dep-audit

Audit project dependencies for known vulnerabilities (CVEs). Supports npm, pip, Cargo, and Go. Zero API keys required. Safe-by-default: report-only mode, fix commands require confirmation.

Dependency Audit Skill

Detect and report known vulnerabilities in your project's dependency tree. Supports npm, pip (Python), Cargo (Rust), and Go out of the box. No API keys. No config. Just point it at a project.

Activation

This skill activates when the user mentions:

• "audit", "vulnerability", "CVE", "dependency check", "supply chain", "security scan"
• Checking dependencies, lockfiles, or packages for issues
• Generating an SBOM (Software Bill of Materials)

Example Prompts

1. "Audit this project for vulnerabilities"
2. "Check all my repos in ~/projects for known CVEs"
3. "Are there any critical vulnerabilities I should fix right now?"
4. "Generate an SBOM for this project"
5. "What dependencies need updating in this project?"
6. "Audit only the Python dependencies"

Permissions

yaml

permissions:
  exec: true          # Required to run audit CLIs
  read: true          # Read lockfiles
  write: on-request   # SBOM generation writes sbom.cdx.json when user asks
  network: true       # Tools fetch advisory DBs

Agent Workflow

Follow this sequence exactly:

Step 1: Detect

Run the detection script to discover lockfiles and available tools:

bash

bash <skill_dir>/scripts/detect.sh <target_directory>

If no target directory is given, use the current working directory (.).

Parse the JSON output. Note which ecosystems have lockfiles and which tools are available.

Step 2: Audit Each Ecosystem

For each ecosystem detected in Step 1:

• If the audit tool is available, run the corresponding script:

  bash

  bash <skill_dir>/scripts/audit-npm.sh <directory>
  bash <skill_dir>/scripts/audit-pip.sh <directory>
  bash <skill_dir>/scripts/audit-cargo.sh <directory>
  bash <skill_dir>/scripts/audit-go.sh <directory>
  

• If the tool is missing, tell the user which tool is needed and the install command from the detect output. Skip that ecosystem and continue with others.

Note: yarn.lock and pnpm-lock.yaml are detected as yarn and pnpm ecosystems respectively. Audit support is npm-only in v0.1.x (package-lock.json). If only a yarn.lock or pnpm-lock.yaml is present, inform the user that dedicated yarn/pnpm audit is not yet supported and suggest running yarn audit or pnpm audit manually.

Each script outputs normalized JSON to stdout.

Step 3: Aggregate

Pipe or pass all per-ecosystem JSON results to the aggregator:

bash

bash <skill_dir>/scripts/aggregate.sh <npm_result.json> <pip_result.json> ... 1>unified.json 2>report.md

The aggregator outputs unified JSON to stdout and a Markdown report to stderr. Capture both: 2>report.md for the Markdown, 1>unified.json for the JSON.

Step 4: Present Results

Show the user the Markdown report from the aggregator. Highlight:

• Total vulnerability count by severity
• Critical and High findings first (these need attention)
• Which ecosystems were scanned vs skipped

If zero vulnerabilities found: report "✅ No known vulnerabilities found." If no lockfiles found: report "No lockfiles found in

. This skill works with npm, pip, Cargo, and Go projects."

Discord v2 Delivery Mode (OpenClaw v2026.2.14+)

When the user is in a Discord channel:

• Send a short first response with totals and only Critical/High findings.
• Keep the first message under ~1200 characters and avoid large Markdown tables up front.
• If Discord components are available, include quick actions:
  • Show Full Report
  • Show Fix Commands
  • Generate SBOM
• If components are unavailable, provide the same options as a numbered list.
• Send long details in short chunks (<=15 lines) to improve readability.

Step 5: Fix Suggestions (only if user asks)

If the user asks to fix vulnerabilities:

1. List every fix command with the package name, current version, and target version.
2. Suggest creating a branch first: git checkout -b dep-audit-fixes
3. Ask for explicit confirmation before running ANY fix command.
4. Never batch-run fix commands silently.

Example interaction:

text

I found these fix commands:
  1. cd /home/user/project && npm audit fix
  2. pip install requests>=2.31.0

I recommend creating a branch first:
  git checkout -b dep-audit-fixes

Shall I run them? (yes/no)

Step 6: SBOM (only if user asks)

bash

bash <skill_dir>/scripts/sbom.sh <directory>

Report the file location and component count.

Error Handling

Situation	Behavior
Tool not found	Print which tool is missing + install command. Continue with available tools.
Audit tool fails	Capture stderr, report "audit failed for [ecosystem]: [error]". Continue with others.
Timeout (>30s per tool)	When timeout/gtimeout is available, report "audit timed out for [ecosystem], skipping". Continue.
Invalid target directory	Report "directory not found or not accessible" and stop that ecosystem scan (do not report false "clean").
No lockfiles found	Report "No lockfiles found" + list supported ecosystems.
jq not available	Detection works without jq. Audit and aggregation require jq — install it first.
Malformed lockfile	Report parse error for that ecosystem. Continue with others.

Aggregation Robustness

• aggregate.sh now tolerates mixed inputs (valid results + error objects).
• Invalid input objects are listed under errors in unified JSON and rendered in a "Skipped / Error Inputs" Markdown section.
• If no valid ecosystem results are provided, aggregate output sets status: "error" instead of crashing.

Safety

• Default mode is report-only. The skill never modifies files unless you explicitly ask for a fix and confirm.
• Audit tools read lockfiles — they do not execute project code.
• Fix commands (npm audit fix, pip install --upgrade) are printed as suggestions. The agent will ask for confirmation before running them.
• This skill checks known advisory databases (OSV, GitHub Advisory DB, RustSec). It does not detect zero-days or runtime vulnerabilities.
• No data is sent to third-party services beyond what the native audit tools do (they query public advisory databases).
• No telemetry. No tracking. No phone-home.