@aegis-scan/skills 0.5.0 → 0.5.1

package/skills/compliance/aegis-native/brutaler-anwalt/hooks/post_write.py

@@ -0,0 +1,315 @@
+#!/usr/bin/env python3
+"""
+PostToolUse-Hook fuer brutaler-anwalt (v4.3.0+).
+
+Triggers nach Write/Edit auf Audit-Output-Files. Erzwingt 3 Quality-Gates:
+
+  1. DISCLAIMER-BLOCK (RDG §2, blocking warning)
+     - "Haftungsausschluss" + "Keine Rechtsberatung" in ersten 40 Zeilen
+     - stderr-Warnung non-blocking (User muss bewusst entscheiden)
+
+  2. FINDING-ID-UNIQUENESS (blocking exit 2)
+     - Audit-Output-Files: pro "## Finding (F-NNN)" Pattern duerfen IDs nicht dupliziert sein
+     - Verhindert versehentliche Dup-IDs nach Persona-Synthese
+
+  3. AZ.-PROVENANCE-GUARD (blocking exit 2, SKILL.md §5)
+     - Jede Az.-Nummer im Output braucht Source-Markierung
+     - Akzeptierte Marker: Source-URL in selber Zeile / Folgezeile
+                         oder "[ungeprueft, manuelle Verifikation vor Schriftsatz erforderlich]"
+                         oder "[secondary-source-verified]"
+                         oder "[primary-source-verified]"
+     - Verdaechtige Pattern (Halluzinations-Indikatoren) werden separat gemeldet:
+       - "1234/22", "9999/22" Placeholder-Style
+       - Az.-Jahr vs. Urteilsjahr divergiert > 2 Jahre
+       - Az. ohne jeglichen Source-Hint im Umkreis
+
+Input:   JSON auf stdin (tool_input.file_path)
+Output:  stderr-Warnungen + Exit-Code 0/2
+
+Aktivierung: ueber .claude-plugin/plugin.json hookSpec
+Disable:    set BRUTALER_ANWALT_HOOK_DISABLE=1 in env
+"""
+from __future__ import annotations
+
+import json
+import os
+import re
+import sys
+from typing import Iterable
+
+
+RELEVANT_PATH_PATTERNS = [
+    re.compile(r"audits?[\\/].+\.md$"),
+    re.compile(r"compliance[\\/].+\.md$"),
+    re.compile(r"legal-audit[\\/].+\.md$"),
+    re.compile(r"BRUTALER-AUDIT.*\.md$"),
+    re.compile(r"COMPLIANCE-AUDIT.*\.md$"),
+    re.compile(r".*-AUDIT-\d{4}-\d{2}-\d{2}.*\.md$"),
+    re.compile(r"ABMAHN-SIMULATION.*\.md$"),
+    re.compile(r"FINDINGS.*\.md$"),
+    re.compile(r"LegalAudit\.md$"),
+    re.compile(r"DSAR-\d{4}-\d{2}-\d{2}.*\.md$"),
+    re.compile(r"AVV-REDLINE.*\.md$"),
+    re.compile(r"references[\\/]streitwerte\.json$"),
+]
+
+
+MAX_FILE_SIZE_BYTES = 5_000_000
+
+FORBIDDEN_PATH_PREFIXES = [
+    "/etc/", "/var/", "/usr/", "/sys/", "/proc/", "/dev/",
+    "/private/etc/", "/private/var/",
+    "/Library/Keychains/", "/Library/Application Support/Keychain/",
+]
+
+DISCLAIMER_HEAD = re.compile(r"Haftungsausschluss|Disclaimer", re.IGNORECASE)
+DISCLAIMER_BODY = re.compile(r"Keine\s+Rechtsberatung|not\s+legal\s+advice|RDG", re.IGNORECASE)
+
+FINDING_ID_RE = re.compile(r"^##\s+Finding\s+(F-\d{3})", re.MULTILINE)
+
+AZ_PATTERNS = [
+    re.compile(r"\bBGH\s+[IVX]+\s+ZR\s+\d{1,4}/\d{2}\b"),
+    re.compile(r"\bBGH\s+[IVX]+\s+ZB\s+\d{1,4}/\d{2}\b"),
+    re.compile(r"\bBVerfG\s+\d\s+BvR\s+\d{1,4}/\d{2}\b"),
+    re.compile(r"\bEuGH\s+C-\d{1,4}/\d{2}\b"),
+    re.compile(r"\bOLG\s+\w+\s+\d{1,2}\s+U\s+\d{1,4}/\d{2}\b"),
