CTI Expert

Cyber threat intelligence and open-source intelligence skill. Turns Claude into a trained CTI/OSINT analyst. Generates precision search queries, interprets public data, builds case timelines, and delivers structured intelligence products — no API keys, no paid subscriptions.

Runs anywhere. Works in Claude Code (Desktop & CLI) and in OpenAI Codex / ChatGPT and other AGENTS.md-aware agents — see AGENTS.md for the cross-agent runtime contract. Throughout this file, $SKILL_DIR = the directory containing this SKILL.md (Claude Code: ~/.claude/skills/cti-expert; Codex/manual clone: the repo you are working in). Resolve it by locating SKILL.md — never hard-assume ~/.claude. Detect the OS once (Windows/macOS/Linux) and prefer uv for all Python — see §13 Tool Auto-Install Policy.

Collection method: agent-browser when available (JavaScript-heavy sites, infinite-scroll, screenshot evidence), with automatic fallback to web search / web fetch / direct URL fetch. Tool limitations are logged as collection gaps — never as case blockers.

---

1. Quick Start

# Full autonomous case — runs every applicable technique
/case target.com

# Guided flow for first-time investigators
/flow person

# Summary of what's been found so far
/brief

Append --yolo to any command to skip all interactive prompts and confirmations. The analyst makes every decision autonomously.

---

2. AEAD Case Lifecycle

Every investigation follows four phases:

Phase	What Happens
Acquire	Collect raw data — /sweep, /query, /username, /phone, /email-deep, /breach-deep, /subdomain, /webpivot + /icp (domain/URL targets), /dork-sweep · /docleak · /github-osint//secrets, /cn-corp · /iban · /hash-id on discovery
Enrich	Recursive pivot loop — the pivot orchestration engine treats every discovered identifier as a new seed and expands the graph hop-by-hop (/branch, /crossref, /link-subjects, /signatures) automatically until the frontier is exhausted, no approval prompts (autonomy=auto). Each discovered identifier auto-fires its leak/breach/OSINT/dork legs — email→/breach-deep+/intelx (breach dumps·infostealer logs·pastes·darknet), username→/username+socials, name→/dork-sweep+/docleak, apex→/intelx --phonebook+/secrets+/github-osint — see §"Leak / breach / infostealer auto-fire" + the Dork/GitHub auto-fire matrices. Acquire↔Enrich iterate, not run once.
Assess	Score and verify — /exposure, /threat-model, /validate, /coverage, /verify-finding. Judgments carry likelihood terms, coverage gets the 5W1H pass, attributions get an ACH matrix (handbook/analytic-standards.md). If the case has not converged (frontier still open after the pivot loop + deterministic pipeline) and posture is active, /case auto-escalates to the /harness deepening loop — keyless-first (the CLI's own model), egress hard-gated on hostile infra; --no-harness opts out
Deliver	Package output — /report, /brief, /render, /workspace save — first ASKS whether to import more evidence from manual investigation (merged into the report JSON before anything is built), always auto-saves the base data bundle (.md + .json + .csv + IOC bundle: .stix.json/.txt/.csv/.jsonl), then ASKS which presentation report to render — (a) PDF · (b) DOCX · (c) HTML · (d) all (both prompts skipped under --yolo/guided-auto, which default to HTML). When CHONGLUADAO_API_KEY is set, the IOC bundle also attaches CLD's STIX + MISP indicator feed as companion artifacts (cld_api.py feed stix2|misp --raw → loadable bundle, not merged into the case graph). Deep-layer persist (automatic, ZERO extra egress): when /backend is live, /case reuses the pivots it already collected — never re-fetches — to persist the versioned case at $SKILL_DIR/intel_engine/cases/<CASE-ID>/ and correlate it cross-case; see the auto-chain note below. See connectors/chongluadao-api.md

Run /progress at any point to see which phase you're in and what's pending.

/case and web-infra pivoting. For a domain or URL target, /case includes web-infrastructure pivoting (/webpivot) in the Acquire phase. It runs keyless by default (crt.sh + passive DNS + anonymous urlscan) and upgrades automatically when premium keys are set via /apikeys (Shodan/Censys/FOFA/Hunter.how/DNSLytics/SecurityTrails/urlscan-PRO/WhoisXML). With keys the pipeline also: reads the urlscan Pro hostname lifecycle (pre-registration NS/A eras on the timeline, verdict rows in Appendix B), runs the MO-neighbour pivot on the estate's non-CDN origin (co-tenants WHOIS-verified; only a registrant join-key ever seeds, same-MO personas render as a rung-10 Related personas table — --mask-personas to aggregate), measures entitlement (meta.capability.plans, per-case capability_plans.json; Censys' search is its own probe), and fires SecurityTrails DNS-history + DSL reverse-WHOIS, DNSLytics reverse-IP (co-tenancy-filtered), a once-per-case Censys cert search, Shodan cert/JARM search and IntelX (loop: --full only). Every metered leg is --free-only/no_spend-gated; IntelX selectors, WhoisXML/SecurityTrails reverse-WHOIS terms, MO-neighbour origins, DNSLytics reverse-IP and the Censys cert search are bought once per CASE (on-disk memo), while per-host legs (urlscan lifecycle, SecurityTrails subdomains/history, Shodan search) stay per host under per-case caps. Because /webpivot can fetch the target directly, for hostile infrastructure it prefers passive capture (urlscan/Wayback) — see techniques/web-pivot.md. It is not run for username/phone/person targets.

⭐ ChongLuaDao is the first-party premium upgrade. With CHONGLUADAO_API_KEY set (/apikeys set chongluadao <KEY>), Acquire/Enrich fold CLD's own datasets into /scam-check, /threat-check, /phone, /breach-deep, /email-deep, /vuln-check and /impersonate, and /cld <target> is the direct entry point. Your client connects only to ChongLuaDao, never to the target (provable from scripts/cld/cld_api.py); for URL/AI/IP checks CLD does any target fetch server-side — the safe first-touch verdict on a live scam funnel before a direct pivot. Full catalog + AEAD placement: connectors/chongluadao-api.md.

Archive IOC harvest runs by default too. For domain/URL targets the Acquire phase also runs wayback_harvest.py <domain> --indicators (add --urlscan when URLSCAN_API_KEY is set), harvesting emails, phones, crypto wallets, tracking/verification IDs, SaaS-operator IDs, and socials from the entire Wayback history — not just the live page — with first-seen/last-seen per selector. It writes case-schema indicators[] to <case>/raw/harvest.indicators.json, which merge into the case and flow into the auto-saved IOC bundle at Deliver. This is the step that recovers selectors a network later scrubbed — across the whole snapshot corpus, not just the live page. Passive by construction — only web.archive.org (+ urlscan.io if keyed), never the target.

The five v2.6 commands are in the pipeline too — no flags. /icp runs for every domain/URL/org target (and an IP's resolved hostname); /cn-corp, /iban and /hash-id fire the moment a company name/USCC, payment detail, or hash appears — and all three feed their yields back into the recursive pivot loop as new seeds, so an ICP licence serial or a reused bank account expands the graph like any other node. /redact is the exception: it is opt-in (--redact), because a redacted report is a weaker artifact and that should always be a deliberate choice. Full trigger table: §Technique Activation Matrix. Narrow with --no-cn.

Two layers, one skill: broad collector → deep pipeline. cti-expert is the broad collector — the wide net of Acquire/Enrich commands (/webpivot, /sweep, /subdomain, /icp, /username, /email-deep, /breach-deep, …) that pull artifacts from anywhere. The intel_engine engine is now vendored in-repo under intel_engine/ (intel_engine/harness/, intel_engine/tools/, intel_engine/WebPivot/, intel_engine/IntelGraph|IntelReport|BinaryPivot|IntelAnalysis/) and supplies the pipeline chains + deeper pivoting logic: a persistent knowledge base (intel_engine/knowledge/), versioned cases (cases/), cross-case correlation, calibrated assessment, and rendering.

The chain (automatic for /case, and it NEVER re-fetches): broad collection (cti-expert) already ran pivot_extract per host during Acquire. When /backend resolves to Tier 1/2 and the run produced ≥1 host seed (domain/URL/IP), /case hands what it already collected to the deterministic pipeline in reuse mode (--no-collect) — one command, a complete case dir, and nothing touches the target again:

1. Write each host's already-collected pivot JSON to $SKILL_DIR/intel_engine/cases/<CASE-ID>/raw/<host>.json and the host list to an absolute <CASE-ID>-seeds.txt, both anchored at the engine root (ROOT=$SKILL_DIR/intel_engine) — never a CWD-relative path.
2. intel.py pipeline open <CASE-ID> <abs-seeds.txt> --no-collect — --no-collect skips the live fetch and runs the rest of the existing pipeline over the raw you just wrote: ingest → prior-overlap (/recall) → risk (/risk) → shared cluster seeds → clusters.json → case_graph.json → ICD-203 assessment.md. Every step is a KB/local read: zero egress, zero metered calls.

The persisted case lands at $SKILL_DIR/intel_engine/cases/<CASE-ID>/ — raw/, shared.txt, clusters.json, case_graph.json, assessment.md — NOT the current working directory, and it is a COMPLETE case that intel.py pipeline status <CASE-ID> accepts (not the partial dir a hand-rolled op sequence would leave). <CASE-ID> is the same id as the report filenames (mint CASE-YYYYMMDD-NN when none is given). Its cluster/risk/assessment fold back into the auto-saved report. Tier 3 (stateless), or a person/username/phone target with no host seeds → the handoff is skipped silently and the tier noted; broad collection + report are unaffected.

Collecting pipeline open (no --no-collect) stays MANUAL — /case never runs it. intel.py pipeline open without the flag re-fetches every seed directly (collect_many on https://<host>; the egress gate at collect_core.py:169 only fires when hostile is set, which the open path never sets), so a blind auto-open would be a second live round against infra /case may have just classified hostile/no-touch — which is why the automatic handoff uses --no-collect. Run collecting mode by hand only for a fresh case with no prior collection, after setting the egress posture (/scope, /cti-proxy, or --passive).

/harness deepening AUTO-ESCALATES on non-convergence (--no-harness opts out). After the --no-collect pipeline, if the case has not converged — intel.py convergence <case> reports status ≠ converged (cold = no free leads left = exhausted), or intel.py frontier <case> still lists open leads and posture is active (not --passive, infra not classified hostile/no-touch) — /case runs the harness Collect→Correlate→Assess loop to close the gap. Egress stays safe by construction: the harness's own audit.py PreToolUse gate turns hostile=True into a hard denial of outbound collection (harness/README.md), so an escalation can never re-touch no-touch infra — on hostile infra it deepens correlation/assessment only. Model — the CLI's own agent: run interactively in Claude Code, the IntelHarness skill front-end drives the same pipeline on your subscription with no separate LLM key; it falls back to HARNESS_BACKEND=local (Ollama/vLLM/LM Studio, keyless) or an API key only for unattended SDK runs. No reasoning backend reachable and non-interactive → the escalation is skipped and noted as a collection gap, never a blocker. A converged case, --passive, or hostile-only infra → no escalation and the deterministic result stands.

Self-contained & self-resolving. /backend resolves to SELF (in-repo) — no external setup. The bundled installer (scripts/install.{sh,ps1}) provisions the deep layer; or by hand: uv venv && uv pip install -r requirements.txt (harness SDK/MCP + IntelGraph renderers; the collector + KB + deterministic pipeline are stdlib and need none). An explicit $INTEL_HOME still overrides for a shared external KB. Full architecture, the op map, and the evidence-envelope schema: connectors/intel-backend.md.

---

2.5. Pivot Priority & False-Positive Control (CRITICAL)

Two failure modes ruin a cluster: asserting a link that isn't there, and missing one that is. This section governs both. Apply it in Enrich, before anything reaches a report.

Pivot priority ladder

Work down this ladder. Never assert same-operator on a lower rung when a higher rung is available or contradicts it. Tag every asserted link in the report with the rung it rests on.

Rung	Indicator	Strength
1	Registrant email / phone / org — including historic WHOIS	decisive
2	One domain carrying two identities across its own WHOIS history	decisive — proves an alias
3	Site-verification token (Google Search Console, etc.)	decisive — proves account control
4	Shared TLS certificate / SAN cross-cover	strong
5	Nameserver delegation to a host the operator runs themselves	strong — proves zone control
6	APK signing certificate	strong
7	Distinctive favicon / analytics / tracker / backend tenant ID	moderate — verify below
8	Co-tenancy on a dedicated host (few tenants)	moderate
9	Site template / framework / kit	weak — kit-level, never operator-level
10	Co-tenancy on shared/reseller hosting; managed-provider nameservers	information, not a link

Reverse-WHOIS is the highest-yield pivot here. Always mode=preview first — the count is free. A term returning hundreds is shared boilerplate; do not purchase it.

Mandatory false-positive control

Before any indicator becomes a cluster edge, run /reference check <value>. If it returns UNKNOWN, decide and record it with /reference add so the next case inherits the judgement.

Six traps, all of which have produced real false clusters:

Trap	Why it fools you	Test
Commodity site kit	A template sold to hundreds of unrelated fraud operators	Search the template path in urlscan/FOFA — a large population means kit-level
Privacy-proxy contacts	The registrar's boilerplate phone/email, shared by every customer of that service	Reverse-WHOIS it; a spread of unrelated domains means noise
Shared/reseller hosting IP	A 20+-tenant cPanel box links nothing	Count tenants before clustering
Managed-provider nameservers	Cloudflare/GoDaddy/Gandi/Wix NS are shared by millions	Self-hosted NS is rung 5; provider NS is rung 10
Org-name collision	A registrant org string that also matches a real, unrelated company	Reverse-WHOIS the org; inspect what comes back before attributing
Shared analytics / tag container	Often one web developer reusing a container across unrelated clients	Check domain creation dates — a decade-old business sharing a tag with a new fraud domain is a third party

Never put an unvalidated indicator into a report that recommends abuse reporting. Naming an uninvolved business is the most damaging error this skill can produce. When a cluster rests on a single rung-7-or-below indicator, label it candidate, single-indicator — not a cluster member.

Never submit the case's own sample to a public sandbox (CRITICAL)

/anyrun is lookup-only: it reads detonations that already happened (anyrun_lookup). The engine does carry a detonation path — anyrun_submit (T1) / bp_anyrun.py submit — but it is gated four ways, and every gate is code, not convention:

1. Per-submission analyst confirmation. anyrun_submit without confirm=true returns the risk briefing and sends nothing; bp_anyrun.submit() refuses unless confirm=True is passed (a function parameter — references/anyrun.json can only make the policy stricter). Show the briefing, ask, and only on an explicit yes to this submission call again with confirm=true. Consent to "analyze this sample" is not consent to detonate it.
2. Private by default, public refused. Privacy defaults to owner; public is refused even with confirm unless allow_public is separately authorised — per call, or as the analyst's standing ANYRUN_ALLOW_PUBLIC=1 in the gitignored .env. With that set, a plan that cannot go private (gate 3 denied) is downgraded to a public task explicitly — the result carries public_task, downgraded_from, public_authorized_by — never silently, and never when the plan can go private. Auto-delete defaults to a week.
3. Free plan fails closed. Pre-flight, before any POST, the engine checks the key's own account record: /user → limits.private (observed live; 0 in any window = denied, and the analyst attestation below cannot override that positive evidence; -1/positive = entitled), else a prior non-public task in the account's history. Neither → refused + plan_evidence, unless the analyst explicitly attests to a paid plan (allow_unverified_plan=true / --i-have-a-paid-plan) — never set it on your own.
4. Post-submit read-back. The task record carries no privacy while the sandbox is running (observed live: a status: "in progress" stub for ~2 min), so after the POST the engine polls the report — bounded to sandbox timeout + 60 s, max 240 s — and, once finished, reads the privacy ANY.RUN actually applied. A forbidden mode is withdrawn (task deleted) and the result says exposed: true. If the task outlives the wait, the result says privacy_verified: null with a verify_command; run it — bp_anyrun.py verify-privacy <uuid> / anyrun_submit with verify_task=true, target=<uuid> — to finish the check and the withdrawal. Either way this is detected, not prevented: deletion does not un-publish what the feed already showed, so treat the infrastructure as tipped.

The harness adds a fifth: audit.gate() denies anyrun_submit outright unless the run was launched with HARNESS_ALLOW_SUBMIT=1. tests/test_no_sample_submission.py asserts the gate is both marked and enforced and that no upload machinery exists outside it; intel_engine/tools/eval/test_intelx_anyrun.py §7b–7c exercises every refusal, the plan proof, the attestation path and the read-back.

Do not work around any of it. Uploading the case's own APK / installer / archive to ANY.RUN — or VirusTotal, or any public sandbox — is an outbound, irreversible act:

• A public task is world-readable: the file, its hash, screenshots and full network log.
• Operators watch for their own samples. The standard response is to rotate the backend, revoke the signing key and re-skin the front — destroying the infrastructure the case is built on, often days before a takedown or referral can land.
• It cannot be recalled. Deleting a task does not un-publish what was already seen.
• A URL detonation fetches the live target from published sandbox egress: the operator learns it was sandboxed, and a "clean" verdict may be the decoy served to datacenter IPs.

Try first, and say what you tried: static analyze_artifact, then an existing detonation of the hash (anyrun_lookup, VirusTotal, MalwareBazaar, Triage, Koodous). Prefer the downloaded FILE over the live URL. Never put a case ID or an analyst/client name in tags or the filename. Never on standing permission inferred from an earlier approval, never as a side effect of a pivot. The same reasoning governs --submit (urlscan/Wayback): a public urlscan scan of a live scam funnel is visible to the operator too.

A permuted email is a hypothesis, never a finding (CRITICAL)

When a case yields a real person's name or a username, and you already hold a domain that matters to the case, run /email-permute. An operator's mailbox is almost never published, but it is usually derivable — mail hosts use a small set of local-part conventions, and the operator's own domain is the highest-yield thing to permute against.

That value comes with a matching hazard, so this rule is absolute:

• Permute against the case's own domains. Name × the operator's domain is a narrow, high-prior question. Name × gmail.com is volume with no prior behind it — --free exists, is capped, and should be a deliberate choice, not a reflex.
• Never ingest a candidate into the KB, cite one in a report, or contact one. A fabricated address that reaches kb_ingest becomes a shared indicator, and a shared indicator merges two operator clusters. A permutator wired straight into correlation does not enrich a case — it silently names an innocent party. This is the same failure RULE 5 exists to prevent.
• Candidates are not seeds. They never enter the spider-map frontier. Only an address in the tool's promote list — corroborated by independent evidence (Gravatar registration, breach corpus, a GitHub commit, a page/DOM hit, a dork) — may be treated as a real email seed, and that promotion is an analyst decision.
• Never validate over SMTP. RCPT TO probing connects to the target's mail server, which the egress posture exists to prevent on a hostile case; and a catch-all domain answers 250 for every address ever tried, so it manufactures confidence instead of measuring it. Use --verify, which gates on MX (RFC 7505 null MX included) and checks Gravatar — both keyless, neither touching the target.

