Architecture¶
The platform is built on a skill-based architecture with micro-agent parallel execution.
System Architecture¶
graph TB
subgraph "Input & Initialization"
CLI["CLI Commands<br/>/pentest, /intake, /context"]
Brief["Brief.json<br/>Business Context"]
end
subgraph "Skill-Based Execution"
Skills["31 Sub-Agent Skills<br/>10 Test Domains"]
Router["Smart Router<br/>Endpoint → Scope Mapping"]
end
subgraph "Parallel Execution"
L1["L1: Docker Parallel<br/>Tools: subfinder, nuclei, katana"]
L2["L2: Claude -p Parallel<br/>AI Reasoning: 3 agents/wave"]
end
subgraph "Primary Engine — Claude"
Opus["Opus 4.6<br/>Creative/Critical Skills"]
Sonnet["Sonnet 4.6<br/>Passive Analysis"]
Haiku["Haiku 4.5<br/>Tool Execution"]
end
subgraph "Advisory Engine — Codex"
Codex["GPT-5.4 xhigh<br/>Cross-model verification"]
end
subgraph "Output & Reporting"
Findings["Finding Database<br/>FINDING-NNN.md"]
Report["Professional Report<br/>CVSS 4.0, HWG Format"]
end
CLI --> Brief
Brief --> Skills
Skills --> Router
Router --> L1
Router --> L2
L1 --> Opus
L2 --> Opus
L2 --> Sonnet
L2 --> Haiku
Opus --> Findings
Sonnet --> Findings
Findings -->|dual-verify| Codex
Codex -->|verdicts| Findings
Findings --> Report
style CLI fill:#4a148c,color:#fff
style Brief fill:#6a1b9a,color:#fff
style Skills fill:#7b1fa2,color:#fff
style Router fill:#8e24aa,color:#fff
style L1 fill:#9c27b0,color:#fff
style L2 fill:#ab47bc,color:#fff
style Opus fill:#0277bd,color:#fff
style Sonnet fill:#00838f,color:#fff
style Haiku fill:#00695c,color:#fff
style Codex fill:#01579b,color:#fff,stroke:#00e5ff,stroke-width:2px
style Findings fill:#004d40,color:#fff
style Report fill:#4a148c,color:#fff
AI Agent Codebase¶
The platform's core is 236 files / 99,661 lines of AI agent instructions — 46% of the entire project:
- 70 SKILL.md files (54,999 lines) — agent decision logic, test methodology, wave orchestration
- 140 knowledge packs & helpers (41,583 lines) — 25+ vulnerability classes with techniques, payloads, bypass patterns
- 1 CLAUDE.md (233 lines) — master orchestration: 6 phases, 14 safety rules, completeness gates
- 25 schemas, hooks, configs (2,846 lines) — Codex structured output, doc sync, automation
Key Components¶
- Skill System — All functionality lives in
.claude/skills/ - Micro-Agent Architecture — L1 Docker parallel + L2
claude -pparallel - Model Routing (V3 Pragmatica) — 3-tier model assignment with per-scope compound keys
- Agent Decomposition — 3 levels of parallelism (scope, sub-scope, endpoint split)
- Wave Coordinator — 12-wave dispatch with health checks
- Dual-Engine: Claude + Codex — Cross-model verification with multi-agent parallel dispatch
- Context Engineering — Research terms, exemplars, knowledge packs
- Token Optimization — Context fork, file piping, result isolation