+    re.compile(r"\bLG\s+\w+\s+\d{1,2}\s+O\s+\d{1,4}/\d{2}\b"),
+]
+
+SOURCE_MARKER_RE = re.compile(
+    r"\["
+    r"(ungeprueft|secondary-source-verified|primary-source-verified|verified)"
+    r"[^\]]*\]"
+    r"|https?://(?:juris\.bundesgerichtshof\.de|bundesgerichtshof\.de|curia\.europa\.eu"
+    r"|eur-lex\.europa\.eu|dejure\.org|openjur\.de|rewis\.io|nrwe\.justiz\.nrw\.de"
+    r"|medien-internet-und-recht\.de|bundesanzeiger\.de|gesetze-im-internet\.de)",
+    re.IGNORECASE,
+)
+
+SUSPICIOUS_AZ_PLACEHOLDERS = [
+    re.compile(r"\b1234/\d{2}\b"),
+    re.compile(r"\b9999/\d{2}\b"),
+    re.compile(r"\b1111/\d{2}\b"),
+    re.compile(r"\b0000/\d{2}\b"),
+    re.compile(r"\b5678/\d{2}\b"),
+    re.compile(r"\b(?:0123|1234|2345|3456|4567|5678|6789)/\d{2}\b"),
+    re.compile(r"\b(?:1000|2000|3000|4000|5000|6000|7000|8000)/\d{2}\b"),
+]
+
+BYPASS_AZ_CHECK_RE = re.compile(
+    r"<!--\s*brutaler-anwalt:\s*bypass-az-check"
+    r"(?:,\s*reason=(?:\"|')([^\"']+)(?:\"|'))?\s*-->",
+    re.IGNORECASE,
+)
+BYPASS_DISCLAIMER_RE = re.compile(
+    r"<!--\s*brutaler-anwalt:\s*bypass-disclaimer"
+    r"(?:,\s*reason=(?:\"|')([^\"']+)(?:\"|'))?\s*-->",
+    re.IGNORECASE,
+)
+BYPASS_ALL_RE = re.compile(
+    r"<!--\s*brutaler-anwalt:\s*bypass-all"
+    r"(?:,\s*reason=(?:\"|')([^\"']+)(?:\"|'))?\s*-->",
+    re.IGNORECASE,
+)
+
+
+def main() -> int:
+    if os.environ.get("BRUTALER_ANWALT_HOOK_DISABLE") == "1":
+        return 0
+
+    if hasattr(sys.stdout, "reconfigure"):
+        sys.stdout.reconfigure(encoding="utf-8")
+    if hasattr(sys.stderr, "reconfigure"):
+        sys.stderr.reconfigure(encoding="utf-8")
+
+    try:
+        data = json.loads(sys.stdin.read() or "{}")
+    except Exception:
+        return 0
+
+    file_path = (data.get("tool_input") or {}).get("file_path") or ""
+    if not file_path or not os.path.isfile(file_path):
+        return 0
+
+    try:
+        real_path = os.path.realpath(file_path)
+    except (OSError, ValueError):
+        return 0
+
+    for forbidden in FORBIDDEN_PATH_PREFIXES:
+        if real_path.startswith(forbidden):
+            print(
+                f"[brutaler-anwalt] SECURITY: Hook refused to read system-path {real_path!r} "
+                f"(matches forbidden prefix {forbidden!r}). Pfad-Traversal-Verdacht.",
+                file=sys.stderr,
+            )
+            return 0
+
+    normalized = real_path.replace("\\", "/")
+    if not any(p.search(normalized) for p in RELEVANT_PATH_PATTERNS):
+        return 0
+
+    try:
+        size = os.path.getsize(real_path)
+    except OSError:
+        return 0
+    if size > MAX_FILE_SIZE_BYTES:
+        print(
+            f"[brutaler-anwalt] WARNUNG: {real_path} ist {size} bytes > "
+            f"{MAX_FILE_SIZE_BYTES}-byte-Cap. Hook ueberspringt Validierung. "
+            f"Bitte File splitten oder bewusst akzeptieren.",
+            file=sys.stderr,
+        )
+        return 0
+
+    try:
+        with open(real_path, "r", encoding="utf-8") as f:
+            text = f.read()
+    except Exception:
+        return 0
+
+    file_path = real_path