State the status in the turn. "12 candidates, 0 corroborated" is an honest result; presenting those 12 as discovered addresses is not.

Dead seed? Do not stop

Zero pivots, a parked page, or NXDOMAIN is not an answer. Run /fallback <domain> — crt.sh, the full Wayback timeline, archive.today, and the local KB. A parked apex frequently has live subdomains: enumerate CT and the Wayback CDX host histogram before writing a seed off. Report an empty result as empty; a collector that returned nothing is a finding, not something to omit.

Egress control — proxy / rotation

On a hostile case your egress IP is a selector too — a direct fetch of scam infrastructure exposes your real address to the operator, and repeated lookups from one IP get you rate-limited or fingerprinted. The /cti-proxy layer routes every HTTP(S) request the collectors make (keyless crt.sh, Wayback/CDX, urlscan, the CLD connector, WHOIS, analytics reverses, /apikeys test) through a configured proxy — or a rotation pool with automatic failover — so collection egresses from an IP you choose, and successive calls can egress from different ones. It also tunnels the collector's raw-socket TLS cert probe (/cert-pivot leaf fingerprint) via CONNECT, failing closed rather than dialling direct.

Raw-socket TLS probes are handled too — nothing dials the target directly behind a proxy. The cert-SHA probes (wp_pssl.py, wp_recon.py) and JARM (jarm.py) all take their socket from cti_proxy.proxied_connection: under an HTTP pool it is CONNECT-tunnelled and fails closed (never a direct dial); under a SOCKS pool the in-process socket hook carries it (installed at import via wp_common, and jarm.py's own bootstrap) when PySocks is present in that interpreter — if it is not (e.g. the intel.py pipeline runs tools under $INTEL_PY), the hook is absent and these probes fail closed instead of leaking, so install pysocks there; with no proxy it dials direct. As a policy choice the /webpivot analyze path additionally skips JARM under an HTTP pool (ten tunnelled handshakes are slow) and runs it under SOCKS / no proxy — that gate honors the env/pool proxy, not just an explicit --proxy.

uv run scripts/proxy/proxy.py add http://user:pass@host:3128 --label res-1
uv run scripts/proxy/proxy.py add 1.2.3.4:8080        # bare host:port -> http://
uv run scripts/proxy/proxy.py rotation round-robin    # | random | sticky | off
uv run scripts/proxy/proxy.py test                    # confirm each proxy's egress IP
uv run scripts/proxy/proxy.py status                  # pool + policy + toggles
uv run scripts/proxy/proxy.py disable                 # back to the real IP

• It is opt-in and additive — with no proxy configured the skill runs exactly as before, from your real IP. A pool, once added, is enabled by default.
• No-leak default: with a pool set, a failed pool is not silently retried direct — turn that on deliberately with allow-direct on.
• Precedence: env CTI_PROXY / CTI_PROXIES (and standard HTTPS_PROXY) override the stored pool for a one-off session; the store lives in scripts/proxy/proxies.json (gitignored, chmod-600 — it may hold credentials).
• Coverage: the broad collectors (scripts/…) get full in-process rotation + failover; the deep pipeline (/backend, /pipeline, /harness) inherits the egress for every tool it spawns (one proxy per run). For an ad-hoc tool call or the MCP server, export first: eval "$(python3 scripts/proxy/proxy.py use)".
• Formats: add accepts a full URL, a bare host:port, a provider host:port:user:pass export, a user:pass@host:port authority, or a pasted http_proxy="…" line. HTTP/HTTPS get the full rotation + failover + no_proxy behavior; socks5:///socks5h:// auto-install PySocks on add but rotate only per run — no in-process failover, and no_proxy is not enforced (the global socket hook routes everything).
• Full reference: /cti-proxy (commands/cti-proxy.md).

---

3. Command Reference

How to read this table — check the marker before you announce a command.

Marker	Meaning	What you may say
T2: / T1: shown	Backed by a real CLI op and/or MCP tool.	Call it, then report what it returned.
[model]	No code behind it, and none is needed — it names a way for you to work (a summary style, a checklist, a KB read-back).	Do the thing. Never claim a tool ran.
[unimplemented]	The tradecraft is documented but nothing executes it yet.	Say so, then follow the linked technique by hand. Do NOT narrate it as a tool call.

A command with no marker and no T2:/T1: line has not been triaged yet — treat it as [unimplemented]. Announcing a tool call that cannot happen is the failure this table exists to prevent: the output looks identical to real collection and is not.

3.0 Entry point & registered commands

/cti <target> is the single entry to this skill. It routes any target type — domain, IP, email, username, phone, wallet, hash, APK — through recall → collect → cluster → assess. Plain English works identically ("analyze example.com and pivot the infrastructure"); the command form just removes ambiguity.

/case <target> is an alias of /cti <target> — the same full pipeline run; /cti is the canonical entry (and the only form that works from a cold prompt). Prefer /cti.

Nine commands are registered with Claude Code by bash scripts/register.sh and work from a cold prompt in any project:

Command	Does	Equivalent T2 op	Equivalent T1 tool
/cti <target>	entry point — routes by target type	(whole chain)	(whole chain)
/cti-recall <seed>	seen before? run first, always	recall	domain_verdict, which_cases
/cti-case <ID> <seeds>	full deterministic pipeline	pipeline open	(none — CLI only)
/cti-pivot <url|ip>	collect one target	pivot-extract	pivot_extract
/cti-cluster <domain>	correlate & expand	kb, cert-overlap	kb_cluster, cert_overlap
/cti-check <indicator>	false-positive control	reference check	reference_check, reference_add
/cti-report <ID>	render graph + PDF/DOCX	graph, report	render_diagram, render_report
/cti-status	backend / MCP / credits health	backend.py status	api_usage
/cti-proxy [op]	egress proxy / rotation pool for all outbound calls	(none — CLI only)	(none — CLI only)

Every other /command in §3 is a convention read from this file, not a registered command. Once the skill is loaded they are unambiguous instructions; typed at a cold prompt they do nothing. When in doubt use /cti and describe the goal.

Three layers, one operation. The same capability is reachable three ways and the names differ by layer — T0 uses kebab-case after a slash, T2 uses kebab-case ops, T1 uses snake_case tools. The table above is the canonical mapping; when you add a capability, add a row here in the same commit or the layers drift apart again.

Capabilities that are not registered commands still carry their layer mapping inline in the §3 tables. The engine's WebPivot/BinaryPivot collectors add these: /capabilities (T2 capabilities, T1 capability_check), /impersonate (T2 impersonate, T1 impersonation_hunt), /search-pivot (T2 search-pivot, T1 search_pivot), /censys (T2 censys, T1 censys), /intelx (T2 intelx, T1 intelx_search) and /anyrun (T2 anyrun, T1 anyrun_lookup).

---

Commands grouped by AEAD phase.

Acquire

Command	What It Does	Example
/case [target]	Full pipeline — runs every applicable technique (alias of /cti) T2: intel.py case <seed> (= pipeline)	/case example.com
/sweep [target]	Multi-vector recon on any target type T2: intel.py sweep <target> (= pipeline)	/sweep @username
/query [subject]	Builds 12–15 advanced search operator queries T2: intel.py query <indicator>	/query example.com
/username [handle]	Enumerate handle across 3000+ platforms T2: intel.py username <handle>. T1: username_enum — HYPOTHESES, not findings	/username johndoe
/phone [number]	Carrier, line type, reputation, public associations, infostealer exposure (Hudson Rock); VN scam-phone reports (ChongLuaDao) when keyed T2: intel.py phone +<E164>. T1: phone_osint — carrier/line-type NOT determined	/phone +84901234567
/email-deep [email]	Accounts, breach history, infrastructure; breach/exposure records (ChongLuaDao data-leaks) when keyed T2: intel.py email-deep <email>. T1: deep_profile — metered steps planned, not fired	/email-deep u@domain.com
/subdomain [domain]	CT logs, brute-force, passive enumeration; flags admin/sensitive subdomains (admin,adm,kef,ador,panel…) per handbook/admin-endpoint-indicators.md T2: intel.py subdomain <domain> (keyless certspotter + hackertarget + crt.sh; names any source that was down) · intel.py subenum <apex> (subfinder auto-keyed from .env, amass, assetfinder, findomain → cases/<id>/subenum/<apex>.json). T1: subdomain_enum — the case-persisting form; live names are queued for the next collection round by case_frontier	/subdomain example.com
/breach-deep [email]	Multi-source breach lookup with context — Hudson Rock, IntelX, ChongLuaDao data-leaks/exposure when keyed T2: intel.py breach-deep <email>. T1: deep_profile (mode=breach)	/breach-deep u@domain.com
/traffic [domain]	Traffic estimation, ranking, audience data T2: intel.py traffic <domain>. T1: traffic_rank — Tranco only; no paid-panel estimates	/traffic example.com
/visitors [domain]	Full visitor intelligence: tech, geo, sources, analytics T2: intel.py visitors <url>. T1: pivot_extract (trackers)	/visitors example.com
/techstack [domain]	Technology fingerprint (CMS, analytics, CDN, server) T2: intel.py techstack <url>. T1: pivot_extract (tech_fingerprint)	/techstack example.com
/competitors [domain]	Competitor & related site discovery [unimplemented]	/competitors example.com
/secrets [target]	Exposed credentials in repos and paste sites T2: intel.py secrets <target>. T1: github_osint (secrets=true) — code search needs auth, so queries are EMITTED	/secrets github.com/org
/github-osint [target]	GitHub user/org/repo recon: profiles, repos, code search, commits, forks. Deterministic committer-identity harvest built in — `wp_github.py <login	org
/cld [target] ⭐	ChongLuaDao first-party premium connector (scripts/cld/cld_api.py, needs CHONGLUADAO_API_KEY). Now also wired into the deterministic pipeline open: wp_cld.py runs per collected domain host in enrich_live — the denylist checkurl verdict + IoC-URL analyzer land in live_results["cld"] and ingest as reputation FACTS (cld_verdict/cld_denylisted/cld_reputation_score, never a cluster edge; an empty-evidence label is flagged, not adopted), and for .vn hosts CLD WHOIS is the PRIMARY source (WhoisXML has no .vn coverage; RDAP/port-43 fall back) so the Domain Summary registrar/registrant/dates fill in. Metered → gated by --free-only/no_spend; CLD fetches the target server-side (posture-safe). Auto-routes any indicator (url/domain/ip/hash/email/phone/asn/CVE/.onion; non-indicators are skipped, never a blind metered call) to CLD's own datasets: URL verdict vs a ~20M denylist, deep AI URL analysis (risk 1–10 + findings), IoC verdict+evidence, denylist/brand-lookalike search, data-leak/breach exposure + full data-leak module (machines, stolen/exposed creds, cookies, leaked-accounts, devices, full-export — async start→poll), CVE/KEV + actor feeds, STIX/MISP export. Your client connects only to CLD, never to the target; CLD fetches server-side. 30-min timeout ceiling (--timeout); 403/404 → skip, not fail. Subcmds: route|checkurl|analyze|denylist|checkphone|whois|burner|ioc|exposure|leaks|breaches|machines|stolen-credentials|exposed-credentials|cookies|leaked-accounts|devices|device-detail|device-credentials|full-export|brand-domains|vulns|actors|onion|feed. See connectors/chongluadao-api.md	/cld https://scam-site.top
/threat-check [target]	IP/domain/URL/hash threat intelligence — ChongLuaDao IoC verdict + evidence (registration, reputation, threat feeds/reports) when keyed T2: intel.py threat-check <indicator>. T1: threat_check	/threat-check 185.1.1.1
/scam-check [domain]	Phishing/scam/malicious domain check — upgraded by ChongLuaDao checkurl (20M-URL denylist verdict) + analyze (deep AI, risk 1–10); client talks only to CLD, which fetches the target server-side T2: intel.py scam-check <domain>. T1: threat_check (mode=scam)	/scam-check susp-site.xyz
/webpivot [url]	Web-infra pivoting — extract favicon mmh3 / GA-GTM-AdSense / wallet / SaaS-operator artifacts from a page's DOM → ranked pivot queries (Shodan/PublicWWW/urlscan/FOFA). Flags: --render, --crawl, --history (Wayback GA), --fetch (pull archived page content — WebFetch can't reach Wayback), --harvest (full-IOC harvest across whole archive history → emails/phones/wallets/IDs/socials), --whois, --graph (cluster), --rank (score same-operator relations), --cert (cert-fingerprint pivot), --suggest, --wallets, --paths. See techniques/web-pivot.md (reverse-lookup engines per artifact → handbook/pivot-services.md) T2: intel.py webpivot <url>. T1: pivot_extract	/webpivot https://scam-site.top
(automatic — no flag)	Four layers now run on every collection and need no command. Asset layer: fetches the page's own JS bundles and re-runs every extractor over the source — the fix for SPA/white-label kits where the shell HTML is empty; yields off-apex api_endpoint/websocket_endpoint (the backend survives a front-end re-skin), build_env:<KEY> tenant tokens, js_bundle_sha256, and via sourceMappingURL the operator's own dev_username/dev_project. SPA route table: reads the app's router literals — spa_route:admin, spa_route:funnel, and a spa_route_signature that survives a re-skin. Zero extra requests, routes are leads only and are never fetched. Well-known/policy files: a fixed standards list (never a wordlist, no path brute-forcing) → adstxt_publisher, apple_team_id, security_contact. JARM: TLS-stack fingerprint of the server. Suppress with --no-assets / --no-well-known; cap fetches with --assets-max N	(runs inside /cti-pivot)
/capabilities	Run this first, and again before reporting any "nothing found". Which optional API keys are configured, and for each absent one the evidence class that went unqueried plus the free path that substitutes. A keyless run extracts every artifact but cannot reverse most of them — so "no sibling domains" with no FOFA/urlscan key is a fact about the credentials, not about the operator. Every collection also records this in meta.capability; carry the limitation statement into the assessment and cap confidence accordingly. T2: capabilities · T1: capability_check	/capabilities
/impersonate [domain]	Hunt lookalike / typosquat domains of a seed — typosquat permutations (omission, insertion, adjacent-key, transposition, homoglyph, hyphenation, combosquat) + a curated scam-heavy TLD sweep + a crt.sh keyword hunt, then existence-checked by live DNS. Output separates confirmed registered lookalikes (each an impersonation:candidate — run /cti-pivot on it and compare) from an unregistered monitoring watchlist. FREE (crt.sh + DNS); --fofa / --urlscan add the metered sweeps. Never live-fetches the lookalike infra. Tune the TLDs/affixes per campaign in intel_engine/WebPivot/references/impersonation.json. T2: impersonate · T1: impersonation_hunt	/impersonate example.com
/search-pivot [indicator]	Multi-engine search-engine pivot — the general-web complement to FOFA/PublicWWW, which only see served HTML. Takes any indicator (domain, slogan, tracking ID, wallet, Telegram/Zalo handle) and emits ready-to-open, URL-encoded dork queries across Google/Yandex/DuckDuckGo/Bing/Brave. It does not scrape: fire the queries with WebSearch, or WebFetch the DuckDuckGo html URL, then feed new hosts back into /cti-pivot. FREE, no keys. T2: search-pivot · T1: search_pivot	/search-pivot "distinctive slogan"
/censys [mode] [value]	Censys Platform — the server-side view FOFA/urlscan don't give. cert <sha256> returns every hostname on that exact leaf certificate (near-decisive cross-brand same-operator evidence, and it works on a free plan); host <ip>, webproperty <host> also free-plan. query <kind> <value> builds the CenQL offline and keyless; budget reports the balance. ⚠️ 100 credits/MONTH per account, no rollover — a lookup is 1, a search 5, and running the emitted CenQL in the web UI costs the same 5. Prefer handing the analyst the query over spending a search. Needs CENSYS_PAT. T2: censys · T1: censys	/censys cert 1a2b3c…
/intelx [selector]	Intelligence X — search ONE strong selector across a corpus nothing else here indexes: breach dumps, infostealer logs, pastes, darknet mirrors, historical WHOIS. Takes an email / domain (*.apex wildcard ok) / URL / IP / phone / wallet / IBAN — never a brand or person name (soft terms are refused and still cost a unit; classify_selector() blocks them locally). --phonebook <domain> inventories every email, subdomain and URL under an apex — the highest-value call, PAID-only. Grading is not optional: a hit in a breach dump or stealer log is EXPOSURE, flagged NOT clusterable — two addresses in one combolist share victims, not an operator. Only whois / pastes / darknet hits may carry a same-operator edge. Keyless ≈ 50%: it still types the selector and hands you the intelx.io URL. T2: intelx · T1: intelx_search	/intelx registrant@example.com
/anyrun [indicator]	ANY.RUN TI Lookup — READ-ONLY. What samples carrying this indicator did when other people detonated them: contacted domains/IPs/URLs/ports, family label, Suricata context, public task links. Run it after /binary on the sample's sha256, backend host or ip:port. It is the only way to recover a packed sample's real endpoints — those exist only at runtime, so a thin string sweep plus a binary:protection finding is exactly the cue. A shared family is same-KIT, never attribution on its own. Keyless ≈ 50%: composes the query + UI link. ⚠️ This tool never submits a sample — see the box below. T2: anyrun · T1: anyrun_lookup	/anyrun <sha256>
/webamon [domain|url]	Webamon scan corpus — the seed's latest sandbox scan (server IPs, certs, resources, risk score) and its kit fingerprints: one query per fingerprint key (dom_band/dom/links/scripts/dom_structure) returns the other lures built from the same template, with per-key agreement as corroboration and a per-key prevalence guard. Plus reverse-IP hosting eras (servers index), NRD/removed-feed dates, and a count-only infostealer exposure. pivot_extract runs all of this automatically when WEBAMON_API_KEY is set. Same kit ≠ same operator: siblings are related_leads / add_fact — never a frontier seed, never an edge, never corroboration. scan <url> submits to Webamon's sandbox behind the same two-step gate as /anyrun (--confirm-submission / confirm=true, HARNESS_ALLOW_SUBMIT=1): the fetch is Webamon's egress (posture-safe), but the scan is PUBLIC and permanent in the corpus. Starter plan: /search only — campaigns/clusters need research_lab. T2: webamon · T1: webamon_scan (submit; the read-only layer rides pivot_extract)	/webamon scam-site.top · /webamon scan https://scam-site.top/login
/cert-pivot [domain]	Cert-fingerprint pivot — other hosts serving the same TLS cert + SAN siblings (keyless; Shodan/Censys with keys). T2: intel.py cert-pivot <domain>. T1: cert_pivot	/cert-pivot scam-site.top
/sensitive-paths [list]	Classify a Wayback/URL list for exposed paths (.git/.env/backups/configs) — severity + per-year timeline. Pure matching, no request reaches the target. T2: intel.py sensitive-paths --file <list>. T1: sensitive_paths	/sensitive-paths waymore_index.txt
/email-hygiene [email]	Grade an email domain 0–100 + A–F (disposable / MX / free / role). An RFC 7505 null MX (0 .) scores as undeliverable, not valid. T2: intel.py email-hygiene <email>. T1: email_hygiene	/email-hygiene admin@site.top
/vuln-check [query]	CVE/vulnerability lookup (CIRCL + NVD; ChongLuaDao CVE/KEV threat-feed when keyed) T2: `intel.py vuln-check CVE-…	--product <p>. **T1:** vuln_check`
/ransomware-check [org]	Check if org is a ransomware victim T2: intel.py ransomware-check <domain>. T1: threat_check (mode=scam)	/ransomware-check "Acme Corp"
/stealer-log [folder]	Triage an infostealer-log folder — stealer-family attribution, victim-vs-operator profiling, cross-log actor correlation, IOC extraction (raw passwords/cookies/autofill/history shown)	/stealer-log ./logs
/gdoc [url]	Extract metadata/owner from Google document T2: intel.py gdoc <url>. T1: doc_metadata	/gdoc https://docs.google.com/...
/msftrecon [domain]	M365/Azure tenant recon — tenant ID, federation, MDI, SharePoint T2: intel.py msftrecon <domain>. T1: msft_recon	/msftrecon example.com
/icp [domain|serial]	ICP filing (工信部备案) → registered PRC entity + licence number; reverse the licence serial to sibling domains under the same filing (same-operator, HIGH). See techniques/china-recon.md T2: intel.py icp <domain>. T1: cn_recon — MIIT is CAPTCHA-walled; gates are named	/icp scam-site.top
/cn-corp [name|USCC]	PRC corporate registry chain — GSXT (ground truth) → TianYanCha/QCC/Aiqicha → 信用中国 blacklist → UBO; officers, shareholders, subsidiaries, revoked-status flags T2: intel.py cn-corp --company "<name>". T1: cn_recon — GSXT/TianYanCha gated	/cn-corp 深圳市某某科技有限公司
/iban [value]	Validate + decompose a bank account as a selector — mod-97 checksum, country, BBAN split, bank code, jurisdiction-mismatch signals. See techniques/fiat-payment-osint.md T2: intel.py iban <IBAN>	/iban GB29NWBK60161331926819
/hash-id [hash]	Identify a hash's algorithm before lookup — separates file hashes from credential material (32 hex = MD5 or NTLM) so it routes to the right service T2: `intel.py hash-id <hash> [--context file	credential]. **T1:** hash_id**T2:**intel.py hash-id <hash> [--context file
/appliance-scan [domain|ip]	Fingerprint internet-facing edge/VPN appliances (Citrix/F5/Cisco/Ivanti/Forti/PAN/Exchange) + exposed services → CISA KEV/CVE mapping. Passive-first (Shodan InternetDB/Censys); feeds /vuln-check + /threat-model. See techniques/fx-edge-appliance-recon.md [unimplemented]	/appliance-scan vpn.example.com
/saas-map [domain]	Map SaaS tenancy + identity fabric — DNS-TXT tenancy tokens, non-Microsoft IdP fingerprint (Okta/Auth0/OneLogin/Ping/Keycloak/ADFS), unauth API/GraphQL/spec discovery. See techniques/fx-saas-identity-recon.md T2: intel.py saas-map <url>. T1: pivot_extract (saas_ids)	/saas-map example.com
/sharelink [url]	Extract sharer identity from share link T2: intel.py sharelink <url>. T1: sharelink_resolve — contacts the final host	/sharelink https://vm.tiktok.com/ABC
/binary [file|url]	Built-in. Static IOC extraction from a scam/fraud binary (sideloaded APK, desktop trading .exe/.dmg, bundled .jar) via the in-repo BinaryPivot/ — signing-cert SHA-256, package name/permissions, embedded C2/backend hosts, Firebase/S3 tenants, wallets, Telegram/WhatsApp handles. Output is WebPivot-shaped → clusters the app with web infra in the shared KB. See connectors/intel-backend.md §7	/binary ./trader.apk

