How a turn is processed
Every message you send goes through the same pipeline:
- Slash-command parsing. If your message starts with or contains
/<cmd>, the matching skill is activated immediately — no router guess. - Router pass. Otherwise the router scores all skills against your prompt and picks the top few that get their tools wired in for this turn.
- Tool execution. Activated tools run in parallel — passive DNS, breach lookups, sandbox queries, mindmap renderers, etc. Each one is a real network call.
- Synthesis. Findings are streamed back as short prose, structured cards, and code blocks where useful.
Conversations are auto-titled after the second turn and stored against your account so you can resume them.
Composer affordances
Slash picker
Type / to open the picker. Filter by typing, confirm with ↵. The picker shows the label, description, and an example for every command.
Server mentions
Type @ to open the registered-server picker. Selecting one inserts @<name>as an unambiguous target — the model will not ask “which server?” when a mention is present.
Multiple commands in one message
You can chain commands. /scan /checkleak acme.com runs both on the same target. The model interprets each one and merges the answer.
Slash command catalogue
Single source of truth for what each command does and what argument it expects. Generated directly from SLASH_COMMANDS in packages/ai/src/slash-commands.ts.
/flow-perimeterFlow · Perimeter auditskill: flowsFull external attack surface — domain → subdomains → IPs → ports → services → certs → fingerprints.
expects: domain› /flow-perimeter acme.com/flow-leaksFlow · Leak auditskill: flowsSweep breach indices, paste sites, and code hosts for credentials and tokens tied to a domain or brand.
expects: domain› /flow-leaks acme.com/flow-squatFlow · Typosquat watchskill: flowsGenerate lookalike-domain set, resolve each, score for live phishing infrastructure, draft takedown packets.
expects: domain› /flow-squat acme.com/flow-vulnsFlow · Vuln sweepskill: flowsMatch recent CVEs against the detected stack on the target. Sorted by exploitability × severity.
expects: domain (or stack tags)› /flow-vulns acme.com/flow-kickoffFlow · Engagement kickoffskill: flowsFirst-hour brief — perimeter + leak audit + light typosquat + draft threat model. Run it after creating a project.
expects: domain› /flow-kickoff acme.com/flow-iocFlow · IOC triageskill: flowsIndicator → impact in one pass: reputation, sandbox, passive-DNS pivot, ASN/cert correlation, incident packet.
expects: IP, hash, URL, or domain› /flow-ioc 1.1.1.1
/scanRecon scanskill: reconMap a target — IPs, ports, services, subdomains, technology fingerprint.
expects: domain or IP› /scan acme.com
/checkleakBreach / leak checkskill: intel-breachSearch dark-web breaches and HIBP for an email or domain.
expects: email or domain› /checkleak [email protected]/cveCVE lookupskill: intel-vulnResolve a CVE ID — CVSS, vendor advisories, exploits.
expects: CVE-YYYY-NNNNN› /cve CVE-2024-3094/threatThreat intelskill: intel-threatsReputation lookup — VirusTotal, AbuseIPDB, GreyNoise, urlscan.
expects: IP, hash, URL, or domain› /threat 1.2.3.4/brandBrand OSINTskill: brand-osintExtract a site's logos, color palette, and brand identity.
expects: URL› /brand https://stripe.com
/payloadGenerate payloadskill: offensiveReverse / bind shells, msfvenom, web shells.
expects: type + LHOST/LPORT (optional)› /payload reverse_shell lhost=1.2.3.5 lport=4444/ctfCTF web triageskill: ctf-webApproach a CTF web challenge — recon, vuln triage, exploit.
expects: URL› /ctf http://1.2.3.6/login/postPost-exploit guidanceskill: post-exploitPrivesc, lateral movement, persistence, AD attacks.
expects: context (Linux/Windows/AD details)› /post i have a shell as www-data on debian/revReverse engineeringskill: reverse-engStatic / dynamic binary analysis — Ghidra, GDB, radare2.
expects: paste disassembly or describe the binary› /rev pasted Ghidra decompilation here/runRun on serverskill: remote-execPropose a command for one of your registered SSH servers. Combine with @<server>.
expects: @server + command› /run @kali-docker df -h
/sigmaSigma / detection ruleskill: blue-teamGenerate a Sigma / KQL / SPL detection rule for a TTP or pasted log.
expects: technique or pasted log› /sigma LSASS access via WinAPI/huntThreat hunt queryskill: blue-teamHunt query for credential dumping, lateral movement, C2.
expects: TTP description› /hunt unusual outbound DNS volume per host
/decodeHash / cipher IDskill: cryptoIdentify a hash (md5/sha*/bcrypt) or decode common encodings.
expects: hash, base64, or hex blob› /decode 5f4dcc3b5aa765d61d8327deb882cf99
/mindmapMindmapskill: threat-modelGenerate a Mermaid mindmap (center node + radial branches) for a topic.
expects: topic to map› /mindmap web app pentest checklist/diagramDiagramskill: threat-modelRender a Mermaid diagram — flowchart, sequence, kill chain, DFD.
expects: what to diagram› /diagram OAuth2 PKCE flow/cvssCVSS scoreskill: reportingCompute a CVSS v3.1 vector + score from a description.
expects: vulnerability description› /cvss SQL injection in /api/login/reportPentest findingskill: reportingDraft a vulnerability finding — title, impact, repro, remediation.
expects: finding details› /report XSS reflected in search param/threatmodelThreat modelskill: threat-modelSTRIDE / PASTA / attack-tree analysis of a system.
expects: system description› /threatmodel SaaS app with public API + mobile clients
Output blocks you'll see
- Prose — short, dense paragraphs. No disclaimers, no padding.
- Findings cards — labelled key/value rows for recon, intel, breach hits, etc., with severity dots.
- Code blocks — payloads, queries, Sigma / KQL / SPL rules, JSON samples. Always copy-paste-ready.
- Command cards — when the remote-exec skill is involved, the model emits a card you click to run on a registered server.
- Mermaid diagrams — mindmaps, flowcharts, kill chains, DFDs from
/mindmapand/diagram.
A worked example
Sample turn: combining recon and breach intel on one message.
/scan /checkleak acme.comTwo skills activate. Recon enumerates subdomains, ports, and stack fingerprint; intel-breach checks paste indices and known breach sources. The reply weaves both into a single answer with separate findings cards and a closing recommendation.