+
+    first_lines = "\n".join(text.splitlines()[:40])
+    exit_code = 0
+
+    bypass_all_match = BYPASS_ALL_RE.search(text)
+    if bypass_all_match:
+        reason = bypass_all_match.group(1) or "<no reason given>"
+        print(
+            f"[brutaler-anwalt] BYPASS-ALL aktiv in {file_path}: {reason!r}. "
+            f"Alle Quality-Gates uebersprungen — User-eigene Verantwortung.",
+            file=sys.stderr,
+        )
+        return 0
+
+    bypass_disclaimer = bool(BYPASS_DISCLAIMER_RE.search(first_lines))
+
+    # Gate 1: Disclaimer (non-blocking warn)
+    if not bypass_disclaimer and (
+        not DISCLAIMER_HEAD.search(first_lines) or not DISCLAIMER_BODY.search(first_lines)
+    ):
+        print(
+            f"[brutaler-anwalt] WARNUNG: {file_path} fehlt der verpflichtende "
+            f"Disclaimer-Block (Haftungsausschluss + Keine Rechtsberatung / RDG §2) "
+            f"in den ersten 40 Zeilen.",
+            file=sys.stderr,
+        )
+        print(
+            "[brutaler-anwalt] Vorlage:\n"
+            "  > **Haftungsausschluss**: Diese Analyse ist keine Rechtsberatung\n"
+            "  > i.S.d. § 2 RDG und ersetzt keinen zugelassenen Rechtsanwalt.",
+            file=sys.stderr,
+        )
+
+    # Gate 2: Finding-ID-Uniqueness (blocking)
+    ids = FINDING_ID_RE.findall(text)
+    seen: set[str] = set()
+    dupes: set[str] = set()
+    for fid in ids:
+        if fid in seen:
+            dupes.add(fid)
+        seen.add(fid)
+    if dupes:
+        print(
+            f"[brutaler-anwalt] FEHLER: {file_path} enthaelt doppelte Finding-IDs: "
+            f"{sorted(dupes)}. Findings muessen eindeutig durchnummeriert sein.",
+            file=sys.stderr,
+        )
+        exit_code = 2
+
+    # Gate 3: Az.-Provenance (blocking)
+    is_json = file_path.endswith(".json")
+    unsourced_az = collect_unsourced_az(text, is_json=is_json)
+    suspicious_az = collect_suspicious_az(text)
+
+    if unsourced_az:
+        print(
+            f"[brutaler-anwalt] FEHLER: {file_path} enthaelt {len(unsourced_az)} "
+            f"Az. ohne Source-Marker (SKILL.md §5 Az.-Provenance-Pflicht):",
+            file=sys.stderr,
+        )
+        for az_str, line_no in unsourced_az[:10]:
+            print(f"  - Zeile {line_no}: {az_str!r}", file=sys.stderr)
+        if len(unsourced_az) > 10:
+            print(f"  ... ({len(unsourced_az) - 10} weitere)", file=sys.stderr)
+        print(
+            "[brutaler-anwalt] Markiere jede Az. mit einem der Tags:\n"
+            "  [primary-source-verified]   — Az. aus juris/curia/dejure-Volltext verifiziert\n"
+            "  [secondary-source-verified] — Az. aus etablierter Sekundaerquelle (medien-internet-und-recht etc.)\n"
+            "  [ungeprueft, manuelle Verifikation vor Schriftsatz erforderlich] — Skill konnte nicht verifizieren\n"
+            "Oder fuege die Source-URL in der selben oder Folgezeile ein.",
+            file=sys.stderr,
+        )
+        exit_code = 2
+
+    if suspicious_az:
+        print(