<!-- dork-integration:phase-05 start -->

| /dork-sweep [target] [--telegram\|--docs\|--filetype\|--all] [--after DATE] [--clean] | Zero-auth dork sweep: Telegram ecosystem, 18 doc-hosts, filetype families; 4-tier fallback cascade T2: intel.py dork-sweep <target> | /dork-sweep example.com --filetype | | /docleak [target] [--platform list] [--severity high] | 18-platform document leak hunt with severity classification (CRITICAL/HIGH/MEDIUM/LOW) T2: intel.py docleak "<target>". T1: dork_builder — emits queries, never runs them | /docleak "Acme Corp" |

<!-- dork-integration:phase-05 end -->

| /dns-history [domain] | Historical DNS record changes (A, NS, MX) via passive DNS T2: intel.py dns-history <domain>. T1: wayback_ga | /dns-history example.com | | /cert-history [domain] | SSL/TLS certificate timeline from CT logs (crt.sh) T2: intel.py cert-history <domain>. T1: passive_ssl | /cert-history example.com | | /proton-check [email] | Proton Mail account creation date via PGP key [unimplemented] | /proton-check user@proton.me | | /pgp-lookup [email] | PGP key search — creation date, UIDs, signatures [unimplemented] | /pgp-lookup dev@example.com | | /wifi [ssid] | WiFi SSID geolocation via Wigle.net T2: intel.py wifi "<ssid>". T1: wifi_ssid — needs a WiGLE account; discloses the gap | /wifi "HomeNetwork" | | /wifi --bssid [mac] | Exact AP lookup by MAC address | /wifi --bssid AA:BB:CC:DD:EE:FF | | /register [name] | Add a subject to the case workspace | /register JohnDoe | | /snapshots [url] | List/fetch archived Wayback snapshots. WebFetch is blocked from web.archive.org (robots.txt) — this reads the archive instead, so the request never reaches the target. T2: intel.py wayback-fetch <url> [--near latest\|earliest\|YYYY] [--list]. T1: wayback_fetch. See analysis/archive-explorer.md | /snapshots example.com | | /archive-harvest [domain] | Sweep a domain's whole Wayback history for indicators an operator has since scrubbed — the GA ID that clusters the estate is often only in an old capture. T2: intel.py wayback-harvest <domain> --indicators [--from YYYY --to YYYY]. T1: wayback_harvest | /archive-harvest site-a.example | | /fallback [domain] | Dead-seed recovery (§2.5) — crt.sh + full Wayback timeline + archive.today + local KB when a seed returns zero pivots / parked / NXDOMAIN; enumerates CT + Wayback host history before a seed is written off. T2: fallback · T1: fallback_probe | /fallback scam-site.top |

Enrich

Command	What It Does	Example
/branch [data]	Expand a discovered identifier laterally [model]	/branch john@mail.com
/pivot-suggest	Rank "what to pivot on next" from findings — leet/variant/reuse/temporal/domain clusters. T2: intel.py pivot-suggest <findings.json>. T1: pivot_suggest	/pivot-suggest
/email-permute [name|handle]	Derive email candidates from a person name or username against case domains. VN/CN/KR family-name-first aware; folds diacritics Unicode won't. --verify = MX gate + Gravatar. Output is hypotheses — see the rule below	/email-permute "Nguyen Van A" --domain example.com --verify
/rank-relations	Score + rank same-operator relations across analyzed pages (noise-filtered). Mechanizes one artifact = lead, two = cluster — run it before asserting a cluster. T2: intel.py rank-relations cases/<CASE>/raw/*.json. T1: rank_relations	/rank-relations
/crypto-balance [addr]	On-chain balance + lifetime flow for a wallet, valued at spot. T2: intel.py crypto-balance <addr>. T1: crypto_balance	/crypto-balance 1ExampleBitcoinAddressDoNotUse
/timeline [subject]	Assemble dated event sequence	/timeline Company Inc
/crossref	Detect shared identifiers across subjects T2: intel.py crossref [--case <id>]. T1: kb_crossref	/crossref
/link-subjects [A] [B]	Define a connection between two subjects [model]	/link-subjects John Jane
/show-connections	Display all logged connections [model]	/show-connections
/show-trail [subject]	Show the evidence chain for a subject [model]	/show-trail JohnDoe
/watch [subject]	Add subject to active tracking list [model]	/watch example.com
/record-finding	Log a finding with source and confidence [model]	Paste data after command
/show-findings	List all recorded findings [model]	/show-findings
/graph	Full ASCII subject relationship map	/graph
/pathfind [A] [B]	Discover connection path between subjects [model]	/pathfind A B
/diff [url]	Diff archived versions of a URL [model]	/diff example.com/page

Assess

Command	What It Does	Example
/exposure [target]	Composite exposure score (0–100) T2: intel.py exposure --set k=v. T1: exposure_score	/exposure domain.com
/threat-model	Build threat model from findings; every attribution claim carries an ACH matrix (competing hypotheses scored by inconsistency, runner-up named) per handbook/analytic-standards.md §3. Backend hook (Assess): if /backend is up, calibrate confidence on your own priors first — intel.py operators list + intel.py risk --case <id> + read knowledge/{calibration.jsonl,analyst_profile.md} — instead of scoring from scratch. See connectors/intel-backend.md §6 [model]	/threat-model
/signatures	Surface recurring behavioral patterns T2: intel.py signatures --set k=v. T1: signature_scan — evaluates, does not observe	/signatures
/validate	Quality audit — score 0–100 [model]	/validate
/coverage	Coverage matrix with identified gaps — technique matrix plus the 5W1H substantive pass (Why/How unanswered blocks Deliver-ready) [model]	/coverage
/verify-finding [id]	Re-check a specific finding's sources [model]	/verify-finding 12
/subject [name]	View or create subject record [model]	/subject JohnDoe
/lookup [name]	Retrieve a registered subject [model]	/lookup JohnDoe
/modify [name]	Update a subject record [model]	/modify JohnDoe
/archive-subject [name]	Remove subject from active tracking [model]	/archive-subject JohnDoe
/find [query]	Search across all subjects [model]	/find domain:example.com
/blind-spots	Prioritized investigation gap analysis [model]	/blind-spots
/source-check	Batch source URL accessibility check [model]	/source-check
/drift [subject]	Temporal risk score tracking T2: intel.py drift <case> [--snapshot]. T1: case_drift	/drift example.com
/clarify [finding]	Plain-language finding explanation [model]	/clarify fnd-003

Deliver

Command	What It Does	Example
/report	Full report — always saves the base data bundle (.md + .json + .csv + IOC .stix.json/.txt/.csv/.jsonl), then asks which presentation to render: (a) PDF · (b) DOCX · (c) HTML · (d) all	/report
/report html	Interactive self-contained HTML report (primary deliverable)	/report html
/report brief	Single-page executive brief	/report brief
/report json	Raw data as JSON	/report json
/report csv	Spreadsheet-compatible export	/report csv
/report docx	Word document in the PDF house style (slate/steel palette, serif body + sans headings, cover/TOC, rich charts + Diagram Design editorial diagrams + cloud figure) — on request	/report docx
/report legal	Evidence-formatted for legal proceedings (adds DOCX/PDF)	/report legal
/report journalist	Source-citation-heavy format	/report journalist
/brief	Plain-language summary (non-technical) [model]	/brief
/render entities	ASCII subject relationship diagram [model]	/render entities
/render timeline	Chronological event chart	/render timeline
/render risk	Exposure heatmap	/render risk
/render network	Network topology of connections	/render network
/stats	Counts and coverage statistics T2: intel.py stats	/stats
/workspace save [name]	Persist case state [model]	/workspace save mycase
/workspace open [name]	Resume a saved case	/workspace open mycase
/workspace list	Show saved cases	/workspace list
/workspace diff [a] [b]	Diff two saved workspaces	/workspace diff case1 case2
/render threat-path	ASCII attack path flow diagram	/render threat-path
/render attack-surface	ASCII attack surface exposure map	/render attack-surface
/report ioc	Export IOCs as STIX 2.1 or flat list	/report ioc --format stix
/redact [file]	Shareable variant of a report — stable numbered placeholders ([EMAIL_1]) + reversible JSON map; .md/.json/.csv. Opt-in — the base data bundle stays unredacted; request with /redact or --redact	/redact REPORT.md

UX & Navigation

Command	What It Does	Example
/flow [type]	Guided step-by-step case workflow [model]	/flow person
/template list	Browse pre-built case templates [model]	/template list
/template run [name]	Run a pre-built template	/template run security-audit
/novice	Toggle simplified, low-jargon mode [model]	/novice
/terms	OSINT term glossary [model]	/terms
/progress	Current case phase and coverage [model]	/progress
/opsec	OPSEC checklist for current task [model]	/opsec
/onboard	Interactive first-time onboarding guide [model]	/onboard
/quality	Investigation quality composite score [model]	/quality

