Engagement phases
Every engagement runs through four phases. The phase is stored in engagement.json, shown as chips in the menu bar, and critically it decides which tools the model can see. Before the engagement reaches a phase, that phase's tool groups are withheld from the model entirely.
The four phases
| Phase | What happens | Tool groups unlocked |
|---|---|---|
| Planning & Pre-Engagement | No active testing. castiel discusses the target, scope, objectives and rules with you; passive OSINT from third-party sources only, no traffic reaches the target. | Shell · Files · Workspace layout · Web · Workspace search · Ask the user |
| Reconnaissance | Information gathering is open. castiel interacts with in-scope targets freely and fills in network.json, web.json and identity.json. | Sub-agents · Kali discovery · Kali launch · Background processes plus every connected MCP tool |
| Scanning & Vulnerability Assessment | castiel probes in-scope targets for weaknesses using the recon data and records each issue in vulnerabilities.json. It never writes a CVSS score by hand: after recording a finding it calls cvss_score with a CVSS v4.0 and a CVSS v3.1 vector. | Vulnerability intel · CVSS scoring · MITRE ATT&CK |
| Exploitation | castiel attempts to exploit confirmed vulnerabilities and reports every attempt and result to you, stopping when told to. | Nothing new, everything is already unlocked; exploitation itself is what you authorize |
You move the phases
castiel can suggest moving to the next phase when it believes the current one is complete, but the decision is always yours: you transition between phases using the phase chips in the menu bar. Nothing scans or exploits anything until you have moved the engagement there.
The full tool-by-tool breakdown, including which groups cannot be switched off, is on the Tools page.