+            f"[brutaler-anwalt] HALLUZINATIONS-VERDACHT: {file_path} enthaelt "
+            f"{len(suspicious_az)} Az. mit Placeholder-Pattern (1234/22, 9999/22 etc.):",
+            file=sys.stderr,
+        )
+        for az_str, line_no in suspicious_az[:5]:
+            print(f"  - Zeile {line_no}: {az_str!r}", file=sys.stderr)
+        print(
+            "[brutaler-anwalt] WebFetch zur Volltext-Verifikation pflicht "
+            "(SKILL.md §5c). Bei keinem Treffer: Az. aus Output entfernen.",
+            file=sys.stderr,
+        )
+        exit_code = 2
+
+    return exit_code
+
+
+def collect_unsourced_az(text: str, is_json: bool = False) -> list[tuple[str, int]]:
+    """
+    Liefert alle Az.-Treffer, denen ein Source-Marker fehlt.
+    Source-Marker = explicit-tag ODER bekannte Quelle-URL ODER Reference-File-Pfad
+    in derselben Zeile oder den N folgenden Zeilen (markdown=3, json=8).
+    Per-Finding-Bypass: HTML-Comment `<!-- brutaler-anwalt: bypass-az-check -->`
+    in der selben Zeile oder den 2 vorausgehenden Zeilen blockt den Check.
+
+    JSON-Files: source_url-Felder liegen oft 4-6 Zeilen entfernt vom Az. (Schema-
+    struktur). Daher groesseres Window fuer JSON.
+    """
+    lines = text.splitlines()
+    findings: list[tuple[str, int]] = []
+    window_size = 10 if is_json else 4
+    for i, line in enumerate(lines):
+        for pat in AZ_PATTERNS:
+            for match in pat.finditer(line):
+                az_str = match.group(0)
+                window = "\n".join(lines[i:min(i + window_size, len(lines))])
+                if SOURCE_MARKER_RE.search(window):
+                    continue
+                if "references/bgh-urteile.md" in window:
+                    continue
+                if "bgh-urteile.md" in window:
+                    continue
+                bypass_window = "\n".join(lines[max(0, i - 2):i + 2])
+                if BYPASS_AZ_CHECK_RE.search(bypass_window):
+                    continue
+                findings.append((az_str, i + 1))
+    return findings
+
+
+def collect_suspicious_az(text: str) -> list[tuple[str, int]]:
+    """Az. mit verdaechtigen Placeholder-Patterns (1234/22, 9999/22)."""
+    lines = text.splitlines()
+    findings: list[tuple[str, int]] = []
+    for i, line in enumerate(lines):
+        for pat in AZ_PATTERNS:
+            for match in pat.finditer(line):
+                az_str = match.group(0)
+                for suspicious in SUSPICIOUS_AZ_PLACEHOLDERS:
+                    if suspicious.search(az_str):
+                        bypass_window = "\n".join(lines[max(0, i - 2):i + 2])
+                        if BYPASS_AZ_CHECK_RE.search(bypass_window):
+                            continue
+                        findings.append((az_str, i + 1))
+                        break
+    return findings
+
+
+if __name__ == "__main__":
+    sys.exit(main())

package/skills/compliance/aegis-native/brutaler-anwalt/hooks/prompt_submit.py

@@ -0,0 +1,144 @@
+#!/usr/bin/env python3
+"""
+UserPromptSubmit-Hook fuer brutaler-anwalt (v4.3.0+).
