vector-workflow
Reference doc for the row-seeding logic inside scripts/offensive.py's derive_rows(), called by
offensive.py board. You do not run this skill as a step in the loop - it explains what already
happens automatically each time board writes the Approach.md 4a matrix, the same relationship a
"how the fingerprint router works" doc has to the router itself.
Wiki
Hub: [[methodology-moc]] - Primary page: [[multi-agent-campaign-orchestration]] REFS: [[multi-agent-campaign-orchestration]]
The pipeline, in order
derive_rows(eng, index, etype) builds the desired 4a rows in this fixed sequence:
- OSINT pre-pass (once, target-level). Before any host/IP exists,
VECTOR_CLASSES["osint"]=osint-subdomain,osint-leaks,subdomain-takeoverare emitted keyed to the engagement name itself (Path(eng).namestands in for "asset" since no asset row exists yet at this phase).subdomain-takeoverlives here (not thewebvector) because it's a subdomain-enumeration-time check against every discovered subdomain, not gated on a per-asset web fingerprint. - Per asset, once
state.mdhas assets (from rustscan+nmap etc.), for each asset row: a. Fingerprint-implied classes first. Every routing-table fingerprint whose\b<fp>\bregex matches the asset'shaytext (built fromtech/services/service/os/notescolumns) contributes its mappedclass, in routing-table order. These rank above vector/base rows. b. Vector baseline, inVECTOR_PRIORITYorder (["web", "ad_windows", "linux"]): for each vector whoseVECTOR_INDICATOR[vector]regex matcheshay, emit every class inVECTOR_CLASSES[vector], then checkVECTOR_EXCEPTIONfor that vector and emit the extra class if its fingerprint substring also appears inhay. c.BASE_CLASSES[etype]fallback fills in any base class not already in the implied list (rows alreadyseenfrom steps a/b are deduped regardless).
An asset can match more than one vector - a host running both SMB and a web app gets both the
ad_windows and web baseline rows. G4 (Deadends.md) suppression applies at every step: a
(asset, class) pair already deadended is never re-emitted. next/done/G1-G9 treat every
vector-seeded row exactly like any other 4a row - no new gate, no new command.
Vector indicators and baseline classes
VECTOR_PRIORITY = ["web", "ad_windows", "linux"]
VECTOR_INDICATOR = {
"web": r"\b(https?|nginx|apache|iis|tomcat)\b",
"ad_windows": r"\b(smb|ldap|kerberos|winrm|rdp|445|389|88|3389|5985)\b",
"linux": r"\b(ssh|22|linux|nfs|rsync)\b",
}
VECTOR_CLASSES = {
"osint": ["osint-subdomain", "osint-leaks", "subdomain-takeover"],
"web": ["content-discovery", "js-extract", "recon-nuclei", "recon-nikto"],
"ad_windows": ["ad", "windows"],
"linux": ["linux-svc-enum"],
}
VECTOR_EXCEPTION = {
("web", "wordpress"): "wpscan-scan", # only fires when "wordpress" is also in hay
}
The pseudo-classes and their real hunt-skill mapping
Every class above resolves through the routing table (scripts/routing.json)
to a real skill/wiki/arsenal, same as any fingerprint-implied class - _class_info() reads it
straight from the compiled index, no special-casing for vector rows. Current mapping (grep
scripts/routing.json for the -baseline rows if this drifts):
| class | hunt skill | primary wiki | arsenal slug | tool (CLASS_TOOL) |
|---|---|---|---|---|
| osint-subdomain | hunt-secrets | web-attack-surface | recon | subfinder |
| osint-leaks | hunt-secrets | secret-hunting | recon-dorks | trufflehog |
| subdomain-takeover | hunt-secrets | subdomain-takeover | recon | nuclei |
| content-discovery | (none - G2-skip baseline row) | web-attack-surface | wordlists | ffuf |
| js-extract | hunt-secrets | javascript-source-map-exploitation | recon-dorks | linkfinder |
| recon-nuclei | (none - G2-skip baseline row) | web-attack-surface | nuclei-arsenal | nuclei |
| recon-nikto | (none - G2-skip baseline row) | web-attack-surface | cve-arsenal | nikto |
| wpscan-scan | (none - G2-skip baseline row) | cms-exploitation | cms-exploitation | wpscan |
| ad | hunt-ad | active-directory | netexec | netexec |
| windows | hunt-windows | windows-privesc | privesc-exploit-arsenal | winpeas |
| linux-svc-enum | (none - G2-skip baseline row) | vuln-assessment | service-enumeration | searchsploit |
The five "(none)" rows match scripts/routing.json's *-baseline entries with
skill: "" - a legal state per routing.json's _doc (vector-baseline rows with no
doctrine match; the row still seeds the board with its wiki/arsenal/tool, it just
names no hunt skill). Content discovery on those rows is owned by Skill(fuzz)
(deterministic wordlist pick + WAF/throttle handling); these rows only seed the board.
ad and windows are real hunt-core classes reused as the ad_windows vector baseline, not
vector-only pseudo-classes - they also fire directly off ad/windows fingerprint matches.
content-discovery/js-extract/recon-nuclei/recon-nikto/wpscan-scan/osint-subdomain/
osint-leaks/subdomain-takeover/linux-svc-enum are the nine pseudo-classes that exist only to
give the vector pipeline (and OSINT pre-pass) a routable row.
When this matters
You don't invoke this skill manually - a fresh engagement's board command runs derive_rows()
automatically and the resulting OSINT + vector rows just appear on the Approach.md 4a matrix
already populated with arsenal/skill/tool. Read this skill when:
- You need to explain why a particular row is on the board (which vector matched, or which fingerprint implied it).
- You're extending
offensive.pyconceptually (not editing it here) and need the exact seeding order without re-deriving it fromderive_rows(). - A vector row's skill/arsenal looks wrong and you need the real mapping to check against
scripts/routing.jsonrather than guessing.
Full loop mechanics (gates, next/note/done, close-out) live in Skill(offensive); this skill
is scoped to the board-seeding pipeline only.