Dependency audit
Check the dependencies a project actually declares against OSV. Start with a
local package-lock.json, npm-shrinkwrap.json, pnpm-lock.yaml or Cargo.lock.
The default reads that file through native admission, reuses sbom-generate for
inventory interpretation, and looks up exact registry package versions. It
returns advisory summaries, affected coordinates, source identity and coverage.
It never installs dependencies or executes the project's code.
Use this before release review or vuln-triage. Use sbom-generate when the job
is an inventory or CycloneDX document; use vuln-triage to decide exposure,
priority and remediation from findings already available. A database match is
evidence about a package version, not proof that an application is exploitable.
Composes
sbom-generate#inventory-filesbom-generate#inventory-sourcesbom-generate#makeweb-fetch#web-fetch
Choose the evidence you already have
For an ordinary local audit, supply path relative to the authorized workspace.
repo_root selects that workspace root; lockfile_type is inferred from a
conventional filename unless explicitly selected. scope defaults to
all-locked. direct-production is available where the lockfile can establish
it; npm lockfile version 1 and Cargo do not record it, so that selection is
refused rather than guessed. These are locked declarations, not an inventory of
what is currently installed.
from-inventory accepts the completed runx/sbom-generate.inventory.v1 packet.
It keeps source identity, locations and coverage without reading or parsing the
lockfile again. from-source accepts complete acquired text and uses the same
inventory owner. from-url fetches a public npm JSON lockfile through web-fetch;
its parsed inventory has canonical JSON identity, while the separate acquisition
record retains the actual response-body digest and final URL. A mutable URL is
an observation of those bytes, not an immutable repository revision.
Only eligible exact npm and crates.io registry coordinates are queried. Local,
Git, unknown-source and unresolved components remain in the retained inventory
and omission counts. Registry identity does not establish that a package name is
public. Before running, apply the project's disclosure policy with exclude, a
list such as npm:@internal/payments. OSV receives only selected ecosystem,
package name and version; source files, paths and credentials are not sent.
Read the result and continue the work
The useful output is scan_result (runx/cve-audit.audit.v1): readable Markdown,
actual findings, changed or withdrawn observations and explicit coverage. The
normal bounded scan queries up to one provider batch and reads up to one advisory
batch. Larger work returns partial and an exact continuation artifact. Invoke
resume with that descriptor until no work remains. Keep the most recent cursor;
the host owns scheduling, retry timing and cursor selection. Reusing an older
cursor deliberately forks the scan and can repeat reads.
The per-turn query, record and byte budgets are operational bounds, not a total dependency or advisory cap. Native artifacts hold the inventory, query plan and immutable result parts. The cursor holds only progress and the current bounded batch, never the accumulated report, completed-id history or a dependency location archive. Query records carry a bounded preview; full components remain in a separate native record artifact. A provider page with a continuation but no records remains pending; stop and respect provider timing instead of spinning. A resource or transport failure is not an empty result. Retained artifact and receipt references identify the failed boundary.
findings_scope says whether the inline findings cover all observed results or
only the current part. result_part locates that part; results reads it and
returns its previous_part reference. Follow those references to consume earlier
parts without provider calls. Workflows can triage each bounded part, preserving
overall coverage and deduplicating coordinate/advisory identities if provider
pages repeat observations. A final part with no findings does not erase earlier
findings. Counts describe observations, not distinct global advisory identities.
Find every affected location
Each finding carries dependency_locations: its component count, a small preview,
a complete flag and the exact components_artifact. The preview contains at most
three complete short locations; it never substitutes shortened text for a path.
Large location sets and long paths stay in the retained component records.
Call locations with that descriptor and the finding's exact_query, then continue
with next_offset until eof. A page with no matches and eof: false is still
pending. max_bytes controls one native page and can be increased for a large
record within native byte ceilings; it does not cap the total location count.
The root scan also exposes the component artifact for inspecting a queried
coordinate with no advisory matches. This read makes no provider call.
Assurance and recovery
Queries and advisory reads use the operator-selected OSV connector. Runx owns
network transport, scoped authority, provider evidence and receipt sealing.
Package code joins exact query coordinates to captured records. It checks ids,
modified times at their shared provider precision and withdrawal state; changed evidence is reported as
needs_refresh and needs a new audit of the affected coordinates. No second
live scan is called independent verification.
Artifact digests establish byte integrity and project containment, not the authenticity of caller-supplied earlier provider claims. Verify the original receipt tree when that assurance is required. Inside a composed workflow the provider steps and final parent receipt are the evidence; do not invent a sealed parent receipt before the workflow finishes.
No active matches means only that OSV returned none for the completed selected coordinates. Missing inventory coverage, exclusions, pending work or changed records prevent a clean-audit conclusion. Source and provider observation times remain distinct. Missing artifact scopes, unavailable credentials, malformed provider records or unsupported source shapes stop at their owning boundary; they never widen authority, silently truncate results or create a Cloud operator.
This skill performs reads and local immutable evidence custody. It does not
repair a dependency, file an issue, publish an advisory or claim independent
proof of safety. Pass its findings and coverage to vuln-triage, then use the
owning development or publication capability for any requested action.