Files
Vyntral 3a4c230aa7 feat: v2.0 full rewrite — event-driven pipeline, AI + Nuclei + proxy
Complete architectural overhaul. Replaces the v0.1 monolithic scanner
with an event-driven pipeline of auto-registered modules.

Foundation (internal/):
- eventbus: typed pub/sub, 20 event types, race-safe, drop counter
- module: registry with phase-based selection
- store: thread-safe host store with per-host locks + deep-copy reads
- pipeline: coordinator with phase barriers + panic recovery
- config: 5 scan profiles + 3 AI tiers + YAML loader + auto-discovery

Modules (26 auto-registered across 6 phases):
- Discovery: passive (26 sources), bruteforce, recursive, AXFR, GitHub
  dorks, CT streaming, permutation, reverse DNS, vhost, ASN, supply
  chain (npm + PyPI)
- Enrichment: HTTP probe + tech fingerprint + TLS appliance ID, ports
- Analysis: security checks, takeover (110+ sigs), cloud, JavaScript,
  GraphQL, JWT, headers (OWASP), HTTP smuggling, AI cascade, Nuclei
- Reporting: TXT/JSON/CSV writer + AI scan brief

AI layer (internal/ai/ + internal/modules/ai/):
- Three profiles: lean (16 GB), balanced (32 GB MoE), heavy (64 GB)
- Six event-driven handlers: CVE, JS file, HTTP response, secret
  filter, multi-agent vuln enrichment, anomaly + executive report
- Content-hash cache dedups Ollama calls across hosts
- Auto-pull of missing models via /api/pull with streaming progress
- End-of-scan AI SCAN BRIEF in terminal with top chains + next actions

Nuclei compat layer (internal/nucleitpl/):
- Executes ~13k community templates (HTTP subset)
- Auto-download of nuclei-templates ZIP to ~/.god-eye/nuclei-templates
- Scope filter rejects off-host templates (eliminates OSINT FPs)

Operations:
- Interactive wizard (internal/wizard/) — zero-flag launch
- LivePrinter (internal/tui/) — colorized event stream
- Diff engine + scheduler (internal/diff, internal/scheduler) for
  continuous ASM monitoring with webhook alerts
- Proxy support (internal/proxyconf/): http / https / socks5 / socks5h
  + basic auth

Fixes #1 — native SOCKS5 / Tor compatibility via --proxy flag.

185 unit tests across 15 packages, all race-detector clean.
2026-04-18 16:48:41 +02:00

102 lines
3.0 KiB
Go

package config
// AIProfile bundles the triage + deep models for a named AI tier. Unlike
// the scan-level Profile (bugbounty/pentest/…), an AIProfile only touches
// model selection — it doesn't flip stealth, recursion, or module enables.
type AIProfile struct {
Name string
Description string
FastModel string
DeepModel string
// MinRAMGB is an advisory (not enforced) hint about the memory footprint
// of both models loaded simultaneously. Printed in the profile help
// banner so users can pick the right tier for their machine.
MinRAMGB int
}
// Built-in AI profiles. The lean tier matches the repository defaults so
// `--ai-profile lean` is always equivalent to "use whatever the defaults
// say". balanced and heavy upgrade deep model to Qwen3-Coder MoE which
// activates only 3.3B parameters per token despite its 30B total.
var (
AIProfileLean = AIProfile{
Name: "lean",
Description: "Runs on 16GB RAM; default. qwen3:1.7b triage + qwen2.5-coder:14b deep.",
FastModel: "qwen3:1.7b",
DeepModel: "qwen2.5-coder:14b",
MinRAMGB: 16,
}
AIProfileBalanced = AIProfile{
Name: "balanced",
Description: "32GB RAM / 24GB VRAM. Upgrades deep to qwen3-coder:30b MoE (3.3B active, 256K ctx).",
FastModel: "qwen3:4b",
DeepModel: "qwen3-coder:30b",
MinRAMGB: 32,
}
AIProfileHeavy = AIProfile{
Name: "heavy",
Description: "64GB+ RAM. Best-quality triage + deep. Slowest; ideal for final analysis passes.",
FastModel: "qwen3:8b",
DeepModel: "qwen3-coder:30b",
MinRAMGB: 64,
}
)
// BuiltinAIProfiles lists every AIProfile in CLI help order.
var BuiltinAIProfiles = []AIProfile{
AIProfileLean,
AIProfileBalanced,
AIProfileHeavy,
}
// AIProfileByName resolves a named profile. Lookup is case-insensitive
// and tolerates the common alias "max" → heavy.
func AIProfileByName(name string) (AIProfile, bool) {
switch normaliseAIProfileName(name) {
case "lean":
return AIProfileLean, true
case "balanced", "balance", "mid":
return AIProfileBalanced, true
case "heavy", "max", "power":
return AIProfileHeavy, true
}
return AIProfile{}, false
}
func normaliseAIProfileName(s string) string {
out := make([]byte, 0, len(s))
for i := 0; i < len(s); i++ {
c := s[i]
if c >= 'A' && c <= 'Z' {
c += 'a' - 'A'
}
if c == ' ' || c == '_' || c == '-' {
continue
}
out = append(out, c)
}
return string(out)
}
// ApplyAIProfile merges p's models into cfg. If cfg.AIFastModel /
// cfg.AIDeepModel were explicitly set by the user (overrideFast /
// overrideDeep true) the profile is ignored for that field. The caller
// is responsible for detecting explicit flags; in practice this comes
// from cobra's cmd.Flags().Changed("ai-fast-model").
func ApplyAIProfile(cfg *Config, p AIProfile, overrideFast, overrideDeep bool) {
if cfg == nil {
return
}
if !overrideFast && p.FastModel != "" {
cfg.AIFastModel = p.FastModel
}
if !overrideDeep && p.DeepModel != "" {
cfg.AIDeepModel = p.DeepModel
}
if cfg.AIProfile == "" {
cfg.AIProfile = p.Name
}
}