Imported from Vulnetix/pix-ai-coding-assistant (AGENTS.md). Install upstream with npx skills add Vulnetix/pix-ai-coding-assistant. Copyright stays with the author.

Vulnetix AI Coding Agent Plugin — Product Requirements

Purpose

This plugin brings vulnerability intelligence into the AI coding agent development loop. Developers get automatic security scanning on commit/install, on-demand vulnerability analysis, exploit intelligence, and AI-driven remediation — all without leaving their editor.

System Context

graph LR
    subgraph "Developer Machine"
        CC["Claude Code"]
        PLG["pix-ai-coding-assistant<br/>(this repo)"]
        CLI["Vulnetix CLI<br/>(../cli)"]
        CC --> PLG
        PLG --> CLI
    end

    subgraph "Vulnetix Cloud (AWS)"
        API["VDB API<br/>(../vdb-api)<br/>Vulnerability data"]
        SITE["VDB Site API<br/>(../vdb-site)<br/>Auth & accounts"]
        DB[(PostgreSQL 18<br/>VDB + Accounts)]
        API --> DB
        SITE --> DB
    end

    subgraph "Infrastructure"
        MGR["VDB Manager<br/>(../vdb-manager)<br/>Terraform IaC"]
        MGR -.->|deploys| API
        MGR -.->|deploys| SITE
        MGR -.->|provisions| DB
    end

    CLI -->|"GET /v1/vuln/*<br/>GET /v2/vuln/*"| API
    CLI -->|"POST /v1/cli/device/*<br/>GET /v1/account"| SITE

    subgraph "Third Party"
        GH["GitHub API<br/>Dependabot · CodeQL"]
        STRIPE["Stripe<br/>Subscriptions"]
    end

    PLG --> GH
    SITE --> STRIPE

Repository Dependencies

Vulnetix CLI (../cli)

The plugin's runtime dependency. Every skill and command shells out to CLI subcommands.

The CLI handles authentication (API key via vulnetix auth login), request signing, and JSON response parsing.

VDB API (../vdb-api)

The vulnerability data backend. Aggregates 12+ sources (CVE.org, NVD, VulnCheck, CISA KEV, GitHub/OSV, EUVD) into enriched CVE 5.1 responses.

Key capabilities the plugin relies on:

• Vulnerability lookup with CVSS v2/v3/v4, EPSS, KEV enrichment
• Exploit records with source attribution (ExploitDB, Metasploit, Nuclei, etc.)
• Fix intelligence with registry versions, distribution patches, source fixes
• V2 remediation plans with workarounds, advisories, CWE guidance, verification steps
• Package vulnerability listings with affected version ranges
• SBOM/manifest scanning endpoints

API versions: V1 (original endpoints), V2 (discrete remediation, scanning, cloud locators, malware detection)

Authentication: CLI exchanges org credentials at VDB Site API for JWT, then uses API key (ApiKey {orgId}:{hexDigest}) for VDB API requests.

VDB Site (../vdb-site)

Account management backend. The plugin depends on it for:

• Authentication: POST /v1/cli/device/authorize + /v1/cli/device/token — RFC 8628 device grant; the user approves in a browser signed in to the Vulnetix identity provider, and the CLI receives an API key
• Account info: GET /v1/account — subscription tier, rate limits
• Usage tracking: GET /v1/usage — request counts, audit log
• Sign-up: browser only, at https://www.vulnetix.com/vdb-register (Community tier, free). There is no unauthenticated endpoint that mints credentials from an email.

Subscription tiers affecting plugin behavior:

Rate limit exhaustion causes CLI commands to fail with 429 errors. The session-summary hook could surface this.

