappings come from the manifests for exact locked npm package versions, not a package's latest manifest. The gate verifies that each repository is public before requesting advisories with the explicit state=published filter. It does not scan private repositories or unpublished advisories.
Known malware is the only blocking class: a malware advisory anywhere in any audited lockfile fails the gate and stops publication, because it marks a compromised package rather than a vulnerability report. Vulnerability advisories of every severity, in every graph, are recorded without blocking: the gate prints each one as a GitHub Actions warning annotation, and the release dependency evidence summary lists them under "Non-blocking advisory findings". A newly published advisory never delays a release; fix it with a normal dependency bump on main after publication. GitHub's medium severity maps to moderate in these reports. Upstream findings match the npm package identity and affected-version range against exact locked versions; only matches absent from the npm result for the corresponding lockfile and graph are added. Reports retain the source lockfile, so a release-tool finding does not imply product runtime exposure. Missing or invalid locks for declared dependency graphs fail the gate.
Release automation reuses its existing standard GH_TOKEN only for GitHub API requests, never for npm registry requests. Local runs without that token use anonymous GitHub requests and their rate limits. No new OpenClaw configuration or operator credential setup is required.
Interpret coverage
Release artifacts and the GitHub Actions step summary show npm coverage as checked, upstream coverage as checked or partial, mapped package-version counts, checked repository counts, coverage issues, and upstream-only findings. Findings identify their source as npm-bulk or github-repository; upstream findings also record matched locked versions. Even checked coverage is limited to these advisory sources, not comprehensive vulnerability clearance.
The upstream scan has fixed bounds: 2,500 exact package versions, 4,000 HTTP requests, and five minutes per run. It uses four concurrent requests, up to five pages of 100 advisories per repository, and at most 10,000 advisories per run. Each response is limited to 2 MiB and each request to 15 seconds. It does not retry or reuse stale cached results.
Matching findings from each repository page reserve request capacity across all workers and are checked against GitHub's reviewed advisory ranges before that task continues discovery. This prevents later repository scans from consuming the budget needed to reconcile already-discovered findings. Missing reviewed evidence preserves the publisher's finding and records partial coverage.
Missing or unsupported repository metadata, malformed affected-version ranges, exhausted request or pagination budgets, and request or rate-limit failures produce partial upstream coverage, not an unaffected result. Confirmed findings remain in the report under the same policy. Inspect the coverage issue subjects and reasons before interpreting a zero-finding result.
Published package behavior
Published OpenClaw plugin packages bundle their runtime dependency files in the tarball by default. Those bytes ship with the plugin and work the same way regardless of whether the operator uses npm, pnpm, or Bun.
Native-heavy plugins opt out of runtime dependency bundling because their dependency trees contain platform-specific or large native artifacts. Those plugins resolve dependencies at install time from exact-pinned direct dependencies. The root openclaw package also resolves dependencies at install time and does not bundle its full dependency tree.
The bundled Anthropic plugin communicates directly with the separately installed claude executable. It does not depend on or copy the Claude Agent SDK into OpenClaw's package. The external ACPX plugin independently declares an ACP adapter that depends on the SDK; ACPX leaves those dependencies to installation from npm instead of bundling them into its published package.
Neither path publishes a lockfile:
- root and plugin tarballs contain neither
npm-shrinkwrap.jsonnorpackage-lock.json; pnpm-lock.yamlremains the reviewed source dependency graph;- npm package locks exist only transiently while OpenClaw validates package graphs or runs
npm cito assemble a bundled plugin.
Validate npm dependency graphs
The npm-lock checker generates package-lock.json in a temporary directory, applies workspace overrides, and rejects any generated registry version absent from pnpm-lock.yaml. It does not write a lockfile into the checkout.
# Root and every publishable package
pnpm deps:npm-lock:check
# Only packages affected by the current changeset
pnpm deps:npm-lock:check:changed
Inspect a plugin tarball
npm pack @openclaw/discord@<version> --json --pack-destination /tmp/openclaw-plugin-pack
tar -tf /tmp/openclaw-plugin-pack/openclaw-discord-<version>.tgz | grep '^package/node_modules/'
tar -tf /tmp/openclaw-plugin-pack/openclaw-discord-<version>.tgz | grep -E '^package/(npm-shrinkwrap|package-lock)\.json$' && exit 1 || true
The node_modules entries prove that the plugin carries its bundled runtime payload. The final check proves that neither npm lockfile format ships in the tarball.
Related
- Release performance sweep - the May 2026 package-size, dependency, and shrinkwrap audit this policy came out of