Skills

A skill is a procedure the agent follows in your repository. Each one is an Agent Skills SKILL.md under skills/, installed by vulnetix skills install.

A skill earns its place only if it needs the repository, needs you, or needs a sequence a model would not compose correctly alone. Fetching one record is none of those, so lookups are MCP tools and vulnetix vdb subcommands, not skills.

Invoking a skill

The agent picks a skill from its description, or you name it. In Claude Code with the plugin, skills are namespaced:

/vulnetix:repo-impact CVE-2021-44228
/vulnetix:fix CVE-2024-29041

Installed through npx skills or gh, the skill is /repo-impact (or however your agent lists skills). Every skill is user-invocable: true.

All skills

Skill	The question	Arguments	Usually followed by
dependency-choice	I need capability X: which package, and what am I trading?	<capability or package name> [--ecosystem npm|pypi|...]	typosquat-check, fix
typosquat-check	Is this name impersonating something, or known-malicious?	[<package-name>] | [--installed]	dependency-choice
dep-resolve	The bump is blocked by peer dependencies: find a compatible set	<package-name> [--target-version X]	dependency-choice, fix, verify-fix
eol-check	What here is past end of life, or will be within 90 days?	[--strict]	dep-resolve, fix
license-check	Does anything here conflict with our licence policy?	[--policy permissive|copyleft-aware|strict]
repo-impact	Does this CVE reach anything in this repository?	<CVE-...|GHSA-...>	fix, dependency-choice
fix	Apply a fix for this finding, here, safely	<vuln-id>	verify-fix, vex-publish
verify-fix	Did the fix land?	<vuln-id>	vex-publish
dashboard	What have we already decided here?	none	repo-impact
sast-scan	Code-level vulnerabilities in changed files	[--rule-id ID] [--paths file1 file2] [--baseline]	secure-code-write, verify-fix
secret-scan	Hardcoded credentials	[--paths file1 file2] [--staged-only]	verify-fix
iac-scan	Terraform, OpenTofu, Nix and Kubernetes misconfigurations	[--paths file1 file2]	verify-fix
container-scan	Dockerfile, Containerfile and compose risks; image CVEs with Trivy or Grype	[--paths Dockerfile1 Dockerfile2] [--image registry/img:tag]	verify-fix
secure-code-write	I am about to write auth, crypto or SQL: what should I know?	[<file-path>] | [<topic: auth|crypto|sql|xss|deser>]	sast-scan
detection-rules	I cannot patch this: what do I deploy instead?	<vuln-id>	exploit-test, verify-fix
exploit-test	Prove this is or is not exploitable against my authorised target	<vuln-id> [--target URL]	verify-fix
sbom-generate	Produce the SBOM a customer or auditor asked for	[--format cyclonedx|spdx|both] [--sign] [--output PATH]	vex-publish
vex-publish	Publish our triage decisions as VEX	[--format openvex|cyclonedx] [--upload]

“Usually followed by” is the skill’s metadata.chain. It is advice to the agent: nothing runs the next skill automatically.

Tool grants

Each skill’s allowed-tools lists only the commands its body runs. No skill is granted bare Bash, and no read-only skill is granted Write or Edit.

Grant	Skills
Bash(vulnetix:*) Read Grep Glob only	dashboard, dependency-choice, eol-check, iac-scan, license-check, repo-impact, secret-scan, secure-code-write, typosquat-check
+ Edit	fix
+ Edit and package managers (npm, yarn, pnpm, pip, cargo, go)	dep-resolve
+ Write	detection-rules
+ Bash(git:*)	verify-fix
+ Bash(semgrep:*)	sast-scan
+ Bash(nuclei:*)	exploit-test
+ Bash(syft:*) Bash(grype:*) Bash(trivy:*)	container-scan
+ Write Bash(syft:*) Bash(cosign:*)	sbom-generate
+ Write Bash(cosign:*) Bash(gh:*)	vex-publish

Each skill also names a model: haiku for dashboard, sonnet for the rest. A host that does not read the key ignores it.

Rules every skill follows

These come from the shared contract, skills/_lib/contract.md:

1. MCP first, CLI otherwise. If vulnetix_* tools are present, use them for data. Anything that touches the working tree uses the CLI even when MCP is there.
2. Install the CLI if it is missing, with Homebrew, Scoop, Nix or the install script, then vulnetix agent install.
3. Find credentials, never ask for a key. The CLI’s own lookup order applies. With none, say once that the session is on the shared Community tier.
4. vulnetix auth login needs you. The skill shows the URL and code and waits; it cannot approve the device grant.
5. Read .vulnetix/capabilities.yaml first where the skill depends on local tools, and skip what is not installed rather than fail. Refresh it with vulnetix agent capabilities.
6. Write state to .vulnetix/, mostly memory.yaml. See Repository files.

Checking the skills yourself

From a clone of the repository:

Command	Does
just validate	skill-validator check --strict against the Agent Skills spec
just evals	Checks the eval corpus (skills/*/evals/); no model, no credential
just evals-full claude	Drives real agents through the corpus. Costs money.