+
+Regex-matcht User-Prompt gegen Trigger aus hooks/triggers.json.
+Bei Match: laedt die zugeordneten reference-Files und injiziert sie
+als additionalContext. Spart vs. Vollload des References-Trees
+typisch 70-90% Tokens pro Audit.
+
+Halt-Condition: wenn KEIN Trigger matcht und der Prompt offensichtlich
+ein Audit-Auftrag ist (Keywords: audit, compliance, dsgvo, abmahnung),
+laed wir als Fallback `audit-patterns.md` immer.
+
+Disable: set BRUTALER_ANWALT_HOOK_DISABLE=1 in env.
+"""
+from __future__ import annotations
+
+import json
+import os
+import re
+import sys
+
+
+AUDIT_FALLBACK_KEYWORDS = re.compile(
+    r"\b(audit|compliance|dsgvo|abmahn|datenschutz|impressum|cookie|agb|"
+    r"avv|drittland|consent|uwg|nis2|ai-act|dsa|tdddg|ttdsg|gobd|art-9|"
+    r"art-13|art-15|art-83|datenpanne|breach|brutaler|anwalt)\b",
+    re.IGNORECASE,
+)
+AUDIT_FALLBACK_FILES = ["audit-patterns.md", "dsgvo.md", "bgh-urteile.md"]
+
+MAX_FILES_PER_PROMPT = 8
+MAX_BYTES_PER_FILE = 80_000
+
+
+def main() -> int:
+    if os.environ.get("BRUTALER_ANWALT_HOOK_DISABLE") == "1":
+        return 0
+
+    if hasattr(sys.stdout, "reconfigure"):
+        sys.stdout.reconfigure(encoding="utf-8")
+    if hasattr(sys.stdin, "reconfigure"):
+        sys.stdin.reconfigure(encoding="utf-8")
+
+    skill_root = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
+    triggers_path = os.path.join(skill_root, "hooks", "triggers.json")
+    refs_dir = os.path.join(skill_root, "references")
+
+    if not os.path.isfile(triggers_path):
+        return 0
+
+    prompt = ""
+    try:
+        raw = sys.stdin.read()
+        if raw.strip():
+            data = json.loads(raw)
+            prompt = data.get("prompt", "") or ""
+    except Exception:
+        return 0
+
+    if not prompt:
+        return 0
+
+    try:
+        with open(triggers_path, "r", encoding="utf-8") as f:
+            cfg = json.load(f)
+    except Exception:
+        return 0
+
+    matched_files: list[str] = []
+    seen: set[str] = set()
+    matched_triggers: list[str] = []
+
+    for trigger in cfg.get("triggers", []):
+        pattern = trigger.get("pattern", "")
+        name = trigger.get("name", "<unnamed>")
+        if not pattern:
+            continue
+        try:
+            if re.search(pattern, prompt, re.IGNORECASE):
+                matched_triggers.append(name)
+                for rel in trigger.get("files", []):
+                    if rel not in seen:
+                        seen.add(rel)
+                        matched_files.append(rel)
+        except re.error:
+            continue
+
+    if not matched_files and AUDIT_FALLBACK_KEYWORDS.search(prompt):
+        for rel in AUDIT_FALLBACK_FILES:
+            if rel not in seen:
+                seen.add(rel)
+                matched_files.append(rel)
+        matched_triggers.append("<fallback-audit>")
+
+    if not matched_files:
+        return 0
+
+    if len(matched_files) > MAX_FILES_PER_PROMPT:
+        matched_files = matched_files[:MAX_FILES_PER_PROMPT]
+
+    blocks: list[str] = []
+    for rel in matched_files:
+        full = os.path.join(refs_dir, rel.replace("/", os.sep))
+        if not os.path.isfile(full):
+            blocks.append(
+                f"### references/{rel}\n\n"
+                f"> [Hinweis] Reference-File noch nicht angelegt.\n"
+            )
+            continue
+        try:
+            with open(full, "r", encoding="utf-8") as f:
+                content = f.read(MAX_BYTES_PER_FILE)
+                if os.path.getsize(full) > MAX_BYTES_PER_FILE:
+                    content += (
+                        f"\n\n> [Hinweis] File >{MAX_BYTES_PER_FILE} bytes — "
+                        "Inhalt abgeschnitten. Manueller `Read`-Tool-Aufruf "
+                        "fuer Volltext."