Configure

Command	What It Does	Example
/apikeys	Manage premium/pro API keys (ChongLuaDao ⭐ first-party, Shodan, Censys, FOFA, SecurityTrails, DNSLytics, urlscan-PRO, WhoisXML, Hudson Rock, IntelX, GitHub, SerpAPI…) — status/set/unset/test/unlocks. Keys upgrade existing techniques (especially /cld + /webpivot); keyless/free stays the default. Stored chmod-600 in $SKILL_DIR/.env (gitignored), env-var override. See handbook/api-keys.md	/apikeys set chongluadao <KEY>
/backend	Detect/report the optional persistent-intelligence backend and pick the tier — Tier 1 typed MCP (intel-harness) → Tier 2 CLI → Tier 3 stateless. Runs scripts/backend/backend.py to resolve $INTEL_HOME (env → .mcp.json → sibling dir → symlink) and print the tier line. All the backend commands below dispatch through scripts/backend/intel.py <op> at Tier 2 (or the typed MCP tool at Tier 1). intel.py list maps all 73 engine ops (full CLI parity — CDN ranges, graph-build, hypothesize, calibration, evidence-report, case-store, cost, deterministic pipeline, …); intel.py mcp prints/writes the .mcp.json that enables Tier 1 ("the server"). See connectors/intel-backend.md	/backend · /backend check
/kb [query]	Built-in. Query the shared knowledge base. T2: intel.py kb --stats/--entity <v>/--cluster <domain>/--shared --min N; intel.py operators list. T1: kb_entity/kb_cluster/kb_query_shared	/kb --entity example.com
/recall [seed]	Built-in. "Have I seen this before?" — check a seed against every prior case before collecting. T1: which_cases/domain_verdict (typed MCP). T2: intel.py recall <seed> (query.py --entity; which_cases/domain_verdict are MCP-only). Surfaces known operators up front	/recall scam-site.top
/risk [case]	Built-in. Score a case's hosts for NRD / bulletproof-hosting / money-trail red flags. T2: intel.py risk --case <id> (or --file <pivot.json>). T1: risk_signals	/risk CASE-0001
/reverse-whois [email|name]	Built-in. Reverse-WHOIS a registrant identity → only high-value pivots; refuses privacy/registrar terms, flags bulk resellers as noise. T2: intel.py reverse-whois --reverse-email <e> --search-type historic --json. T1: reverse_whois	/reverse-whois owner@x.com
/cert-overlap [d1 d2 …]	Built-in. KB-aware TLS/SAN same-operator verdict (SHARED-CERT / SIBLING-OVERLAP / NO-CT-OVERLAP) across 2+ domains — corroborates a cluster at the TLS layer. Complements the keyless /cert-pivot. T2: intel.py cert-overlap a.com b.com. T1: cert_overlap	/cert-overlap a.com b.com
/reference [check|add|list]	Built-in. Curated false-positive control ledger — is a fingerprint BENIGN (common logo/CDN → don't cluster), SIGNAL (distinctive, prior-case → pivot), or UNKNOWN. T2: intel.py reference check <value>. T1: reference_check/reference_add	/reference check favicon:123
/pipeline [open|status] <case> <domains-file>	Built-in. The deterministic chain (no LLM key) — broad collect (cti-expert's pivot_extract) → ingest → prior-overlap → risk → shared-cluster → ICD-203 assessment, persisted under cases/<case>/. Prints collector: cti-expert. --no-collect = reuse mode: skip the live fetch and run the whole chain over pivot JSONs already in cases/<case>/raw/ — the zero-egress handoff /case uses (it re-fetches nothing). T2: intel.py pipeline open <case> seeds.txt [--no-collect] [--no-graph]	/pipeline open case1 seeds.txt
/harness [open|continue|status]	Built-in. The agent-driven whole-case orchestration (IntelHarness) — persistent, versioned, cross-case Collect→Correlate→Assess to convergence. Auto-escalates inside /case when the deterministic pipeline hasn't converged and posture is active (--no-harness opts out). Keyless-first: run interactively in Claude Code it uses the CLI's own model on your subscription; HARNESS_BACKEND=local (Ollama/vLLM/LM Studio, keyless) or an API key are needed only for unattended SDK continue. T2: intel.py harness open CASE-0001 <seeds…> · continue CASE-0001 --depth 4 · status [CASE-0001]. Persists to cases/; status needs no key	/harness status CASE-0001
/graph --render	Built-in. IntelGraph publication-quality render of a case graph → PNG/SVG (distinct from the ASCII /graph). Use it whenever a case has a graph — any /case, /pipeline, /harness or multi-node /report — not only on request; it degrades to a note if the dot/mermaid renderer is absent. ⚠ Fidelity check FIRST: case_graph.json holds only raw-backed hosts — clusters attributed via KB/reverse-WHOIS (no raw pivot) are absent, so an auto-render can silently under-represent the finding; compare its node count against the attributed cluster count (operators.jsonl / /clusters) and, if KB clusters are missing, hand-scope the figure (Mermaid/DOT of the real clusters) or caption it as the raw-backed subset. Large infra: above --max-nodes (default 45) the MAIN figure is reduced to a readable, representative subset (operator anchor + one hub per cluster + top-degree nodes; fan-out infra dropped first) so the picture never collapses into a hairball, and the figure title says so — the full indicator set stays in the IOC bundle (CSV/Excel · Markdown · JSONL). --full forces the whole graph; --split-clusters emits one figure per cluster. T2: intel.py graph <case_graph.json> <out-stem> --legend [--max-nodes N | --full]. T1: render_diagram	/graph --render case_graph.json out
/report pdf	Built-in. IntelReport pandoc render of an assessment .md → polished PDF + DOCX in the same editorial house style (slate/steel/ochre/brick palette, serif body + sans headings, cover/TOC/figures, VN-safe) — the DOCX now uses a matching reference.docx so the Word file reads as the same document as the PDF. Rebuild that reference with IntelReport/scripts/make_reference_docx.py. T2: intel.py report <assessment.md> <out-stem> --pdf --docx. T1: render_report	/report pdf assessment.md out
/clusters [case]	Built-in. Partition a case into same-operator clusters before judging it — the unit of judgment is the cluster, not the case. Shows each binding indicator's KB-wide prevalence, so an indicator that binds 3 domains here but sits on 400 KB-wide reads as noise. Pure KB read. T2: intel.py clusters <case>. T1: case_clusters	/clusters CASE-0001
/frontier [case]	Built-in. The case's unresolved gaps — free next seeds already discovered (crt.sh SAN, passive-DNS co-host, TLS co-SAN, CORS, reverse-WHOIS), the deferred metered leads held for approval, and OPEN enrichment leads (per registrant email + apex: /intelx·/breach-deep·/dork-sweep·/github-osint — leak/breach/OSINT/dork legs the infra pipeline never runs). A case is not done while enrichment leads remain; close each after running it with intel.py enrichment-done <case> --key <key>. reopen re-opens a converged case on new seeds. T2: intel.py frontier <case> · intel.py enrichment-done <case> --key <k> · intel.py reopen <case> <seed…>. T1: case_frontier/case_reopen	/frontier CASE-0001
/loop [case]	Built-in. Collect → assess repeatedly until the case converges, instead of stopping at an arbitrary depth. T2: intel.py loop <case>. T1: case_loop	/loop CASE-0001
/scope [case]	Built-in. Case intake: the no-touch class, victim ownership and the egress gate, derived up front rather than assumed mid-run. A defaulted value is never rendered as an answer. T2: intel.py scope <case>. T1: case_scope	/scope CASE-0001
/liveness [domain]	Built-in. Is it actually alive? A 200 parking/default/suspended/soft-404 page is not live and a 404/403/5xx/bot-wall is not dead — only NXDOMAIN reports dead, and every still-controlled name sets reuse_watch. T2: intel.py liveness <domain>. T1: domain_liveness	/liveness scam-site.top
/pssl [domain|cert]	Built-in. Passive SSL — the historical cert → IP direction that recovers an origin from behind a CDN, with the base-rate rail that keeps a shared CDN certificate out of the clustering. Free (same CIRCL account as passive DNS), so it is on by default in the pipeline. T2: intel.py pssl <target>. T1: passive_ssl	/pssl example.com
/paths [url]	Built-in. The URL path as a clustering indicator (path_kit:) — for an operator who rotates disposable hosts and selects the branded template by directory instead. A generic path (/login, /assets) emits nothing; the base-rate denylist is the whole reason this is safe. T2: intel.py paths <url>. T1: url_paths	/paths https://host/kitname/
/serp [domain]	Built-in. The advertising layer — Google Ads Transparency (who paid: a verified, billed advertiser identity) plus the cloaking probe with its falsification control. Opt-in per run: it spends a SerpApi search per host. T2: intel.py serp <domain>. T1: serp_ads	/serp scam-site.top
/docmeta [url|file]	Built-in. Document/image metadata — PDF /Info + XMP, EXIF (incl. GPS), PNG chunks — the author string an operator forgot to strip. Base-rate filtered on both the pivot and ingest paths. T2: intel.py docmeta <target>. T1: doc_metadata	/docmeta https://site/brochure.pdf
/screenshot [url]	Built-in. A rendered full-page PNG as timestamped, hashed visual evidence — the page as a human sees it (a channel bio naming admins, a members-area panel, a deposit page). Renders post-JS in a real browser, so it captures what the page displays, not what the DOM says. --verify re-hashes a stored capture. T2: intel.py screenshot <url> --case <id>. T1: capture_screenshot	/screenshot https://scam-site.top
/exhaust [case]	Built-in. Which collection layers actually RAN versus silently never fired. A layer that never executed looks identical to a layer that found nothing — this names the difference, so "no wallets" is not read as a fact about the operator when the wallet extractor never ran. T2: intel.py exhaust --file <pivot.json>. T1: collection_gaps	/exhaust CASE-0001
/misp-export [case]	Built-in. IntelShare — build a MISP event from the case's own collected pivots. Local only, no network: it writes the event JSON so you can read every attribute before anything leaves the machine. Sets TLP, distribution, threat level and tags. T2: intel.py misp-export <case> --tlp amber. T1: misp_export	/misp-export CASE-0001
/misp [search|push|publish]	Built-in. Two separate decisions, deliberately. search asks the cheaper question first — is this indicator already known to the instance? push stages the event on your instance, organisation-only and unpublished (a real write, but still deletable). publish syncs it to the community and cannot be recalled — every indicator becomes somebody else's blocking rule, so a false positive blocks innocent infrastructure on networks you will never see. Both write paths prompt via hooks/actionguard.py. T2: intel.py misp keycheck|budget|search|push|publish. T1: misp_search/misp_push/misp_publish	/misp search 1.2.3.4
/pivot-extract <.eml>	Built-in. pivot_extract also takes a victim's saved email. The .eml is parsed to its HTML body, so every HTML-side extractor runs over what the funnel actually sent — then header/CDN-derived selectors (sender domains, sending platform) emit pivots like any other artifact. This is the funnel's first hop, and no live fetch of the landing page can recover it. T2: intel.py pivot-extract ./saved.eml. T1: pivot_extract	/pivot-extract ./phish.eml
/victims [case]	Built-in. Infer the access vector from the victim set, plus demography (country + sector) — who was hit tells you how. T2: intel.py victims --case <id>. T1: victim_profile	/victims CASE-0001
/case-timeline [case]	Built-in. IntelGraph infrastructure-lifecycle timeline — registration/expiry spans, registrant eras, IP hosting windows, cert validity, archive visibility — with an evidence ledger citing every dated fact to an online source. T2: intel.py timeline <case>. T1: case_timeline	/case-timeline CASE-0001
/tool-calls [case]	Built-in. Audit what the model actually called during a run, including the denied calls — intel.py dashboard serves the same data as a loopback-only inspector (cost, trace, tool pairing). T2: intel.py tool-calls <case> · intel.py dashboard. T1: tool_calls	/tool-calls CASE-0001
/login-detect [url]	Built-in. Engage (detection half — passive and free): find the login form, the password field and the registration page, and classify by fields (a confirm-password means register; an invite code is a pivot, not an OTP). T2: intel.py login-detect <url>. T1: detect_login T2: intel.py login-detect <url>. T1: detect_login	/login-detect https://scam-site.top
/engage [url]	Built-in. GATED — outbound, attributable, irreversible. Create a synthetic-persona account and log in to read the members area (panel, deposit/withdraw flow, affiliate tree, support handles). Refuses without explicit confirmation, refuses a non-synthetic persona or direct egress, and stops at a CAPTCHA. Same gate class as a sandbox submission — ask first, always. T2: intel.py persona → intel.py engage <url> → intel.py engage-harvest → intel.py engage-report. T1: make_persona/engage_account/harvest_authenticated/engage_report	/engage https://scam-site.top

---

4. Subject & Connection Model

Reference: engine/case-schema.json, engine/subject-registry.md

Subject Types

Type	Emoji	Examples
Person	👤	Full name, alias
Username	@	Social handle
Email	📧	Address, domain
Domain	🌐	Site, subdomain
IP Address	🖥	IPv4, IPv6
Organization	🏢	Company, group
Phone	📱	E.164 format
Location	📍	GPS, address
Asset	📦	Document, image
Event	📅	Dated occurrence
Device	🖥️	IoT device, server, workstation
Image	🖼️	Photograph, screenshot
Crypto Address	💰	Bitcoin, Ethereum wallet
Bank Account	🏦	IBAN, local account no., BIC
ICP Filing	📋	PRC licence serial (one registrant, many sites)
Custom	🏷️	User-defined entity type

Connection Types

owns         — domain, email, or asset ownership
uses         — platform account or tool usage
works_at     — employment or affiliation
linked_to    — general association
alias        — same identity, different handle
communicated_with — observed contact

Finding Trust Scores

Score	Label	Meaning
5	PRIMARY	Authoritative or official source
4	DERIVED	Confirmed by 2+ independent sources
3	CONFIRMED	Single reliable source, verified
2	ANECDOTAL	Reported but unverified
1	CONTESTED	Conflicting data exists

Source Reliability Scale

Complements numeric trust scores with source-level grading. Trust score rates finding content; source reliability rates the source itself.

Grade	Label	Typical Sources
A	Completely Reliable	Official registries, government records
B	Usually Reliable	Established outlets, corporate sources
C	Fairly Reliable	Known blogs, industry publications
D	Not Usually Reliable	Anonymous forums, unverified claims
E	Unreliable	Known disinformation, fabricated content
F	Cannot Be Judged	Insufficient information to assess

Confidence Levels

Level	Label	Use When
VERIFIED	Direct observation, primary source
STRONG	Multiple corroborating sources
MODERATE	Single reliable source
WEAK	Circumstantial or inferred
TENTATIVE	Analyst deduction only
CHALLENGED	Contradicted by other findings

Likelihood Language (judgments, not findings)

The three scales above grade evidence. An analytic judgment built on that evidence — an attribution, a motive, a forecast — carries a probability-anchored likelihood term instead. Without an anchor, "MODERATE" routinely means a 30-point-different thing to writer and reader.

Term	Band		Term	Band
Almost no chance	1–5%		Likely / probable	55–80%
Very unlikely	5–20%		Very likely	80–95%
Unlikely	20–45%		Almost certain	95–99%
Roughly even chance	45–55%

Likelihood and confidence are orthogonal — report both:

The operator is very likely based in Guangdong (moderate confidence — single registry record, unverified).

Never 0% or 100%. One term per judgment. Never attach a likelihood term to a directly observed fact. findings[].confidence in the report JSON stays an integer describing evidence quality — likelihood lives in the narrative.

Attribution claims additionally require an ACH matrix (competing hypotheses, scored by inconsistency, runner-up named). Full rules — likelihood, the 5W1H coverage overlay, and ACH: handbook/analytic-standards.md.

Map Rendering (ASCII Mandatory)

ALL visualization commands produce ASCII box-drawing art by default. This includes /graph, /render entities, /render network, /render timeline, /render risk, /pathfind, and /show-connections. Mermaid available only with explicit --mermaid flag.

Why ASCII-first: Universal terminal compatibility, renders correctly in .md and .docx exports, no external renderer dependency.

┌─────────────────────────────┐   owns   ┌───────────────────────────┐
│ 👤 John Doe          [3/5] │══════════▶│ 🌐 example.com     [4/5] │
└─────────────────────────────┘           └───────────────────────────┘
         │ works_at                       │ hosted_on
         ▼                                ▼
┌─────────────────────────────┐  ┌───────────────────────────┐
│ 🏢 Acme Corp         [4/5] │  │ 🖥 203.0.113.10    [4/5] │
└─────────────────────────────┘  └───────────────────────────┘

Connection arrows: ═══▶ owns · ───▶ confirmed · ···▶ inferred · ←─▶ bidirectional · ─·─▶ alias · ╌╌▶ works_at Box styles: ┌──┐ confirmed · ┌ ─ ┐ unverified · ╔══╗ target Badge: [n/5] trust score · emoji prefix = entity type

---

5. Finding Framework

Reference: engine/finding-framework.md, engine/conflict-resolver.md

Every finding logged via /record-finding captures:

Source URL / method
Collection method (browser | search | fetch | manual)
Trust score (1–5)
Confidence level (VERIFIED → CHALLENGED)
Timestamp
Linked subjects

Conflict detection (engine/conflict-resolver.md): When two findings about the same subject contradict each other, the system flags a CONTESTED state. Both findings are preserved. Resolution options: accept one, mark both TENTATIVE, or log the conflict as its own finding.

Deviation detection (analysis/deviation-detector.md): Automatically flags behavioral anomalies — account creation gaps, platform presence inconsistencies, metadata mismatches.

Weight engine (analysis/weight-engine.md): Aggregates trust scores across findings to compute subject-level confidence.

---

6. Technique Catalog

Reference directory: techniques/

File	Covers
fx-metadata-parsing.md	EXIF, email headers, document metadata analysis
fx-image-verification.md	Image authenticity and provenance workflow
fx-breach-discovery.md	Breach database methods and paste site search
fx-geolocation.md	GPS extraction, W3W, Plus Codes, MGRS, Street View
fx-social-topology.md	Social graph construction and topology
fx-email-header-analysis.md	Header analysis, SPF/DKIM, SMTP routing
fx-document-forensics.md	Document forensics and metadata extraction
fx-http-fingerprint.md	HTTP fingerprinting and server signature analysis
fx-leak-monitoring.md	Leak and breach monitoring, paste site search

<!-- dork-integration:phase-05 start -->

| fx-dork-sweep.md | Zero-auth Google/Bing dork sweeps — Telegram ecosystem, doc-hosts, filetype families + 4-tier fallback cascade (WebSearch → Bing → DDG → agent-browser) | | fx-document-leak-hunt.md | 18-platform document leak discovery with severity classification, paywall handling, auto-snapshot |

<!-- dork-integration:phase-05 end -->

| username-osint.md | 3000+ platform enumeration with pivot extraction | | phone-osint.md | Carrier lookup, VoIP detection, spam databases, FreeCNAM CallerID, WhoCalld, USPhoneBook reverse lookup | | email-osint.md | Full email investigation: accounts, breaches, infra, Proton API, PGP keys, permutation, manual reference tools | | fx-dns-cert-history.md | Historical DNS records (passive DNS, A/NS/MX changes), SSL certificate timeline (crt.sh CT logs) | | threat-intel.md | AbuseIPDB, GreyNoise, OTX, VirusTotal, URLScan.io, CIRCL CVE, NVD API, ransomware.live | | web-traffic-analysis.md | SimilarWeb/Semrush estimation, audience data | | secret-scanning.md | Credential/secret detection in repos and pastes | | github-osint.md | GitHub user/org/repo profiling, code search, commit metadata, forks, collaboration networks | | domain-advanced.md | Subfinder, Amass, CT log enumeration | | social-media-platforms.md | Twitter/X Snowflake IDs, Discord, Strava, BlueSky, ShareTrace share link analysis | | advanced-geolocation-techniques.md | Overpass Turbo, road sign analysis, reflected text | | web-dns-forensics.md | Zone transfers, Tor lookups, GitHub, Telegram, WHOIS, Xeuledoc Google doc intel | | fx-visitor-intelligence.md | Visitor stats, tech stack, geo, traffic sources, analytics/AdSense/advertising ID cross-domain linking, competitors | | wifi-ssid-osint.md | WiFi SSID/BSSID geolocation via Wigle.net, encryption analysis, travel patterns | | scam-check.md | Phishing/scam domain verification and detection | | cloud-audit.md | Cloud infrastructure security (AWS/GCP/Azure): IAM, network, storage, compute, logging, secrets | | microsoft-tenant-recon.md | M365/Azure tenant enumeration — federation, tenant ID, Azure AD config, MDI detection | | china-recon.md | China/Sinophone layer — ICP filing → PRC entity + licence-serial sibling pivot, GSXT/信用中国/TianYanCha/QCC/Aiqicha registry chain, USCC validation, Quake/ZoomEye/FOFA cyberspace engines, Baidu dorking, CJK pinyin + Traditional variant generation, CN social platforms, access-reality gaps | | fiat-payment-osint.md | Bank accounts as selectors — IBAN mod-97 validation + BBAN decomposition, BIC, VN/SEA non-IBAN rails (VietQR/NAPAS BIN), account-reuse pivot, mule-pattern signals | | fx-edge-appliance-recon.md | Edge/VPN appliance fingerprint → CISA KEV/CVE catalog (Citrix/F5/Cisco/Ivanti/Forti/PAN/Exchange) + exposed-service port-risk matrix (Shodan InternetDB, passive-first) | | fx-saas-identity-recon.md | SaaS tenancy + identity-fabric mapping — DNS-TXT tenancy tokens, IdP fingerprinting (Okta/Auth0/OneLogin/Ping/Keycloak/ADFS/Entra), unauthenticated API/GraphQL/OpenAPI-spec discovery | | dependency-audit.md | Supply chain security: CVE audit, framework-specific vulns, typosquatting, CI/CD security | | disk-forensics.md | Digital evidence analysis: image integrity, Sleuth Kit, file carving, artifact recovery, timeline | | incident-triage.md | Security incident response: NIST 800-61 methodology, containment, evidence preservation, IOC extraction | | owasp-audit.md | OWASP Top 10 (2021) source code audit with grep patterns and CWE references | | prompt-injection-audit.md | AI/LLM security: prompt injection classes, agent/MCP security, permission boundary audit | | stealer-log-analysis.md | Infostealer-log triage: family fingerprinting (RedLine/Vidar/StealC/Lumma/META/traffer), victim-vs-operator profiling, cross-log actor correlation, IOC + attribution extraction (uv run parser, raw artifacts shown) | | agent-browser.md | Interactive browser collection & evidence capture via vercel-labs/agent-browser (CDP, accessibility-tree @eN snapshots, screenshots; primary interactive collector, complementary to Scrapling) | | media-vision-analysis.md | Image/A-V content analysis — vision OCR + sign/landmark/logo/face read (multix gemini, Gemini) and FFmpeg keyframe/audio preprocessing → transcription; extracted text/GPS/entities re-enter the pivot loop (standalone npx multix + ffmpeg, no AgentKit/sibling-skill dep) | | phishing-domain-survival.md | Registration/DNS strategy profiling — maliciously-registered vs compromised classification + takedown-survival outlook from WHOIS/DNS (scripts/phish_domain_survival.py, offline, zero-dep); eCrime 2026 "Built to Last?" | | clickfix-clipboard-hijack.md | ClickFix / PasteJacking clipboard-hijack detection — clipboard-write + lure + OS-command co-occurrence, decodes PowerShell -EncodedCommand → C2 IOCs (scripts/clickfix_detect.py, offline, zero-dep); eCrime 2026 "PasteJacked" | | visibility-aware-html.md | Visibility-aware HTML analysis — hidden credential forms / off-origin links / off-screen brand text a naive parser misses (scripts/html_visibility_analysis.py, offline, zero-dep); eCrime 2026 "Visibility-Aware HTML Analysis" | | apk-permission-scope.md | APK permission-scope risk scoring — dangerous-permission combos (accessibility+overlay+SMS) = on-device-fraud capability; combination is the signal, capability≠guilt (scripts/apk_permission_scope.py, offline, zero-dep); eCrime 2026 "The 'Allow' Reflex" | | kit-template-attribution.md | Phishing kit/template structural fingerprint + similarity → same-kit lineage; commodity-template match graded as noise, never auto-merged (scripts/kit_template_fingerprint.py, offline, zero-dep); eCrime 2026 tree-structured attribution | | renderer-confirmation.md | Renderer-level confirmation of ClickFix + visibility — feeds runtime clipboard writes / computed-hidden elements into the static detectors and reconciles (scripts/render_confirm.py; renderer optional, degrades to a note); eCrime 2026 PasteJacked + Visibility-Aware HTML | | phishtrace-dynamic-features.md | Runtime-trace phishing characterization — redirects / exfil endpoints / cloaking; exfil hosts are IOCs not attribution; thin trace on a flagged page = cloaked, never benign (scripts/phishtrace_features.py, offline, zero-dep); eCrime 2026 "PhishTrace" |

---

7. Workflow Guides

Reference directory: workflows/

Guide	Intended User	File
Journalist Source Verification	Journalists verifying claims	wf-journalist.md
HR Screening	HR professionals running background checks	wf-hr-screening.md
Cyber Threat Intelligence	Security analysts tracking adversaries	wf-threat-analyst.md
Private Investigator	Licensed PIs running person cases	wf-private-investigator.md

Activate via /flow [type] — interactive guided prompts walk through each step.

