Generate Cve JSON

generate-cve-json

This skill produces a CVE 5.x JSON document from a tracking issue in <tracker>, ready to paste into the Vulnogram "#source" tab of the ASF CVE tool. The goal is to eliminate the manual "copy each field from the issue into the right Vulnogram form input" step when you are preparing to publish an advisory.

Project-agnostic by design. All project-specific values (vendor, top-level product / package name, project display map, CNA org id, generator tag, …) are loaded from a TOML config the adopting project ships at <project-config>/tools/cve-tool-vulnogram/cve-json-config.toml. Concrete apache-foo-project-* strings appearing in this document are illustrative examples of how a project with a project-style package layout would configure things; replace them mentally with the adopter's own package taxonomy. The schema is documented in the package README.

Golden rule: the script generates a proposal JSON document. It parses a handful of structured fields from the issue body, but it cannot read the security team member's mind. Always review the generated JSON before pasting, and always do the final review inside Vulnogram before moving the CVE from DRAFT → REVIEW → READY → PUBLIC.

Release-vote gating (opt-in, recommended for ASF projects). The emitted CNA_private.state follows a tri-state state machine:

• DRAFT — the CNA is incomplete or the project has opted into release-vote gating and no vote is in progress yet.
• REVIEW — the CNA is review-ready (CVE ID + title + description + affected versions + CWE + severity + ≥ 1 credit + ≥ 1 reference) and either the project hasn't opted into gating (legacy: ready ⇒ REVIEW) or an RC vote is in progress (signalled by the configured tracker label or a --review CLI flag).
• PUBLIC — the CNA is review-ready and the public advisory has shipped (a vendor-advisory reference is present).

Projects opt into gating by setting [workflow].release_vote_gating = true in their cve-json-config.toml and choosing the label name via [workflow].rc_voting_label (default "rc voting"). The sync skill is responsible for detecting [VOTE] threads on the project's dev list (e.g. dev@<project>.apache.org) and proposing the label add/remove; the generator only reads the label on the tracker. Non- ASF adopters who publish advisories without a separate release-vote step typically leave gating off — the legacy "ready ⇒ REVIEW" behaviour is the right default for that workflow.

Determinism: the same input issue body produces exactly the same JSON bytes on every run. The script uses only the Python standard library, has no timestamps or machine-dependent values in its output, sorts JSON keys, and sorts references alphabetically. This lets you paste the result into Vulnogram, tweak fields in the tool, re-run the script later, and cleanly diff the two to see what the tool has added / what you changed by hand.

Inputs

• Issue number (required) — e.g. 232.
• Optional CLI overrides:

  • --cve-id CVE-YYYY-NNNN+ — override the CVE ID if the issue body's CVE tool link field has not yet been filled in, or retarget the JSON to a different CVE ID.

  • --title "<vendor>: <product>: …" — override the CVE title. Default is the GitHub issue title with the project's <vendor>: <product>: prefix (sourced from the TOML config) when it does not already start with that phrase.

  • --version-start X.Y.Z — override the start of the affected version range (the affected[].versions[].version field). Default is the lower bound parsed from the Affected versions field when it uses >= X, < Y syntax, otherwise "0".

  • --remediation-developer "Name" — append a type: "remediation developer" credit on top of whatever the body's Remediation developer field already lists (auto-populated by the security-issue-sync skill from the linked PR's author). Repeat the flag to add multiple developers; duplicates between the body field and CLI flags are dropped silently. The reporter credit(s) from the Reporter credited as field are always emitted with type: "finder".

  • --vendor / --product / --package-name / --collection-url — override the product identity fields. The defaults come from the project's TOML config (product.vendor, product.default_product, product.default_package_name, product.default_collection_url). They are used as the identity for Affected versions lines that don't start with a recognisable per-package directory name (see the multi-product note below). --collection-url reaches the record only through the purl it is used to derive (see product.purl_type below) — except for a product that opts out of purls, the one case a record still carries a collectionURL.

  • Helm charts — purl_type = "helm", and you must know where the chart is published. A chart is identified by its repository, not by its name: two projects may both ship a chart called superset. So the generator emits

    pkg:helm/<chart>?repository_url=<base URL>

    taking the base URL from product.default_collection_url (or the package's collection_url override), and refuses to emit a purl at all without one rather than producing a chart identifier that matches somebody else's chart.

    If the base URL is not recorded anywhere, ask whoever publishes the chart. It is whatever a user would put in helm repo add — the https://… site serving index.yaml, or the oci://… registry path for a chart published as an OCI artifact. Either is accepted verbatim; a trailing slash is dropped so the same repository does not produce two identifiers. Do not infer it from the project's homepage: charts are routinely served from a different host than the project site.

    Note that helm is not one of the types the purl specification registers. It is used because the scanners that would match this advisory to a deployed chart emit it, and because the two spec-correct encodings — pkg:oci/ for an OCI registry and pkg:generic/ for a classic repository — would give the same chart two different identities depending on how it happened to be published. A strict validator may object to the type; being invisible to every scanner is the worse outcome.