+                    )
+        except Exception as e:
+            content = f"> [Lese-Fehler] {e}"
+        blocks.append(f"### references/{rel}\n\n{content}\n")
+
+    header = (
+        "# brutaler-anwalt — On-Demand-KB-Chunks (via UserPromptSubmit-Hook)\n\n"
+        f"Trigger gefeuert: {', '.join(matched_triggers)}\n"
+        f"Geladene Files: {len(matched_files)}\n\n"
+        "Nutze diese als Primaer-Referenz fuer den aktuellen Prompt. "
+        "Wenn Luecken: manueller `Read` auf weitere references/ oder "
+        "WebFetch (Tier-1-Allowlist in settings.json).\n\n---\n\n"
+    )
+
+    out = {
+        "hookSpecificOutput": {
+            "hookEventName": "UserPromptSubmit",
+            "additionalContext": header + "\n\n---\n\n".join(blocks),
+        }
+    }
+    print(json.dumps(out, ensure_ascii=False))
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())

package/skills/compliance/aegis-native/brutaler-anwalt/hooks/session_start.py

@@ -0,0 +1,57 @@
+#!/usr/bin/env python3
+"""
+SessionStart-Hook fuer brutaler-anwalt (v4.3.0+).
+
+Laedt einen schlanken INDEX (~5 KB) aus references/INDEX.md statt
+des kompletten 5056-Zeilen-References-Trees. Volltext-Chunks werden
+on-demand per UserPromptSubmit-Hook geladen, wenn Keywords matchen.
+
+Disable: set BRUTALER_ANWALT_HOOK_DISABLE=1 in env.
+"""
+from __future__ import annotations
+
+import json
+import os
+import sys
+
+
+def main() -> int:
+    if os.environ.get("BRUTALER_ANWALT_HOOK_DISABLE") == "1":
+        return 0
+
+    if hasattr(sys.stdout, "reconfigure"):
+        sys.stdout.reconfigure(encoding="utf-8")
+
+    skill_root = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
+    index_path = os.path.join(skill_root, "references", "INDEX.md")
+
+    if not os.path.isfile(index_path):
+        return 0
+
+    try:
+        with open(index_path, "r", encoding="utf-8") as f:
+            content = f.read()
+    except Exception:
+        return 0
+
+    header = (
+        "# brutaler-anwalt — Reference-INDEX (via SessionStart-Hook)\n\n"
+        "**Volltext-Inhalte werden on-demand geladen**, wenn dein Prompt "
+        "passende Keywords enthaelt (siehe `hooks/triggers.json`).\n"
+        "Manuelles Laden: `Read references/<file>.md`.\n\n"
+        "Bei Audit-Start IMMER zuerst `references/audit-patterns.md` laden "
+        "(SKILL.md Schritt 0).\n\n---\n\n"
+    )
+
+    out = {
+        "hookSpecificOutput": {
+            "hookEventName": "SessionStart",
+            "additionalContext": header + content,
+        }
+    }
+    print(json.dumps(out, ensure_ascii=False))
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())