---

8. Output Formats

Reference: output/reports/, connectors/

Conversational domain table — show this in chat on every collection turn (CRITICAL)

Whenever a collection command (/cti, /case, /sweep, /webpivot, /subdomain, or the pipeline) returns one or more domains, the reply's first element — before any prose — is a markdown table summarizing each domain, so the operator sees the yield at a glance in the conversation. The §8 file exports are the durable record; this table is the live view and is never skipped, even for a single domain.

Domain	Resolves	Top pivots	Risk	Cluster / peers	Seen before
site-a.example	✓	favicon:123456789 · G-XXXXXXXXXX	NRD, BPH	3 peers	CASE-0001 (Operator A)
site-b.example	✓	registrant@example.com	—	3 peers	new

Columns: Resolves ✓/✗ (collector got a host); Top pivots the 1–3 highest-rung indicators (§2.5 ladder) as kind:value; Risk the risk_signals flags (NRD / BPH / money-trail, or —); Cluster / peers shared-indicator peer count from the KB; Seen before the prior case + operator from /recall, or new. Keep to these columns — detail goes in the prose below. One row per domain; for a large sweep show the top 20 by risk and note how many rows were omitted.

Mandatory File Export (CRITICAL)

Every /report, /brief, and /case command runs four steps at the end of delivery — an import prompt, then the always-saved base bundle, then the presentation prompt, then an export confirmation whenever HTML is among the outputs:

Step 0 — ASK whether to import more manually-collected evidence (before building ANYTHING, so it enriches every artifact). Ask the user — via AskUserQuestion — "Import more evidence data collected by manual investigation? (extra findings, subjects, indicators, selectors, timeline events, sources, screenshots, notes)". If yes, fold it into the report JSON (see Step 1) BEFORE Step A:

• Findings / subjects / indicators / selectors / timeline / sources — append them to the corresponding arrays of the report JSON (same schema as Step 1); each still needs a source_url/source note and a confidence.
• Screenshots / evidence images — base64-embed into evidence_images[] with evidence-images.py: explicit files uv run "$SKILL_DIR/scripts/evidence-images.py" shot1.png shot2.png --caption "…" [--finding FND-003] --into REPORT.json, or a whole case dir … --case <case-dir> --into REPORT.json. Re-run the Step 1 build / dash-normalizer after merging, then continue. Only once the manual evidence is in the JSON do Step A and Step B run — so the base bundle, the IOC bundle, and the chosen presentation report all include it. If no, proceed straight to Step A. (--yolo/guided-auto skip this prompt.)

Step A — ALWAYS auto-save the base data bundle (no prompt, every run):

#	Format	File	Role
1	Markdown	CTI-REPORT-[CASE-ID]-[YYYY-MM-DD].md	Diffable, greppable source of truth; also the input to the HTML/DOCX/PDF generators
2	JSON	CTI-REPORT-[CASE-ID]-[YYYY-MM-DD].json	Structured case data (the report JSON below); feeds the generators and downstream tooling
3	CSV	CTI-REPORT-[CASE-ID]-[YYYY-MM-DD].csv	Findings (and indicators, via the IOC export) for spreadsheets / SIEM lookups
4	IOC / selector bundle	IOC-[CASE-ID]-[YYYY-MM-DD].{stix.json,txt,csv,jsonl}	Comprehensive indicators & selectors — STIX 2.1 + flat + CSV + JSONL

Step B — ASK which presentation report to render, then build the choice on top of the base bundle. This prompt is the default (interactive) behavior:

Choice	Format	Builder
a	PDF — the editorial house report (IntelReport: cover, Roman-numbered sections I–XI, Methodology with the Admiralty × ICD-203 scales, relationship graph, entity relationship map, attribution inference chain, temporal view, registration heatmap + domain × shared-indicator matrix, ICD-203 × Admiralty confidence scatter, a captured landing page per estate host inline in the cluster section, per-domain dossiers, Appendices A–E incl. glossary). For a case dir this is deterministic: python3 "$SKILL_DIR/scripts/backend/intel.py" house-report <CASE-ID> writes cases/<CASE-ID>/report/CTI-REPORT-<CASE-ID>-<date>.{pdf,docx,md} + figures. Egress caveat: hosts with no screenshot on disk are rendered in headless Chromium through the research-egress proxy policy (proxied, or direct only if the store allows it; blocked → stated in the report, never forced); when the live page will not render or renders empty, a public web-scan screenshot or a rendered web-archive snapshot stands in — captioned as such, dated by the archive, linked in Appendix B, and flagged as a previous owner's page when it predates the current registration; --no-screenshots keeps the build fully offline, --no-archive-fallback disables the stand-in, --max-screenshots N / --screenshot-timeout S bound it. The four analytic charts are the dashboard's own (scripts/cti_report_figures.py, shared with generate-cti-docx-hybrid.py), drawn from report/report-data.json that build_report_data.py derives from the case dir. No case dir → generate-cti-docx-hybrid.py … --pdf (dashboard style)	intel.py house-report · fallback generate-cti-docx-hybrid.py --pdf
b	DOCX — the editable twin of (a), emitted by the same house-report run; dashboard-style alternative (charts + confidence matrix + heatmaps) via generate-cti-docx-hybrid.py	intel.py house-report · alt generate-cti-docx-hybrid.py
c	HTML — interactive, self-contained, OFFLINE (charts + 2D entity graph + topology + timeline + indicator panel + search); the primary human-facing deliverable	generate-cti-html.py
d	All — build PDF and DOCX and HTML	all of the above

Step C — CONFIRM before exporting HTML (choice c or d, and the explicit /report html). Before running generate-cti-html.py, get the Blueprint plan — python3 "$SKILL_DIR/scripts/cti_archify.py" REPORT.json --plan prints, offline and without rendering, what auto and force would embed (e.g. auto embed — apex level: 10 of 45 apexes shown for 105 hosts … / force embed — … 23 of 45 apexes …) — state it with the output path, then ask, via AskUserQuestion, "Export the HTML report now? Blueprint: (1) Auto — full graph if it fits, else the compact apex-level fold (recommended) · (2) Force — the widest map Archify's grid holds (CTI_ARCHIFY=force: up to 25 nodes, smaller labels, may need the frame's zoom) · (3) No Blueprint (CTI_ARCHIFY=0) · (4) Skip the HTML export". Run the generator with the matching env var and quote its Blueprint (Archify): line back so the analyst sees what was embedded (or the renderer's exact rejection). --yolo/guided-auto skip this prompt and build with Auto.

Why two PDF/DOCX builders. house-report composes the document — the analyst assessment folded into the house structure with figures rendered from the case data (IntelGraph Mermaid graph with a representative subset above 14 nodes, Graphviz inference chain from assessment.json, temporal view from the raw pivots, temporal-correlation tables from the timeline events JSON, landing pages from screenshots/ + evidence/screenshots/manifest.json with full SHA-256 in the evidence ledger) and pandoc + the LaTeX house template. It scrubs internal tool / vendor / path names to public source classes (house Rule 12) and never emits the impersonated brand's genuine domain as an indicator. generate-cti-docx-hybrid.py is the dashboard rendering of the flat report JSON — keep it for report-JSON-only cases and for the chart appendix. Save location: Current working directory, or ./osint-reports/ subdirectory if it exists.

Attribution (MANDATORY — every exported artifact). Every deliverable MUST credit this skill: Generated by CTI Expert — https://github.com/7onez/cti-expert. The generators inject it automatically — HTML (sidebar + page footer), DOCX (page footer, both generators), and the IOC bundle (flat header comment, CSV header comment, STIX producer identity description + bundle x_cti_generator). For the analyst-authored Markdown report, end the file with a footer line carrying that exact credit, e.g.:

---
*Generated by [CTI Expert](https://github.com/7onez/cti-expert) — OSINT/CTI toolkit.*

The JSON deliverable carries a top-level "generator" field (see the report-JSON schema in Step 1) — generators ignore unknown keys, so it is safe and shipped in the file itself. Never strip the credit — the redactor leaves it intact (it is not PII).

The default set is unredacted — it is the analyst's working record. A shareable variant is opt-in, never automatic, so nothing is ever quietly weakened. Request it with /redact or /case … --redact:

S="$SKILL_DIR/scripts"; R="CTI-REPORT-[CASE-ID]-[YYYY-MM-DD]"
for f in md json csv; do
  uv run "$S/redact.py" "$R.$f" -o "$R.redacted.$f" --map "$R.map.json"
done

One --map across all three files keeps a selector's placeholder identical everywhere. Infrastructure (URL/domain/IP) stays visible even then — in a CTI report the actor's infrastructure is the analysis, not incidental PII; add --all-types to cover it too. Never ship the .map.json — it reverses the redaction.

• Interactive mode (DEFAULT) — three prompts, in order: (1) first ask "Import more evidence data collected by manual investigation?" (Step 0) and merge any supplied data into the report JSON; (2) then, after saving the base data bundle, ask — via AskUserQuestion — "Which report format(s) do you want? (a) PDF · (b) DOCX · (c) HTML · (d) All" (multi-select allowed); (3) when HTML is among them, confirm the export and its Blueprint mode (Step C: Auto · CTI_ARCHIFY=force · CTI_ARCHIFY=0 · skip) before running generate-cti-html.py, then build exactly the chosen presentation format(s). Every choice ships alongside the base bundle (.md + .json + .csv + IOC .stix.json/.txt/.csv/.jsonl).
• --yolo / guided-auto / non-interactive: skip all three prompts — no evidence-import question, no export confirmation (Blueprint runs in Auto), and default to HTML (lightest, zero external toolchain) on top of the base bundle. /report legal defaults to All (evidentiary — a fixed Word/PDF artifact is expected).
• Explicit subcommands bypass the format prompt and emit that format directly on top of the base bundle: /report html, /report docx, /report pdf, plus the machine-only /report json, /report csv, /report ioc. /report html still runs the Step C export confirmation (it is the prompt that lets the analyst pick the Blueprint mode).
• DOCX/PDF caveat: two builders, two toolchains. Case dir → intel.py house-report (IntelReport: pandoc + xelatex; the editorial document with sections I–XI, both confidence scales, relationship graph, entity map, inference chain, temporal view, heatmaps, confidence scatter, landing-page captures with web-scan / web-archive stand-ins, domain dossiers, Appendices A–E; PDF and DOCX are one composition rendered twice; the landing-page step is the only egress — proxy-gated, --no-screenshots to skip, --no-archive-fallback to forbid the stand-ins). No case dir → generate-cti-docx-hybrid.py (python-docx + LibreOffice; the dashboard DOCX with the two-axis ICD-203 × Admiralty confidence matrix and the campaign heatmaps, and --pdf renders that same DOCX to PDF via scripts/cti_docx_pdf.py). Both are the heaviest, most failure-prone step: check pandoc, xelatex, dot, mmdc (house) or LibreOffice (dashboard) before promising a PDF, and fall back to HTML — which needs nothing — when a toolchain is missing. house-report scrubs internal tool/vendor/path names and masks third-party e-mails/phones/case ids (keeping the operator's own join key); the dashboard path relies on the analyst masking the Markdown first.
• Dash normalization (MANDATORY — covers every deliverable). Em/en dashes (—, –) read as machine-authored, so no export may ship them. Two layers guarantee this: (1) the HTML, DOCX and pandoc PDF/DOCX generators normalize their prose automatically via scripts/cti_text_normalize.py; (2) for the on-disk Markdown and JSON deliverables — which no generator rewrites — run the normalizer in place as the final step before confirming files:

  S="$SKILL_DIR/scripts"; R="CTI-REPORT-[CASE-ID]-[YYYY-MM-DD]"
  uv run "$S/cti_text_normalize.py" "$R.md" "$R.json"   # rewrites — → - in place; no-op if clean
  

  Prefer writing plain - while drafting so this step is a no-op. It rewrites only when a dash is present and is safe for .md/.json/.csv alike (typographic dashes occur only inside string content, never in structural syntax).

The HTML, JSON, CSV and IOC outputs all derive from one report JSON. Build it once, then run the generators below.

Step 1 — Build the report JSON file. The generators expect a SPECIFIC flat format (NOT the engine case-schema.json).

• For a case produced by the pipeline (/cti · /case · /harness) — build it deterministically. Do NOT hand-author. Run the converter; it reads raw/*.json, whois/*.json, assessment.md (BLUF + recommendations), evidence/*.json (leak-sweep, estate-seo-sweep) and the operator ledger (knowledge/operators.jsonl) — and consults clusters.json only to flag a multi-operator case — then emits the exact flat schema below: operator/registrant subject with selectors, the full domain estate as network indicators, findings/timeline/connections, and the §2.5 exclusion set (CDN/shared IPs, registrar, nameservers) in ioc_exclude so no excluded value ever leaves as an IOC:

  uv run "$SKILL_DIR/scripts/build_report_data.py" "${INTEL_HOME:-$SKILL_DIR/intel_engine}/cases/[CASE-ID]" -o "CTI-REPORT-[CASE-ID]-[YYYY-MM-DD].json"
  

  Then apply the Step 0 import (merge any analyst-supplied findings/subjects/selectors/timeline) onto that JSON before rendering. The manual schema below is the field reference for that enrichment — and the fallback when there is no case dir (e.g. a report authored purely from pasted evidence).
• Otherwise (no case dir) — construct the JSON by hand matching this exact structure. Reference: scripts/sample-cti-report-data.json.

{
  "generator": "CTI Expert — https://github.com/7onez/cti-expert", // MANDATORY attribution (top-level)
  "case": {
    "id": "CTI-2026-001",          // string, case identifier
    "label": "Case Title",         // string, human-readable name
    "classification": "OPEN SOURCE", // string
    "analyst": "AI-Assisted CTI",  // string
    "date": "2026-04-08",          // ISO date
    "subject": "target.com",       // string, primary subject
    "status": "active"             // string
  },
  "executive_summary": "Full paragraph summarizing investigation findings...",
  "subjects": [
    {
      "id": "SUB-001",            // string ID (not UUID)
      "label": "target.com",      // human-readable name — REQUIRED for display
      "type": "domain",           // lowercase: domain, person, ip, organization, email, username
      "confidence": 95,           // INTEGER 0-100 (not string like "VERIFIED")
      "verified": true,           // boolean
      "aliases": ["alias1"],      // string array
      "first_seen": "2025-01-15", // ISO date string
      "notes": "Primary domain"   // string
    }
  ],
  "findings": [
    {
      "id": "FND-001",            // string ID
      "subject_id": "SUB-001",    // links to subject
      "type": "infrastructure",   // credential, infrastructure, identity, exposure, behavioral, legal
      "weight": "HIGH",           // CRITICAL, HIGH, MEDIUM, LOW, INFO — drives severity colors
      "description": "Full description of the finding...",
      "source_url": "https://...",
      "collected_at": "2026-04-08T10:00:00Z",
      "confidence": 88,           // INTEGER 0-100 (not string)
      "tags": ["tag1", "tag2"]
    }
  ],
  "connections": [
    {
      "id": "CON-001",
      "from_id": "SUB-001",       // subject ID
      "to_id": "SUB-002",         // subject ID
      "relationship": "owns",     // string describing relationship
      "strength": "confirmed"     // confirmed, probable, possible
    }
  ],
  "timeline": [
    {"date": "2025-01-15", "event": "Domain registered"}
  ],
  "sources": [
    {"name": "Source Name", "url": "https://...", "date": "2026-04-08"}
  ],
  "intelligence_gaps": [
    "Gap description string"
  ],
  "recommendations": [
    "Action item string"
  ],
  "visitor_stats": {              // optional — enables visitor intelligence charts
    "domain": "target.com",
    "monthly_visits": 150000,
    "traffic_sources": {"direct": 42, "search": 28, "referral": 15, "social": 10, "paid": 5},
    "top_countries": [{"country": "Vietnam", "share": 60}, {"country": "US", "share": 20}]
  },
  "caveats": ["Caveat string"],  // optional — overrides default methodology notes
  "evidence_images": [           // optional — screenshots that visually back findings
    {
      "caption": "Phishing login impersonating Bank X",  // human label
      "host": "phish.example",                            // optional
      "data_uri": "data:image/png;base64,....",           // REQUIRED — offline-safe base64
      "sha256": "....",                                   // citable; auto-added by evidence-images.py
      "captured_at": "2026-04-08T10:00:00Z",              // optional
      "source_url": "https://phish.example/login",        // optional
      "finding_id": "FND-001",                            // optional — ties the shot to a finding
      "subject_id": "SUB-001"                             // optional
    }
  ]
}

CRITICAL FORMAT RULES:

• confidence on subjects and findings MUST be an integer (e.g., 85), NOT a string (e.g., "VERIFIED")
• findings MUST be a flat top-level array, NOT nested inside subjects
• label is REQUIRED on each subject (this is what displays in the report — not value or display_name)
• weight on findings drives severity coloring — use CRITICAL/HIGH/MEDIUM/LOW/INFO
• recommendations must be an array of strings (not objects with priority/action keys)
• All fields shown above should be populated with actual data — empty strings or "N/A" defeat the purpose
• Populate executive_summary with a full paragraph — this is the most-read section of the report

Optional enrichment fields (backward-compatible — used by the HTML report & IOC export when present):

• subjects[].role — actor | victim | infrastructure | associate | witness (drives the role chips and actor↔victim attribution; otherwise inferred from type/links)
• subjects[].selectors[] — contact/social points attached to a person/org: {type, value, platform, url} (e.g. a victim's phone, an actor's Telegram or LinkedIn) — surfaced in the Indicators panel and IOC export
• indicators[] — analyst-curated indicators to force into the export verbatim: {type, value, category, role, confidence, source_url}
• evidence_images[] — screenshots that back findings, embedded into the HTML Evidence view and the DOCX Visual Analytics → Evidence Screenshots section (base64, offline-safe): {caption, host, data_uri, sha256, captured_at, source_url, finding_id, subject_id}. Capture evidence during the investigation whenever it fits — a confirmed phishing/login page, a scam members-panel, a defacement — via the capture_screenshot tool (harness/MCP; citable sha256, written under cases/<case>/), or the deterministic pipeline's --screenshots flag (intel.py open … --screenshots → cases/<case>/screenshots/). Then build the array mechanically: uv run "$SKILL_DIR/scripts/evidence-images.py" --case <case-dir> --into REPORT.json (base64-embeds each PNG, dedupes by sha256; --finding FND-00X ties shots to a finding). A hostile target with no proxy is auto-skipped by the egress gate, so capture never touches infrastructure it must not.