  • product.purl_type (config only) — the Package URL type for the project's packages (pypi, npm, cargo, …). Every affected[] entry carries a packageURL (pkg:<type>/<packageName>). Per the CVE Record Format the purl never includes a version — the entry's versions[] carries the range.

    The purl is the only package identifier the record carries. The ASF CVE tool treats the Package URL as the recommended identifier and derives the legacy collectionURL / packageName pair from it when the record is serialised for publication, so the generator emits the purl alone. Writing the pair as well would hand the tool two identifiers that can disagree — which it reports rather than silently reconciles.

    A purl is required, not optional. A record needs one to be promoted, so the generator refuses to emit a record it cannot build a purl for rather than producing one the CNA will reject later. Leaving the key unset is fine when the type can be derived from product.default_collection_url — PyPI, npm, crates.io, RubyGems and NuGet hosts are recognised — and only hosts whose type is unambiguous are mapped, because a guessed type yields a valid-looking identifier pointing at the wrong ecosystem.

    Set purl_type = "none" to state deliberately that a product has no package host — a source-only release published to dist.apache.org and nowhere else, for instance. That is an explicit decision rather than an omission, which is the distinction the previous optional behaviour lost. Those entries — and only those — carry collectionURL / packageName instead, because there is no purl for the CVE tool to derive them from. Name normalisation follows the package-url spec per type, and is implemented for the types whose rules have been read from it: pypi (lowercased, _ becomes -) and npm (scope becomes the namespace, so @angular/animation renders as pkg:npm/%40angular/animation; names keep their case, since mixed-case npm packages were grandfathered in). Any other type has its name passed through unchanged, and a name needing namespace semantics is covered by product.purl_namespace below, and without one such a type is an error rather than a guessed purl.

  • product.purl_namespace (optional, config only) — the namespace for purl types that require one: a Maven groupId, a Composer vendor, a Go module prefix. Types the spec gives no namespace (pypi, cargo, gem, nuget) ignore it, and an npm scope carried in the package name itself wins over it. A type that requires a namespace and has none is an error naming this key — a guessed Maven groupId would point at another organisation's artifact, and silently omitting the purl produces an unpromotable record.

  • [packages.overrides."<packageName>"] (config only) — per-package distribution identity, for a project that ships to more than one ecosystem. Accepts product, collection_url and purl_type; anything omitted falls through to the product.* default, so a package that differs only in purl type need not restate its collection URL.

    The case this exists for: a project whose packages are on PyPI and which also ships, say, a Helm chart from its own site. product.* describes only the majority ecosystem, so without an override the odd-one-out inherits it and the record claims pkg:pypi/<chart> — a package that host does not carry. Since purls are what scanners match on, a wrong one is acted upon, unlike the wrong free-text product name it replaced.

    [packages.overrides."apache-example-helm-chart"]
    product = "Apache Example Helm Chart"
    collection_url = "https://example.apache.org/"
    purl_type = "none"

    The key is the resolved packageName, so the package must be one the configured package_pattern matches — extending that pattern is a prerequisite, not an extra. --product-for still wins over the product set here.

  • --product-for PACKAGE=PRODUCT — override the CVE product display name for a specific packageName. Repeat to override multiple packages. Useful when a package is not in the project's project_display_map config, or when an acronym needs different casing from the title-cased fallback. Example: --product-for apache-foo-project-baz='Apache Foo Project Baz'.

  • --org-id <uuid> — override the CNA assigner org id (defaults to the ASF org id).

  • --discovery <word> — override source.discovery (default "UNKNOWN"; valid CVE 5.x values include UNKNOWN, INTERNAL, EXTERNAL, USER).

  • --no-envelope — emit only the inner cna container instead of the full CVE 5.x record (envelope is the default).

  • --review / --draft (mutually exclusive) — force the emitted CNA_private.state to REVIEW or DRAFT regardless of the tracker's labels. Useful in two cases:

    • --review lets a release manager nudge a record forward by hand when the rc voting label is not yet set on the tracker.
    • --draft walks a record back when an RC vote was cancelled or failed and the label is still around. Both flags only matter when release-vote gating is enabled in the project's TOML config (see below); otherwise the state is derived from the CNA's readiness alone and these flags have no effect beyond what the legacy logic produces.

  • --attach — after generating the JSON, embed it at the end of the tracking issue's body (after the CVE tool link field), wrapped in a collapsible <details> block. The block is bracketed by HTML-comment markers (<!-- generate-cve-json: cve=CVE-YYYY-NNNN+ version=v1 --> … <!-- generate-cve-json:end cve=CVE-YYYY-NNNN+ version=v1 -->) that the script uses on later runs to find the existing block and replace it in place, so re-runs update the embedded attachment instead of duplicating it or breaking other body fields. The attachment lives in the body — not as a comment — so it stays above every status-change comment in the timeline (effectively "pinned" without needing any pin mechanism). Requires the positional issue argument; incompatible with --stdin.