VDB Manager (../vdb-manager)

Terraform IaC that provisions and deploys everything the plugin talks to:

• ECS Fargate — runs VDB API and Site API containers
• PostgreSQL 18 (RDS) — stores vulnerability data and accounts with read replicas
• Secrets Manager — JWT secrets, API keys, Stripe keys shared across services
• Stripe products — subscription tier definitions (prices, limits)
• Lambda functions — scheduled data processing (source ingestion, EPSS updates)
• S3 — artifact storage for SBOMs and scan results
• CI/CD — GitHub Actions with AWS OIDC for deployment

Changes to rate limits, tiers, or API infrastructure are made here, not in the API repos directly.

Plugin Feature Matrix

Skills — LLM-Guided Workflows

Hooks — Event-Driven Automation

graph LR
    subgraph "Session Lifecycle"
        SS["SessionStart"] -->|session-summary.sh| D1["Display vuln status"]
        UP["UserPromptSubmit"] -->|vuln-context-inject.sh| D2["Inject vuln context"]
        ST["Stop"] -->|stop-reminder.sh| D3["Remind open vulns"]
    end

    subgraph "Tool Lifecycle"
        PRE1["PreToolUse:Bash"] -->|pre-commit-scan.sh| D4["Scan packages<br/>before commit"]
        PRE2["PreToolUse:Edit|Write"] -->|manifest-edit-scan.sh| D5["Gate manifest<br/>edits with scan"]
        POST["PostToolUse:Bash"] -->|post-install-scan.sh| D6["SBOM after<br/>npm/pip/go install"]
    end

Commands — Deterministic CLI Wrappers

Commands bypass the LLM (disable-model-invocation: true). They run a CLI command, parse JSON, and display structured output. Use these for scripting, quick lookups, or when you want raw data without AI interpretation.

Agents — Autonomous Multi-Turn

The bulk-triage agent runs /vulnetix:exploits in parallel across multiple vulnerabilities, then produces a consolidated triage report sorted by CWSS priority. It coordinates memory writes to avoid race conditions.

Data Architecture

graph TD
    subgraph "Repo Root"
        VD[".vulnetix/"]
        VD --> MY["memory.yaml<br/>Vulnerability state"]
        VD --> SC["scans/<br/>CycloneDX SBOMs"]
        VD --> PC["pocs/<br/>Cached exploit source"]
    end

    subgraph "memory.yaml Sections"
        MY --> MAN["manifests:<br/>Tracked files, scan times"]
        MY --> VUL["vulnerabilities:<br/>Status, decisions, CWSS,<br/>threat models, PoCs"]
    end

    SK["Skills"] -->|read/write| MY
    HK["Hooks"] -->|read/write| MY
    HK -->|generate| SC
    SK -->|cache| PC

Development Guidelines

• Skill implementations are SKILL.md files with YAML frontmatter (name, description, user-invocable, allowed-tools, model)
• Command implementations are markdown files in commands/ with disable-model-invocation: true
• Hook scripts are bash, registered in hooks/hooks.json, must exit 0 for informational output or non-zero to block
• Agent definitions are markdown files in agents/ with frontmatter (name, description, effort, maxTurns, allowed-tools, model)
• Website is Hugo (Go templates) at website/ — every registered feature must have a corresponding doc page under website/content/docs/
• The canonical schema for .vulnetix/memory.yaml is defined in skills/fix/references/memory-yaml-schema.md
• All skills share memory — coordinate writes carefully, especially in the bulk-triage agent

Skill quality gate

Every skill under vulnetix/skills/ must have:

1. A ## Use when body section in SKILL.md.
2. The phrase "Use when" in its frontmatter description.
3. evals/evals.json — schema per Anthropic schemas.md; ≥3 evals; each with ≥5 verifiable expectations (named CLI calls, output fields, or side-effects — never vague).
4. evals/trigger-eval.json — 10 should_trigger + 10 should_not_trigger queries, per Sogl, "Skills Without Evals Are Just Markdown and Hope".

SKILL.md structure is enforced locally by pre-commit (see .pre-commit-config.yaml) and on PRs by .github/workflows/skill-validator.yml — both run agent-ecosystem/skill-validator with --strict --allow-dirs=evals.