Step 2 — Generate the interactive HTML report (PRIMARY human-facing deliverable). Self-contained, OFFLINE, zero toolchain to view — opens in any browser:

S="$SKILL_DIR/scripts"     # $SKILL_DIR = dir containing SKILL.md
uv run "$S/generate-cti-html.py" "REPORT.json" "REPORT.html"   # any OS, zero setup
# no uv installed: python3 "$S/generate-cti-html.py" "REPORT.json" "REPORT.html"   (Windows: py …)

It injects the report JSON into cti-report-template.html and renders, entirely client-side and offline (no CDN, no network calls): KPI cards, a finding-type pie, severity bars, a draggable/zoomable 2D entity graph, infrastructure topology, an event timeline, and the comprehensive Indicators & Selectors panel (network IOCs + contacts + identities + social/messaging handles + wallets + actor↔victim attribution) — with global search, category menus, dark/light themes and a print-to-PDF stylesheet.

The report also embeds an interactive Blueprint view — the case entity graph rendered as an Archify architecture diagram (typed nodes, directional relationships, exposure/finding cards) and inlined into the single offline file via an isolated iframe (theme + PNG/SVG export live inside it). This is the one build step that uses Node.js (a trimmed, zero-dependency Archify copy is vendored at scripts/vendor/archify/). It is best-effort: if Node.js is unavailable, the case has no subjects, or CTI_ARCHIFY=0 is set, the report is still produced and the Blueprint tab is simply omitted — every other view (including the always-present 2D entity graph) is unaffected.

Archify's architecture type draws small maps (≤ 12 nodes · 18 edges, cti_archify.BLUEPRINT_LIMITS), so a dense estate is folded to apex level automatically: every host under its registrable domain (PSL-aware, Estate · N hosts), the long tail of apexes into one +N more apexes node, apexes carrying a finding ranked first, the operator hub placed mid-row with its spokes above and below. The full host list stays in the Network Graph and Editorial views. The CTI_ARCHIFY env var selects the mode and the generator prints the outcome on its Blueprint (Archify): line — quote that line back to the analyst:

CTI_ARCHIFY	Blueprint
1 (default)	full graph when it fits, else the apex-level fold, else skipped with the reason
force	past the density gate: the full graph when it has ≤ 25 nodes, else the apex fold at the widest grid Archify holds (12 spokes above and below the hub, cti_archify.FORCE_MAX_NODES) — more apexes, smaller labels; Archify's own validator still has the last word and any rejection is printed verbatim
0	never embedded

It also renders two print-ready figures shared across HTML, PDF and DOCX: an Editorial view — the entity map and infrastructure topology drawn as Diagram Design editorial SVG (atomic-tangerine accent on 1–2 focal nodes, hairlines, no shadows), inlined as vector in HTML/PDF and rasterized via cairosvg for DOCX; and, only when a case reveals cloud infrastructure, a Cloud Architecture figure applying Diagram AI Generator — real AWS/Azure/GCP/K8s provider icons via the diagrams library + graphviz. Both are best-effort: the editorial figures fall back to the matplotlib DOCX diagrams if cairosvg is unavailable, and the cloud figure auto-skips when no cloud infra is detected, graphviz/diagrams are missing, or CTI_CLOUD_ARCH=0 is set.

Step 3 — Generate the comprehensive IOC / selector bundle.

uv run "$S/generate-cti-iocs.py" "REPORT.json" "IOC-[CASE-ID]-[YYYY-MM-DD]" --format all
# single format: --format stix | flat | csv

Extracts EVERY indicator that profiles or can reach an actor/victim — network IOCs, emails/phones, usernames/names/aliases, social-media profiles, messaging handles, crypto wallets, and the attribution links between subjects. Full spec: techniques/ioc-export.md.

Step 4 — DOCX (on request, or automatically for /report legal). Word is no longer auto-generated by default. When the user asks for it (or for evidentiary reports), build it from the SAME report JSON + MD. The generators carry PEP 723 inline dependency metadata, so the simplest, most portable runner is uv run — it provisions the deps on the fly with zero venv/pip setup, identically on every OS. The generator is also self-healing: it forces UTF-8 output and auto-locates pandoc (including Windows %LOCALAPPDATA%\Pandoc), so no PYTHONUTF8 / PATH prelude is needed. Replace REPORT with CTI-REPORT-[CASE-ID]-[YYYY-MM-DD].

Preferred — uv run (any OS, any agent, zero setup):

S="$SKILL_DIR/scripts"     # $SKILL_DIR = dir containing SKILL.md (Claude Code: ~/.claude/skills/cti-expert; Codex/clone: the repo)
# Primary: HYBRID — full narrative from MD + charts/diagrams from JSON (zero content loss)
uv run "$S/generate-cti-docx-hybrid.py" "REPORT.md" "REPORT.json" "REPORT.docx"
# PDF — DOCX-STYLE document (the DOCX rendered by LibreOffice), NOT an HTML print:
uv run "$S/generate-cti-docx-hybrid.py" "REPORT.md" "REPORT.json" "REPORT.docx" --pdf
# Fallback 1: JSON-only (charts + structured data; no pandoc needed)
uv run "$S/generate-cti-docx.py" "REPORT.json" "REPORT.docx"
# Fallback 2: MD-only (styled narrative, no charts)
uv run "$S/generate-cti-docx-hybrid.py" "REPORT.md" "REPORT.docx"

Windows PowerShell: set $S = "$env:USERPROFILE\.claude\skills\cti-expert\scripts" (Claude Code) or "<repo>\scripts" (Codex/clone), and use backslash paths.

Fallback — no uv installed. Use the OS interpreter; the script's ensure_deps() installs the libs on first run (via uv if present, else pip):

• macOS / Linux (Bash): python3 "$S/generate-cti-docx-hybrid.py" "REPORT.md" "REPORT.json" "REPORT.docx"
• Windows (PowerShell): py "$S\generate-cti-docx-hybrid.py" "REPORT.md" "REPORT.json" "REPORT.docx" — the Store python3 stub will not run; use py or the venv python
• Last resort (no styling/charts): pandoc "REPORT.md" -o "REPORT.docx" --from markdown --to docx --standalone

How the hybrid generator works:

1. Phase 1: pandoc converts the MD file to a base DOCX (preserving ALL narrative content — tables, lists, formatting)
2. Phase 2: python-docx post-processes to add CTI professional styling, prepend cover page + TOC, and inject charts/diagrams from JSON at matching section headings

The MD file is the primary content source. It carries the full narrative (detailed person profiles, infrastructure tables, wallet addresses, corporate structure, legal history, etc.). The JSON file provides structured data for visual elements (charts, diagrams, the ICD-203 × Admiralty confidence matrix, and the campaign heatmaps). Using both together produces a complete report with zero content loss.

Rich hybrid DOCX includes: Cover page titled "CTI REPORT", table of contents, all narrative content from MD (every paragraph, table, list, code block), pie chart (finding types), bar chart (severity), two-axis ICD-203 × Admiralty confidence matrix, campaign heatmaps (malicious-domain registration timeline, domain×indicator correlation, domain×domain possible-relation), timeline chart, entity relationship diagram, network topology diagram, traffic/geo charts, CTI-themed styling (navy headings, styled tables), header/footer with classification and page numbers. No overall exposure score / risk gauge (removed). The PDF (--pdf) is this same DOCX rendered by LibreOffice.

After saving, confirm all files to the user:

📄 Base data bundle saved (always):
   → CTI-REPORT-CASE001-2026-03-30.md
   → CTI-REPORT-CASE001-2026-03-30.json
   → CTI-REPORT-CASE001-2026-03-30.csv
   → IOC-CASE001-2026-03-30.stix.json / .txt / .csv / .jsonl   (indicators & selectors)

📄 Presentation report (your choice) — e.g. when (c) HTML was chosen:
   → CTI-REPORT-CASE001-2026-03-30.html   (interactive — open in any browser, fully offline)

Report Formats

Format	Command	Audience
Interactive HTML	/report html · choose c at the prompt	Everyone — analysts to execs; the primary deliverable
Technical INTSUM	/report	Analysts, security teams
Executive Brief	/report brief	Decision-makers, management
Plain-Language Summary	/brief	Non-technical stakeholders
Legal Evidence Format	/report legal	Attorneys, compliance teams (auto-adds DOCX/PDF)
Journalist Format	/report journalist	Reporters, media
JSON Export	/report json	Downstream tools, pipelines
CSV Export	/report csv	Spreadsheets, databases
IOC / selector bundle	/report ioc	SIEM/TIP ingest, threat-intel sharing
Word document	/report docx · b/d at the prompt	Formal sharing
PDF document	/report pdf · a/d at the prompt	Formal sharing / print (the DOCX rendered by LibreOffice)

Every narrative report always auto-saves the base data bundle (.md + .json + .csv + IOC bundle: .stix.json/.txt/.csv/.jsonl — see Mandatory File Export above), then asks which presentation format to render — (a) PDF · (b) DOCX · (c) HTML · (d) all. /report legal defaults to all; --yolo/guided-auto default to HTML. Machine-only subcommands (json, csv, ioc) emit their native format directly.

Visual Outputs

Type	Command	Format
Subject relationship map	/render entities	ASCII (default) — --mermaid for Mermaid
Chronological timeline	/render timeline	ASCII Gantt
Exposure heatmap	/render risk	ASCII
Network topology	/render network	ASCII

All visual outputs use ASCII box-drawing by default. Mermaid only on explicit --mermaid flag.

Diagram tradecraft: output/visuals/diagram-patterns.md — compile-check a Mermaid diagram before presenting it, and pick the right diagram type per CTI question (attack → sequence, lifecycle → state, handoffs → swimlane, infra → graph).

The interactive HTML report (the primary human-facing deliverable, and the --yolo/guided-auto default) renders all of these as live, explorable visuals — a draggable/zoomable 2D force-directed entity graph, infrastructure topology, an event timeline, and SVG charts (pie/bar/gauge/donut) — alongside the ASCII versions in the .md.

IntelGraph — use it whenever a case has a graph. For any multi-node case (/case, /pipeline, /harness, or a /report with an entity/infra graph), render the publication-quality figure with /graph --render and embed it in the report; do not reserve it for explicit requests. It degrades to a note when the dot/mermaid renderer is missing, so calling it is always safe. But check fidelity first: case_graph.json contains only raw-backed hosts, so KB/reverse-WHOIS-attributed clusters can be absent — if the graph's node count is below the attributed cluster count, hand-scope the figure (Mermaid/DOT of the real clusters) or caption it as the raw-backed subset, rather than auto-rendering a figure that under-represents the finding.

Large infrastructure — show a representative example, point to the full list. A case with hundreds of nodes is unreadable in one figure. Above --max-nodes (default 45) IntelGraph automatically renders a representative subset — the operator anchor, one hub per cluster, and the highest-degree nodes, dropping fan-out infrastructure (nameservers, registrars) first — and stamps the figure title representative subset: N of M nodes. The complete indicator set is never lost: it lives in the IOC bundle (generate-cti-iocs.py → .csv openable in Excel, plus the report .md and the JSON/JSONL exports). Always tell the reader, in the figure caption, that the graph is an example and the full details are in the IOC list. Use --full to force the entire graph, or --split-clusters for one readable figure per cluster.

Connectors

Tool	File	What It Exports
Maltego	connectors/maltego-export.md	GraphML entity graph
Obsidian	connectors/obsidian-setup.md	Linked markdown notes
Notion	connectors/notion-schema.md	Structured database
Intel backend	connectors/intel-backend.md	Optional persistent KB + cross-case correlation via the intel_engine engine (MCP/CLI). Absent → stateless as normal. Enables /backend, /kb, /recall, /binary
ChongLuaDao ⭐	connectors/chongluadao-api.md	First-party premium connector (scripts/cld/cld_api.py) — IoC/denylist/breach/data-leak/AI/feeds. Enables /cld; upgrades /scam-check,/threat-check,/phone,/breach-deep,/email-deep,/email-hygiene,/vuln-check,/impersonate. Needs CHONGLUADAO_API_KEY

---

9. Skill Tiers & Customization

Reference: experience/skill-tiers.md, experience/layered-detail.md

Tiers

Tier	Command	What Changes
Novice	/novice	Jargon removed, steps explained, glossary auto-linked
Practitioner	(default)	Standard output, moderate detail
Specialist	/novice off	Full technical detail, raw findings, internal signals

Switch tiers at any point — output adapts immediately.

Guided Flows

experience/guided-flows/ contains step-by-step interactive flows:

• person-investigation.md — Full guided person case
• domain-reconnaissance.md — Guided domain sweep
• email-investigation.md — Guided email tracing
• rapid-case.md — 10-minute abbreviated sweep

Activate: /flow person · /flow domain · /flow email · /flow quick

Case Templates

experience/case-templates/ contains pre-built starting configurations:

• due-diligence.md — Corporate partner vetting
• security-audit.md — Organization exposure audit
• background-check.md — Individual background research

Activate: /template run [name]

---

10. Ethics & Boundaries

This skill operates strictly within publicly available information.

Permitted

• Journalists verifying facts about public figures or institutions
• Security professionals auditing their own organization's exposure
• Individuals reviewing their own digital footprint
• Corporate due diligence on business partners
• Academic research and educational demonstrations

Prohibited

• Stalking, harassment, or doxing of any individual
• Accessing accounts or systems without authorization
• Social engineering or deception campaigns
• Any activity violating applicable law

Ethical reminders are issued automatically when the investigation approaches sensitive territory. Public data is not a license to cause harm.

---

11. Autonomous Mode (--yolo)

Append --yolo to any command or activate at session start.

What changes:

• No clarifying questions — analyst infers context and proceeds
• No confirmation prompts — scope expands automatically on new discoveries
• Guided flows skip Q&A — reasonable defaults applied
• Both /report and /brief generated without asking — the presentation-format prompt and the HTML export confirmation are skipped; defaults to HTML (Blueprint in Auto mode) on top of the always-saved base data bundle (.md + .json + .csv + IOC .stix.json/.txt/.csv/.jsonl)

What stays the same:

• Ethics and legal boundaries — always enforced
• Trust scores on every finding
• Source citations on every claim
• /validate and /coverage run before final delivery

Activate per-command: /case target.com --yolo Activate for session: /cti-expert --yolo