Prerequisites on the tracking issue

For the generated JSON to be useful, the issue body should already be filled in through a prior security-issue-sync run. In particular:

• Short public summary for publish — becomes the CVE description.

• Affected versions — becomes the CVE affected[] list. The script understands the common version-expression shapes (< 3.2.2, >= 2.0.0, < 3.2.2, <= 3.2.1, a bare version like 3.1.5, and a bare lower bound like >= 2.0.0). Multi-product CVEs are supported — put one package per line, prefixing each with the package directory name as it appears in the adopter's repo, and the script emits one affected[] entry per line with the right product and package identity. Example (illustrative — using a hypothetical apache-foo project's sub-project layout):

  apache-foo-project-alpha <=6.5.0
  apache-foo-project-beta <=1.9.0

  Known package directory names are resolved to the vendor-preferred display casing via the project's packages.project_display_map config table; unknown packages fall back to title-cased dash-split and can be overridden with --product-for. A line without a package prefix (or a single-line field) falls back to the --product / --package-name defaults, which preserves the single-product behaviour.

  < NEXT VERSION placeholder — multi-package trackers don't know which package version will ship the fix until the wave's release manager picks it during a release cut. Until then, the Affected versions lines use the literal token NEXT VERSION as the upper bound, e.g.:

  apache-foo-project-alpha < NEXT VERSION
  apache-foo-project-beta < NEXT VERSION

  The generator strips < NEXT VERSION before parsing each line and emits a versions[] entry without lessThan (open-ended upper bound — "affected from <low> onwards, no fix released yet"). When the wave ships and the version is known, the security-issue-sync skill replaces each NEXT VERSION with the actual < X.Y.Z and the next regen produces a fully-bounded entry. Case-insensitive; combines with a lower bound (e.g. >= 2.0.0, < NEXT VERSION becomes {version: "2.0.0", status: "affected"}).

• Security mailing list thread — internal navigation reference only; the script does not export URLs from this field into references[]. Keep whatever the reporter or triager put there.

• Public advisory URL — each URL in this field is extracted and added to references[] with tags: ["vendor-advisory"]. Populated by the release manager (or the security-issue-sync skill) once the advisory is archived on <users-list>. The --advisory-url CLI flag still exists for ad-hoc overrides.

• PR with the fix — each URL in this field becomes a reference URL. Multiple URLs are supported: paste them on separate lines, as a bullet list, or comma-separated — the script extracts every https?://… token it finds.

• Reporter credited as — each line becomes one CVE credit entry with type: "finder". Multiple credits are supported: put each person on their own line. Full Name, Affiliation on a single line is treated as one credit, not two, so the common Jed Cunningham, Astronomer pattern works as expected. Bullets (- , * , 1. ) are stripped. Blank lines are ignored. If you need to credit many people::

  Jed Cunningham
  Saurabh Banawar
  selen (Huntr bounty 3e88d364-5047-4768-a52c-6568f21ef35b)
  

• Remediation developer — each line becomes one CVE credit entry with type: "remediation developer". Same parsing rules as Reporter credited as (newline-separated, Full Name, Affiliation is one credit, bullets stripped). Auto-populated by the security-issue-sync skill from the linked PR's author the first time PR with the fix is set; manual edits survive subsequent syncs (the skill only proposes appending names that aren't already there). The --remediation-developer CLI flag adds further names on top of whatever the body already lists.

Bot / AI credit policy. This generator is intentionally neutral on credit content: whatever a tracker's Reporter credited as or Remediation developer field carries is what lands in credits[]. The filtering of obvious bot / AI accounts (e.g. dependabot[bot], *-scanner, automated-*) happens upstream in the skills at extraction time — see bot-credits-policy.md for the detection rule and the per-skill enforcement sites. Keeping the filter upstream means an intentional human override (typed directly into the field) survives every JSON regeneration without needing a special bypass flag here.

• CWE — CWE-285: Improper Authorization style works; so does a bare CWE-285 or a plain sentence. The script extracts the CWE-\d+ token for the cweId field and uses the rest as the human-readable description.
• Severity — None, Low, Medium, High, Critical (case-insensitive) are emitted as the text content of a metrics[].other block. Vulnogram lets you replace this with a CVSS vector in its form if you want a numeric score.
• CVE tool link — the ASF CVE tool URL, e.g. https://cveprocess.apache.org/cve5/CVE-2026-40913. The script extracts the CVE-YYYY-NNNN+ token from this field. If the field is still _No response_, pass the CVE ID with --cve-id.

If one of these fields is missing, the JSON still generates, but the reviewer will need to fill the gap in Vulnogram. The skill surfaces any empty field in the proposal so nothing is silently skipped.