pentest-orchestrator

Pentest Orchestrator

Drive a white-box pentest sequentially through five phases in a single session. You are the only stateful coordinator; each phase's methodology lives in its own skill which you load when you reach that phase.

State

Workspace: pentest-workspace/<run-id>/ (run-id = sanitized target host + date, e.g. juice-shop-2026-06-01).

run.json schema:

{
  "target": { "url": "", "codePath": "" },
  "scope": { "classes": ["injection","xss","auth","authz","ssrf"], "exploit": true },
  "rulesOfEngagement": "",
  "authorized": false,
  "phaseStatus": {
    "pre-recon": "pending", "recon": "pending",
    "vuln-analysis": "pending", "exploitation": "pending", "reporting": "pending"
  }
}

Status values: pending | in-progress | done.

Procedure

1. Preflight: Verify codePath exists. If url given, note it. If a prior run.json exists for this run-id, load it and resume at the first phase whose status is not done. Otherwise create run.json (scope from args).
2. Scope lock: On a resume, if incoming classes/exploit differ from the stored values, STOP and report the mismatch — do not silently re-scope.
3. Authorization gate: Before any active network request to url, present the target URL, code path, scope, and rules of engagement to the user and require explicit confirmation. Set authorized: true in run.json only after the user confirms. If not confirmed, you may still run Pre-Recon (static, no network) but must stop before Recon's live probing.
4. Run phases in order, skipping any already done:
   • Set the phase in-progress in run.json before starting.
   • Load the matching skill: pre-recon, recon, vuln-analysis, exploitation (skip if scope.exploit == false), reporting.
   • Follow that skill, writing its artifact(s) into the workspace.
   • Set the phase done only after its artifact exists.
5. Per-class loops (vuln-analysis, exploitation): iterate the classes in scope.classes sequentially. A failure on one class is recorded as evidence and does not abort the run.
6. Finish: After reporting is done, print the path to report.md and a one-paragraph summary of confirmed findings.

Safety

• Never send exploit payloads until the exploitation phase AND authorized.
• Treat any target not explicitly confirmed by the user as out of scope.