---

12. Architecture Reference

cti-expert/
├── SKILL.md                    This file
├── README.md                   User-facing overview
│
├── engine/                     Case data model and state management
│   ├── case-schema.json        Subject and finding data structures
│   ├── subject-registry.md     How subjects are tracked and versioned
│   ├── finding-framework.md    Finding lifecycle, trust scores, evidence chains
│   ├── pivot-orchestration.md  Recursive spider-map pivot engine (BFS loop, edge matrix, gating)
│   ├── workspace-format.md     Workspace serialization spec
│   ├── workspace-manager.md    Save/open/list workspace logic
│   └── conflict-resolver.md    CONTESTED finding resolution
│
├── analysis/                   Pattern detection and intelligence engines
│   ├── deviation-detector.md   Behavioral anomaly detection
│   ├── auto-branch-rules.md    Automatic pivot trigger rules
│   ├── drift-monitor.md        Subject state change tracking
│   ├── cross-reference-engine.md Shared identifier detection across subjects
│   ├── archive-explorer.md     Wayback Machine integration and diff
│   ├── signature-catalog.md    Behavioral pattern library
│   ├── exposure-model.md       Exposure score calculation framework
│   ├── risk-trend-tracker.md   Temporal risk score tracking (/drift)
│   ├── pattern-library.md      Username, email, bot detection patterns
│   └── weight-engine.md        Finding aggregation and confidence weighting
│
├── techniques/                 Collection techniques and module specs
│   ├── fx-metadata-parsing.md  EXIF, headers, document metadata
│   ├── fx-image-verification.md Image authenticity and provenance
│   ├── fx-breach-discovery.md  Breach database and paste site methods
│   ├── fx-geolocation.md       GPS, W3W, Plus Codes, Street View
│   ├── fx-social-topology.md   Social graph construction and topology
│   ├── fx-email-header-analysis.md Header analysis, SPF/DKIM
│   ├── fx-document-forensics.md Document forensics and extraction
│   ├── fx-http-fingerprint.md  HTTP fingerprinting and signatures
│   ├── fx-leak-monitoring.md   Leak and breach monitoring
│   ├── username-osint.md       Platform enumeration (3000+)
│   ├── phone-osint.md          Phone carrier/VoIP/spam lookup
│   ├── email-osint.md          Deep email investigation
│   ├── threat-intel.md         Threat intelligence free lookups
│   ├── web-traffic-analysis.md Traffic estimation methods
│   ├── secret-scanning.md      Credential/secret detection
│   ├── github-osint.md         GitHub profiles, repos, code, commits, forks
│   ├── domain-advanced.md      Subdomain enumeration methods
│   ├── social-media-platforms.md Platform-specific techniques
│   ├── advanced-geolocation-techniques.md Overpass Turbo, road signs, reflected text
│   ├── wifi-ssid-osint.md      WiFi SSID/BSSID geolocation via Wigle.net
│   ├── web-dns-forensics.md    DNS, GitHub, Telegram, WHOIS
│   ├── fx-visitor-intelligence.md Visitor stats, tech stack, geo analysis
│   ├── scam-check.md           Phishing/scam domain verification
│   ├── cloud-audit.md          Cloud infrastructure security audit
│   ├── microsoft-tenant-recon.md M365/Azure tenant enumeration
│   ├── china-recon.md          ICP filings, PRC registries, CN cyberspace engines, CJK variants
│   ├── fiat-payment-osint.md   IBAN/BIC/bank accounts as selectors, VN-SEA rails
│   ├── fx-edge-appliance-recon.md Edge/VPN appliance fingerprint → KEV/CVE catalog + port-risk matrix
│   ├── fx-saas-identity-recon.md SaaS tenancy + IdP fingerprint + API/GraphQL/spec discovery
│   ├── dependency-audit.md     Supply chain security audit
│   ├── disk-forensics.md       Digital evidence analysis
│   ├── incident-triage.md      Security incident response
│   ├── owasp-audit.md          OWASP Top 10 source code audit
│   ├── prompt-injection-audit.md AI/LLM security audit
│   ├── stealer-log-analysis.md Infostealer-log triage, actor attribution & IOC extraction
│   ├── agent-browser.md        Interactive browser collection & evidence capture (vercel-labs/agent-browser)
│   └── ioc-export.md           IOC export (STIX 2.1, flat list)
│
├── experience/                 UX, tiers, and guided flows
│   ├── skill-tiers.md          Novice/Practitioner/Specialist spec
│   ├── layered-detail.md       Progressive disclosure rules
│   ├── guidance-system.md      How guided flows work
│   ├── case-progress.md        Progress tracking logic
│   ├── guided-flows/           Interactive step-by-step flows
│   │   ├── flow-person-lookup.md Person investigation guided flow
│   │   ├── flow-domain-sweep.md Domain reconnaissance guided flow
│   │   └── flow-image-check.md Image verification guided flow
│   ├── case-templates/         Pre-built case configurations
│   │   ├── tpl-index.md        Template index and descriptions
│   │   ├── tpl-due-diligence.md Due diligence case template
│   │   ├── tpl-security-review.md Security audit case template
│   │   └── tpl-background-check.md Background check case template
│   ├── tutorial.md             First-time onboarding guide (/onboard)
│   ├── feedback-system.md      Investigation quality feedback loops
│   └── accessibility/          Glossary and accessibility settings
│       ├── glossary.md         OSINT term glossary
│       └── accessible-mode.md  Low-jargon mode settings
│
├── output/                     Report and visualization specs
│   ├── reports/                Report format templates
│   │   ├── format-catalog.md   Report format specifications
│   │   ├── leadership-brief-template.md Executive brief template
│   │   ├── export-specs.md     Export format specifications
│   │   └── citation-guide.md   Source citation standards
│   └── visuals/                Chart and visualization specs
│       ├── chart-templates.md  Chart rendering templates
│       ├── ui-components.md    UI component library
│       ├── render-engine.md    ASCII render engine spec
│       ├── case-dashboard.md   Dashboard layout spec
│       ├── attack-path-diagram.md  Attack path flow visualization (/render threat-path)
│       └── attack-surface-map.md   Attack surface exposure map (/render attack-surface)
│
├── scripts/                    Cross-platform install + HTML / IOC / DOCX report generation
│   ├── platform-setup.md            Cross-platform reference: OS detection, uv-first install matrix, gotchas
│   ├── install.ps1                  Windows installer (uv-first: uv venv/pip/tool; winget + pip/pipx fallback)
│   ├── install.sh                   macOS/Linux/Git-Bash/WSL installer (uv-first; brew/apt + pip/pipx fallback)
│   ├── stealer_log_parse.py         Infostealer-log analyzer — attribution, profiling, IOCs (PEP 723 / `uv run`, zero-dep)
│   ├── iban_analyze.py              IBAN validate + decompose (ISO 13616/7064) → bank code, risk signals (PEP 723, zero-dep)
│   ├── phish_domain_survival.py     Phishing-domain registration/DNS profiling → maliciously-registered vs compromised + survival outlook (PEP 723, zero-dep)
│   ├── clickfix_detect.py           ClickFix / PasteJacking clipboard-hijack detector + `-enc` decode → C2 IOCs (PEP 723, zero-dep)
│   ├── html_visibility_analysis.py  Visibility-aware HTML analysis — hidden credential forms / off-origin links / off-screen text (PEP 723, zero-dep)
│   ├── apk_permission_scope.py      APK permission-scope risk scoring (BinaryPivot ext) — combo-based on-device-fraud capability; UTF-8/UTF-16 AXML decode (PEP 723, zero-dep)
│   ├── kit_template_fingerprint.py  Phishing kit/template structural fingerprint + similarity + commodity-trap grading (PEP 723, zero-dep)
│   ├── render_confirm.py            Renderer-level confirmation — reconciles static + rendered ClickFix/visibility evidence (PEP 723; optional Playwright/agent-browser)
│   ├── phishtrace_features.py       PhishTrace dynamic-feature characterization from a runtime trace → verdict + exfil IOCs (PEP 723, zero-dep)
│   ├── cld/cld_api.py               ChongLuaDao premium API client — IoC / denylist / breach + full data-leak module (async jobs) / AI URL analysis / STIX-MISP feeds (PEP 723 / `uv run`, zero-dep, X-API-Key)
│   ├── redact.py                    Reversible PII redaction — stable placeholders + exportable map; md/json/csv (PEP 723, zero-dep)
│   ├── cti-report-template.html     PRIMARY: interactive HTML report template — self-contained & OFFLINE (charts + 2D entity graph + topology + timeline + indicator panel + search; dark/light + print-to-PDF)
│   ├── generate-cti-html.py         HTML report generator — injects the report JSON into the template + builds the embedded Archify Blueprint (`CTI_ARCHIFY=1|force|0`; PEP 723 / `uv run`, zero-dep, self-heals UTF-8)
│   ├── cti_archify.py               CTI case → Archify `architecture` IR converter (drives the report's inline Blueprint diagram; folds dense estates to apex level + hub-and-spoke placement; stdlib-only, reuses the engine's PSL reducer when present)
│   ├── vendor/archify/              Vendored zero-dep Archify render subset (v2.16.0, MIT) — renders the Blueprint diagram; needs Node.js (see vendor/archify/VENDOR.md)
│   ├── cti_diagram_design.py        CTI case → Diagram Design editorial SVG (entity + topology); cairosvg rasterizer for DOCX (stdlib SVG)
│   ├── cti_cloud_arch.py            CTI case → Diagram AI Generator cloud figure (diagrams+graphviz, cloud-gated, opt-in; isolated uv subprocess)
│   ├── vendor/diagram-design/       Vendored Diagram Design style-guide + self_check + LICENSE (MIT) — editorial token/rule source
│   ├── vendor/diagram-ai-generator/ LICENSE + VENDOR note (MIT) — cloud figure applies its spec/approach via the `diagrams` library
│   ├── generate-cti-iocs.py         Comprehensive IOC/selector exporter → STIX 2.1 / flat / CSV (network IOCs + contacts + identities + social/messaging + wallets + attribution; PEP 723 / `uv run`, zero-dep)
│   ├── generate-cti-docx-hybrid.py  Hybrid MD+JSON DOCX generator — on request / `/report legal` (PEP 723 / `uv run`; self-heals UTF-8 + pandoc)
│   ├── generate-cti-docx.py         Fallback: JSON-only generator (PEP 723 / `uv run`)
│   ├── cti_docx_postprocess.py      Post-processing: styling, chart injection, cover page
│   ├── cti_docx_charts.py           Chart rendering (pie, bar, timeline, traffic, geo, ICD-203×Admiralty confidence matrix)
│   ├── cti_docx_heatmaps.py         Campaign heatmaps (registration timeline, domain×indicator correlation, domain×domain relation)
│   ├── cti_docx_pdf.py              DOCX→PDF via LibreOffice (document-style PDF; the --pdf path)
│   ├── cti_docx_diagrams.py         Entity relationship + network topology diagrams
│   ├── cti_docx_sections.py         Report section formatting (used by JSON-only generator)
│   ├── cti_docx_styles.py           Document styling, colors, cover page, header/footer
│   ├── requirements.txt             Python dependencies
│   └── sample-cti-report-data.json  Example JSON report data
│
├── workflows/                  Professional workflow guides
│   ├── wf-journalist.md
│   ├── wf-hr-screening.md
│   ├── wf-threat-analyst.md
│   └── wf-private-investigator.md
│
├── handbook/                   Reference material
│   ├── operator-queries.md     Search operator catalog
│   ├── quick-report.md         Rapid reporting reference
│   ├── discovery-paths.md      Per-target-type search paths
│   ├── report-template.md      INTSUM format specification
│   ├── admin-endpoint-indicators.md  Admin-panel / sensitive-endpoint detection vocab & rules
│   ├── analytic-standards.md    Likelihood bands, 5W1H coverage overlay, ACH (competing hypotheses)
│   ├── aam-actor-modeling.md   AAM actor-state overlay for /threat-model — OODA faces + Mirror/Twin/Opposite/Lever (eCrime 2026 "Modeling Adversaries Through Chaos")
│   ├── pivot-artifacts.md      Pivot-artifact catalog (favicon, trackers, wallets, certs…)
│   ├── pivot-services.md       Reverse-lookup engines per artifact — hash algo, cost, API/key notes
│   ├── api-keys.md             Premium/pro API key management and unlocks
│   └── tool-cascade-reference.md Tool priority and fallback chains
│
├── guides/                     Worked case walkthroughs
│   └── walkthroughs/           Step-by-step investigation examples
│       ├── walkthrough-person-lookup.md
│       ├── walkthrough-domain-sweep.md
│       └── walkthrough-username-trace.md
│
├── validation/                 Quality assurance
│   ├── coverage-matrix.md      Investigation area coverage tracking
│   ├── quality-scoring.md      Scoring methodology
│   └── verification-checklist.md Finding verification steps
│
└── connectors/                 External tool integrations
    ├── maltego-export.md
    ├── obsidian-setup.md
    ├── notion-schema.md
    ├── intel-backend.md
    └── chongluadao-api.md

---

Technique Activation Matrix

Which techniques activate per target type in a /case run:

Technique	Person	Domain	Org	Username	Email	IP
/sweep	✅	✅	✅	✅	✅	✅
/query	✅	✅	✅	✅	✅	✅
/username	✅	—	✅*	✅	—	—
/email-deep	✅	—	✅*	—	✅	—
/phone	✅	—	✅*	—	—	—
/breach-deep (LeakCheck + HudsonRock)	✅	✅	✅	✅	✅	✅
/intelx (breach dumps · infostealer logs · pastes · darknet · historic WHOIS; --phonebook apex inventory)	—	✅	✅*	—	✅	✅
/subdomain	—	✅	✅	—	—	—
/traffic	—	✅	✅	—	—	—
/threat-check	—	✅	✅	—	—	✅
/secrets	—	✅	✅	✅	—	—
/github-osint	✅*	✅	✅	✅	✅*	—
/scam-check	—	✅	✅	—	—	—
phish-domain-survival (registration/DNS class)	—	✅	✅	—	—	—
clickfix-detect (page clipboard-hijack)	—	✅	✅	—	—	—
html-visibility (hidden-content evasion)	—	✅	✅	—	—	—
/branch	✅	✅	✅	✅	✅	✅
/gdoc	—	✅	✅	—	—	—
/sharelink	✅	—	✅	✅	✅	—

<!-- dork-integration:phase-05 start -->

| /dork-sweep | ✅ | ✅ | ✅ | ✅ | ✅ | ✅* | | /docleak | ✅ | ✅ | ✅ | ✅* | — | — |

<!-- dork-integration:phase-05 end -->

| Social media platforms | ✅ | — | ✅ | ✅ | — | — | | Metadata forensics | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | | Photo verification | ✅ | — | ✅* | ✅ | — | — | | Network analysis | — | ✅ | ✅ | — | — | ✅ | | Advanced geolocation | ✅ | — | — | ✅ | — | — | | Web & DNS forensics | — | ✅ | ✅ | — | ✅ | ✅ | | /timeline | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | | /exposure | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | | /threat-model | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | | /wifi (SSID/BSSID) | ✅ | ✅ | ✅ | — | — | ✅ | | Visitor intelligence | — | ✅ | ✅ | — | — | ✅ | | Cloud audit | — | ✅ | ✅ | — | — | ✅ | | MSFTRecon (M365/Azure tenant) | — | ✅ | ✅ | — | — | — | | /icp (ICP filing → PRC entity) | — | ✅ | ✅ | — | — | ✅ | | /cn-corp (PRC registry chain) | ✅* | ✅ | ✅ | — | — | — | | /iban (payment-rail selector) | ✅ | ✅ | ✅ | ✅ | ✅ | — | | /hash-id (hash typing) | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | | Dependency audit | — | ✅ | ✅ | — | — | — | | Disk forensics | — | — | — | — | — | — | | Incident triage | — | ✅ | ✅ | — | — | ✅ | | OWASP audit | — | ✅ | ✅ | — | — | — | | Prompt injection audit | — | ✅ | ✅ | — | — | — | | /snapshots | — | ✅ | ✅ | — | — | ✅ | | Archive IOC harvest (wayback_harvest.py) | — | ✅ | ✅ | — | — | — | | /diff | — | ✅ | ✅ | — | — | ✅ | | /drift | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | | /render threat-path | — | ✅ | ✅ | — | — | ✅ | | /render attack-surface | — | ✅ | ✅ | — | — | ✅ | | /blind-spots | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | | /source-check | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | | /report ioc | — | ✅ | ✅ | — | — | ✅ | | /report + /brief | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | | Shodan InternetDB (ports/tags/vulns) | — | ✅ | ✅ | — | — | ✅ | | GreyNoise Community (noise/threat class) | — | ✅ | ✅ | — | — | ✅ | | URLScan.io passive (scan history) | — | ✅ | ✅ | — | — | — | | Disposable email check (kickbox) | ✅ | — | ✅* | — | ✅ | — | | URLhaus (malware URL hosting) | — | ✅ | ✅ | — | — | ✅ | | ThreatFox (IOC/C2 lookup) | — | ✅ | ✅ | — | — | ✅ | | MalwareBazaar (hash → malware family) | — | — | — | — | — | — | | ipwho.is (geo + ASN + ISP) | — | ✅ | ✅ | — | — | ✅ | | DMARC/SPF/DKIM check (DNS) | — | ✅ | ✅ | — | ✅ | — |

✅* — runs for discovered key personnel within the organization MalwareBazaar — reached via /hash [value], but only after /hash-id has typed the value

The five v2.6 commands in /case

Four fire automatically as part of the standard pipeline — no flag needed. /redact is opt-in.

Command	Fires	Phase
/icp	Unconditionally for every domain/URL/org target, and on an IP target's resolved hostname. One cheap passive lookup — and a missing filing on CN-hosted infrastructure is itself a finding, so there is no CN-nexus precondition. A discovered licence serial re-enters the pivot loop as its own node and reverse-searches to sibling domains (same operator, HIGH).	Acquire
/cn-corp	Automatically on any company name or USCC the case surfaces — from the ICP filing, WHOIS registrant org, page footer, or personnel discovery. Runs the GSXT → aggregator → 信用中国 chain; officers/shareholders/subsidiaries re-enter the loop as new nodes.	Acquire → Enrich
/iban	Automatically on any payment detail that surfaces — page DOM, Wayback harvest, victim statement, invoice, stealer log. Validation is arithmetic on a string, so it is free and touches nothing. Valid accounts become financial/iban IOCs; a checksum-invalid account on a payment page is logged as a behavioural finding instead.	Acquire → Enrich
/hash-id	Automatically on every discovered hash, and always before /hash. Decides file hash (→ MalwareBazaar/VT) vs credential material (→ /breach-deep, never a public service). 64-hex is queued for both the cert-fingerprint and file-hash readings, since it is genuinely ambiguous.	Acquire → Enrich
/redact	Opt-in, not automatic — pass --redact (or run /redact later). Emits REPORT.redacted.{md,json,csv} + REPORT.map.json alongside the unredacted set. A redacted report is a weaker artifact, so producing one stays a deliberate choice.	Deliver

Registries needing mainland egress (TianYanCha/QCC/Aiqicha) are logged as collection gaps, never blockers (techniques/china-recon.md §7). Non-IBAN rails (VN transfer, VietQR/NAPAS BIN, card BIN, e-wallet) follow techniques/fiat-payment-osint.md §4.

Flags: --no-cn skips /icp+/cn-corp (both run by default). --redact adds the redacted variant (off by default). --no-harness disables the non-convergence auto-escalation to the harness deepening loop (on by default; see the deep-layer note above).

Recursive pivot orchestration (the spider-map). /case is not a one-pass collector — it runs a recursive BFS pivot engine: every discovered identifier becomes a new seed, and the relationship graph expands hop by hop until the frontier is exhausted or a budget cap is hit. The state machine — identifier typing, dedup / cycle prevention, per-node depth, the identifier→pivot edge matrix, confidence gating, and per-depth checkpoints — is scripts/pivot_orchestrator.py; the full spec is engine/pivot-orchestration.md. The orchestrator plans and tracks; the agent executes each hop's technique commands and feeds results back via --ingest.

• Defaults: posture=active (may fetch/scan targets; still passive-first for hostile infra), reach=exhaustive (pivot till the frontier empties), autonomy=auto — the loop runs to closure unattended, no per-depth approval prompts. Depth summaries are still printed as they happen, so the expansion stays auditable. Safety caps: max_nodes=500, max_depth=6.
• Gating (reuses analysis/auto-branch-rules.md): exact-match links (≥95% — shared GA ID / cert / favicon / registrant email, handle exact-match) auto-pursue unbounded; HIGH/MEDIUM capped per type; LOW held unless corroborated; PII (person/phone) auto-expands by default (hold with --authorization unconfirmed); visited nodes and past-depth-cap nodes suppressed (loop-safe).
• Example hops: an email auto-pivots via reverse-WHOIS→domains, /breach-deep, /github-osint; a domain discovered from a person (high-confidence link) continues via /webpivot+wayback_harvest+whois_enrich+cert_pivot+subdomains; a shared GA ID reverse-pivots to sibling domains; a discovered document (.pdf/office) or image (.jpg/.png) is itself a node — metadata/authorship→person/email/org and EXIF GPS + reverse-image/face→person/domain (face matches held pending corroboration) — each new node re-enters the loop.
  • Media content is analyzed, not just fingerprinted. An image/document node is vision-analyzed (multix gemini — OCR, sign/landmark/logo read, structured extraction) and A/V is FFmpeg-preprocessed (keyframes + audio track) then transcribed, before it is treated as terminal — see §Media Evidence Analysis. Extracted text / GPS / entities re-enter the loop as new seeds; face matches stay held pending corroboration.
• Control flags (all narrowing — the defaults are already maximal): /case <t> --passive|--passive-first, --reach balanced|focused, --checkpoint (pause for approval after each depth level), --depth N, --budget N, --authorization unconfirmed (re-hold PII), --no-cn, --no-harness (skip harness auto-escalation); --redact opts in to the redacted variant.
• Termination → if the loop converged (frontier exhausted): emit edges → graph_build.py → interactive HTML force-graph + topology + timeline; findings/indicators roll into the auto-saved report + IOC bundle. If it stopped on a cap (max_nodes/max_depth/--budget) with the frontier still open — i.e. not converged — /case auto-escalates to the harness deepening loop (posture-gated, keyless-first; --no-harness opts out) before Deliver, per the deep-layer note above.

Legacy one-hop note (still true, now a subset of the loop): if /sweep on a domain finds an email, /email-deep and /breach-deep trigger on it automatically.

Leak / breach / infostealer auto-fire in /case:

• Every discovered email → /breach-deep + /email-deep (LeakCheck·HudsonRock·CLD), then /intelx <email> for breach dumps, infostealer logs, pastes and darknet mirrors (logs-first pass is ~50% keyless). A stealer-log hit that ties the credential to a machine also holding the registrar/hosting/CMS login is direct attribution, not a self-declared WHOIS assertion — the single highest-value deanonymization pivot.
• Every discovered domain/apex → /intelx --phonebook <apex> to inventory every email, subdomain and URL IntelX has seen under it (recovers selectors the live site scrubbed).
• Every discovered phone / wallet / IBAN → /intelx <selector> (strong selectors only; a name is refused locally and still costs a unit).
• A local infostealer-log folder → /stealer-log for family attribution, victim-vs-operator profiling and IOC extraction.
• These are the leak/breach legs the deterministic infra pipeline does not run; they are required for a full /cti, gated only by --quick (skips) and --passive (keyless corpora only — IntelX/Wayback never touch the target). Results (new emails, machines, credentials, siblings) re-enter the recursive pivot loop as seeds. State credits spent.
• Pipeline auto-fire (deterministic path): intel.py pipeline open now appends --intelx to the collector whenever an IntelX key is present and the case is not a no_spend posture; pipeline loop appends it only under --full — the default loop stays free-only and never spends an IntelX search on its own.

GitHub OSINT auto-fire in /case:

• Domain/Org target → run /github-osint on the org name, primary domain, discovered GitHub orgs/repos, and developer-platform hits from /query or /dork-sweep.
• Username target → run /github-osint directly when the handle has a GitHub profile or GitHub search hit.
• Person target → run /github-osint only after discovering a likely GitHub handle, commit email, repo author, or developer profile link.
• Email target → run /github-osint only after discovering commit attribution, GitHub noreply patterns, profile links, or repo references.
• Every /github-osint run executes the committer harvest first (github_harvest / intel.py github <target>), on the user or org AND on each discovered repo, before any code search; the harvested e-mails and former logins re-enter the pivot loop as seeds (e-mail → /breach-deep·/intelx·reverse-WHOIS; login → /username).
• Results feed into /secrets, /branch, /timeline, /crossref, /exposure, and final /report automatically.

<!-- dork-integration:phase-05 start -->

✅* dork coverage notes: /dork-sweep on IP runs against reverse-DNS hostname once resolved (graceful skip if no rDNS); /docleak on Username targets document-author/uploader fields on scribd, slideshare, academia.edu, researchgate.

Dork auto-fire matrix — every /case target type gains coverage:

• Person → /dork-sweep --telegram --docs + /docleak on full name
• Domain → /dork-sweep --filetype --docs + /docleak on domain + org name
• Org → /dork-sweep --filetype --docs --telegram + /docleak on org + primary domain
• Username → /dork-sweep --telegram --docs + /docleak (author-angle)
• Email → /dork-sweep --telegram --docs on email + @domain
• IP → /dork-sweep on rDNS-resolved hostname (skipped if no rDNS)
• Domain/Org → GrayHatWarfare open-bucket check (/secrets <label> — the GrayHatWarfare layer; keyless dork fallback) — emitted by the pipeline as a per-apex exposure lead and rendered in the assessment's Exposure (leak surface — not attribution) section. A bucket carrying the brand label is a thing to READ, never a same-operator pivot and never a frontier seed.

Adaptive fan-out: discovered emails → Telegram dork; discovered personnel → /docleak; discovered subdomains → filetype dork; discovered usernames → Telegram + doc sweep; discovered IPs → rDNS → dork-sweep.

<!-- dork-integration:phase-05 end -->

When /case or /sweep runs on a Domain or Org target, it inspects the MX record and SPF TXT record. If MX ends in protection.outlook.com OR SPF contains spf.protection.outlook.com, /msftrecon auto-fires as part of the Acquire phase. Results feed back into the subject registry as infrastructure findings (tenant ID, federation type, MDI presence) and into /exposure scoring.

/case pipeline walkthrough (M365-hosted Domain/Org): (a) standard DNS/WHOIS/subdomain/traffic/scam-check/breach-deep checks run first, (b) if M365 indicators present → /msftrecon fires automatically with no extra flag, (c) tenant ID discovered becomes a pivot for /branch in Enrich phase (search other domains under the same tenant). No user intervention required.

Parallel enrichment (3+ subjects): When Acquire discovers 3+ subjects, enrichment commands fan out in parallel via AgentFlow DAG orchestration. Each subject's enrichment runs independently, results merge with dedup before Assess phase. Disable with --sequential flag. See techniques/agentflow-enrichment.md.

---

Tool Priority & Fallback

Primary interactive collector: agent-browser (vercel-labs) — a fast native-Rust CDP browser that returns accessibility-tree snapshots (@eN element refs) + screenshots; no API key for core automation; cross-platform; also an MCP server. Full how-to + per-command usage in techniques/agent-browser.md. It is complementary to Scrapling, not in conflict (different ecosystems — Rust binary via npm/brew/cargo vs Python via pip — each manages its own browser): use agent-browser to interact with and witness a page (logins, clicks, screenshots, JS render) and Scrapling to fetch and parse pages programmatically.

1. Check agent-browser first (agent-browser --version; load its guide via agent-browser skills get core; install per the auto-install policy if missing)
2. Use agent-browser for: screenshot evidence, logins/interactive UI, JS-rendered/SPA pages, complex multi-step browser flows
3. Use Scrapling DynamicFetcher for: JS-heavy sites, SPA content, auto-escalation from static (programmatic)
4. Use Scrapling StealthyFetcher for: anti-bot bypass, Cloudflare-protected targets
5. Use Scrapling Fetcher for: fast static page collection, HTML parsing (~2ms)
6. Fall back to web search → web fetch → direct curl — no investigation blockers
7. Tag each finding with collection method: [browser] · [scrapling-dynamic] · [scrapling-stealth] · [scrapling-static] · [vision] · [vision-local] · [media] · [search] · [fetch] · [manual] · [whois-lib] · [whois-cli] · [whois-api]

Media Evidence Analysis — vision (multix/Gemini) + video (FFmpeg)

Screenshots, photos, scanned documents, and audio/video are nodes, not dead ends — analyze the content, don't just fingerprint the file. These tools are native to this skill (no AgentKit / sibling-skill dependency, so a bare clone works): the vision layer is the standalone multix CLI run via npx, the media layer is ffmpeg/imagemagick, provisioned by the Tool Auto-Install Policy. Keyless by default: with no GEMINI_API_KEY, OCR/sign-read/transcription fall back to the host agent's own multimodal read + tesseract OCR + local Whisper (whisper-ctranslate2) — a key only upgrades quality/scene-reading. Full workflow: techniques/media-vision-analysis.md. A missing key/binary logs a collection gap, never blocks.

• Images / screenshots / documents → vision (multix gemini, Gemini; reads GEMINI_API_KEY from env or the skill's $SKILL_DIR/.env). Read text a parser can't (OCR a Telegram-bio screenshot), read signage / landmarks / reflected text for geolocation, describe logos and faces for correlation, or pull structured fields:
  • npx --yes --prefer-online --package=@mrgoonie/multix@latest -- multix gemini analyze --files EVIDENCE.png --prompt "OCR all text; describe location cues, signage, landmarks, logos, faces" --format markdown --output vision.md
  • npx --yes --prefer-online --package=@mrgoonie/multix@latest -- multix gemini extract --files EVIDENCE.png --prompt "Extract handles, emails, phones, wallet/IBAN, org names as JSON" --format json --output vision.json
• Video / audio → preprocess with FFmpeg, then vision. Extract keyframes and the audio track first, then analyze/transcribe them:
  • Keyframes → ffmpeg -i clip.mp4 -vf "fps=1/5" frame_%03d.png, then feed each frame to multix gemini analyze
  • Audio → ffmpeg -i clip.mp4 -vn -c:a copy audio.m4a, then ... multix gemini transcribe --files audio.m4a --prompt "Transcript with timestamps" --format markdown --output transcript.md
• Feedback into the pivot loop. OCR'd handles/emails/phones, extracted GPS/landmarks, transcribed names, and read logos become new seeds and re-enter the Enrich loop (§Recursive pivot loop). Face matches stay held pending corroboration.
• Tags: mark such findings [vision] (multix) and [media] (ffmpeg/imagemagick preprocessing).
• No Gemini key? Keyless fallback (own models). OCR + sign/landmark/logo/face read → the host agent's own vision (read the image in-session — Claude Code/Codex are multimodal) plus tesseract for dense OCR; A/V transcription → local Whisper (whisper-ctranslate2, offline) after the same ffmpeg audio extraction. A GEMINI_API_KEY only upgrades quality. Tag these findings [vision-local].

---

Tool Auto-Install Policy

Before running any CLI tool, detect the OS, check whether the tool exists, and if not, install it automatically and silently using the method correct for that OS. This ensures zero-friction investigation on Windows, macOS, and Linux alike — the analyst never stalls on a missing dependency and never runs a Linux-only command on Windows.

Full cross-platform reference: scripts/platform-setup.md — OS detection, $PY/shell conventions, package managers, the complete per-tool × per-OS install matrix, and known gotchas. Consult it whenever this summary is not enough.

Step 0 — Detect the platform (once per session)

Determine the OS before running anything, and cache it for the rest of the session. In Claude Code the environment block already reports it (e.g. Platform: win32 → Windows). Otherwise probe: PowerShell $IsWindows/$IsMacOS, or Bash uname -s (Darwin=macOS, Linux=Linux, MINGW*/MSYS*/CYGWIN*=Windows/Git Bash). Then fix these conventions:

	Windows	macOS / Linux
Shell	PowerShell	Bash
Python runner ($PY)	uv run (preferred) · else venv …\.venv\Scripts\python.exe · else py	uv run (preferred) · else venv …/.venv/bin/python3 · else python3
"exists?" check	Get-Command <tool> -ErrorAction SilentlyContinue (or where.exe <tool>)	command -v <tool>
System pkg manager	winget (→ choco/scoop)	brew (macOS) · sudo apt/dnf/pacman (Linux)

On Windows, python3/python in the Bash tool is often a non-functional Microsoft Store stub. Prefer uv (it brings its own Python and sidesteps the stub); otherwise use py via PowerShell.

Step 0.5 — Ensure uv (the primary Python toolchain)

uv is the preferred way to install and run everything Python in this skill. It is a single fast, cross-platform tool that replaces pip, pipx, venv, and pyenv, manages its own Python (so the Windows Store-stub problem disappears), and resolves script dependencies on the fly. Using uv also collapses the per-OS split for Python tools — the same command works on Windows, macOS, and Linux.

• Check: uv --version
• Install if missing:
  • Windows: winget install --id astral-sh.uv — or powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex"
  • macOS / Linux: curl -LsSf https://astral.sh/uv/install.sh | sh — or brew install uv
  • Any OS that already has pip: python -m pip install uv

If uv genuinely cannot be installed, fall back to the per-OS pip/pipx/venv path — nothing here hard-requires uv.

Auto-Install Protocol

1. Check: OS-correct existence test from the table above (or <$PY> -c "import <module>" for Python modules)
2. Install: If missing, run the OS-correct install command (see dispatch below / platform-setup.md)
3. Verify: Confirm the tool resolves before proceeding — re-check; on Windows a fresh install may need a new shell or a probe of its install dir
4. Log: Note [auto-installed] in the finding's collection method tag
5. Continue: Proceed with the investigation — never block on tool availability

Install Dispatch by Category

Python tools — uv, identical on every OS (the big win: no per-OS split). CLIs use uv tool; libraries go into the skill venv via uv pip. No-uv fallback in the last column.

Python tool(s)	Install (any OS, uv)	No-uv fallback
CLIs — maigret, sherlock-project, holehe, h8mail, waymore, xeuledoc	uv tool install <pkg>	pipx install <pkg>
Libraries — cloudscraper, oletools, whoisdomain, scrapling	uv pip install --python <venv> <pkg>	<$PY> -m pip install <pkg>
Scrapling headless	uv tool install "scrapling[fetchers]" then scrapling install	<$PY> -m pip install "scrapling[fetchers]" then scrapling install
AgentFlow	uv pip install --python <venv> --no-deps agentflow	<$PY> -m pip install --no-deps agentflow
Git-only — theHarvester, msftrecon, blackbird, sharetrace	uv tool install "git+https://…/theHarvester.git" · uv pip install "git+https://…/msftrecon.git" · clone + uv pip install -r requirements.txt	pipx install "git+https://…" · clone + <$PY> -m pip install -r requirements.txt
Run a generator script	uv run <script.py> ARGS (deps auto via inline metadata)	<$PY> <script.py> ARGS

<$PY> = py (Windows) / python3 (macOS/Linux), or the venv python. On PEP-668 Linux add --break-system-packages to the pip fallback.

System binaries — OS package manager (uv does not manage these):

Tool(s)	Windows	macOS	Linux
git, gh, jq, exiftool, pandoc, poppler/pdfinfo, qpdf, whois	winget install <Id>	brew install <pkg>	sudo apt install -y <pkg>
Go toolchain	winget install GoLang.Go	brew install go	sudo apt install -y golang
mat2 (metadata strip)	n/a → exiftool -all= -overwrite_original <file>	brew install mat2	sudo apt install -y mat2
agent-browser (interactive browser)	npm i -g agent-browser or cargo install agent-browser → agent-browser install	brew install agent-browser → agent-browser install	npm i -g agent-browser (or cargo install) → agent-browser install
ffmpeg, imagemagick (magick), rmbg-cli — A/V + image evidence preprocessing (keyframes, audio extract, resize, bg-removal)	winget install Gyan.FFmpeg ImageMagick.ImageMagick + npm i -g rmbg-cli	brew install ffmpeg imagemagick + npm i -g rmbg-cli	sudo apt install -y ffmpeg imagemagick + npm i -g rmbg-cli

Go tools (after Go is present — identical on all OSes): go install <module> for subfinder, amass, gau, gitleaks, httpx. PhoneInfoga and TruffleHog → GitHub release binary per OS/arch (go install rejects TruffleHog's module for its replace directives; the PyPI trufflehog is the abandoned v2 Python tool and does not accept v3 syntax). ASN → Git Bash/WSL bash <(curl -sL …/nitefood/asn/master/asn) on Windows, native bash on macOS/Linux, or RDAP/ipwho.is HTTP fallback.

Vision / A-V analysis — multix CLI (standalone, npx, OS-identical): no install step — run via npx --yes --prefer-online --package=@mrgoonie/multix@latest -- multix .... Needs Node.js 20+ (winget install OpenJS.NodeJS.LTS / brew install node / sudo apt install -y nodejs npm) and GEMINI_API_KEY (read from env or the skill's own $SKILL_DIR/.env; env-var override applies, same as every other key). Pre-warm the npm cache with ... multix --version before network-restricted runs. Missing key/Node → log a collection gap and fall back to exiftool metadata.

The exact winget IDs, brew formulae, apt packages, uv commands, import names, and Go module paths for every tool are tabulated in scripts/platform-setup.md §5. To provision a fresh machine in one shot, run the bundled installer for the detected OS:

• Windows: powershell -ExecutionPolicy Bypass -File "$env:USERPROFILE\.claude\skills\cti-expert\scripts\install.ps1"
• macOS / Linux / Git Bash / WSL: bash ~/.claude/skills/cti-expert/scripts/install.sh

(both accept --headless/-Headless, --go/-Go, --all/-All)

Behavior Rules

• uv first (anything Python): prefer uv — uv tool install for CLIs, uv pip install for libraries, uv run for scripts. It behaves identically on every OS, so reach for it before per-OS pip/pipx. Fall back to <$PY> -m pip / pipx only when uv cannot be installed.
• Detect before install (system binaries): dispatch by the OS detected in Step 0 — never run apt on Windows or winget on Linux.
• Silent install: do not ask permission — install and proceed. Tool installation is a normal part of the investigation workflow.
• Go tools: require Go installed. If go is missing, install it via the OS package manager (winget install GoLang.Go / brew install go / sudo apt install -y golang), or note the gap and fall back to the next tool in the cascade.
• Linux privileges: system installs need sudo unless running as root.
• Windows specifics: winget may prompt UAC; a freshly installed tool may not be on PATH until the shell is reopened (probe its install dir or restart the shell before declaring failure). The DOCX generator self-heals UTF-8 output and pandoc location — see platform-setup.md §6.
• Git-based install: for tools without a PyPI package (msftrecon, blackbird, sharetrace), clone the repo and install its requirements.txt with <$PY> -m pip install -r requirements.txt.
• Fallback on install failure: if installation fails, log a collection gap and skip to the next tool in the cascade — never block the investigation.