sentinel: log subprocess failures + AI-busy retry/callback + allowlist all ban-paths

Three issues from J 2026-04-30:

1. Silent fail2ban-client / nginx subprocess failures
   Pre-fix _execute_ban called fail2ban-client with capture_output=True
   but threw away the result. _nginx_ban had bare `except: pass`
   swallowing everything. So a non-zero fail2ban exit (jail not
   configured, IP already banned, IPv6 quirk) or PermissionError
   on /etc/nginx/banned_ips.conf logged "AI_BAN" while the attacker
   walked through unimpeded. The "errors in logs" J was seeing.

   Now every subprocess failure surfaces:
   - FAIL2BAN_FAILED rc=N stderr=... — non-zero exit
   - FAIL2BAN_TIMEOUT — client didn't return in 5s
   - FAIL2BAN_NOT_INSTALLED — binary missing
   - NGINX_BAN_WRITE_DENIED — permission error on conf file
   - NGINX_RELOAD_FAILED rc=N stderr=... — systemctl reload non-zero
   - NGINX_RELOAD_TIMEOUT / NGINX_RELOAD_NO_SYSTEMCTL — runtime gaps
   sec_log.error catches these so journalctl -u llm-team-ui shows
   the actual reason a ban didn't stick.

2. AI auto-scan failure callback when model is busy
   Pre-fix Ollama unreachable / busy / timeout silently preserved
   log position + skipped the scan. Operator only learned about
   the gap by manually checking sentinel-status. Now:
   - 1 retry inside same scan after SENTINEL_AI_RETRY_DELAY_SECS
     (30s) on connection error / timeout / 429 / 503
   - 4xx errors that won't recover (404 model missing, 400 bad
     prompt) fail fast without retrying
   - consecutive_ai_failures counter in _sentinel_stats
   - On 3+ consecutive failures, send_security_alert() fires —
     "Sentinel AI unreachable" email with last error + endpoint
     + model name. One alert per outage (ai_busy_alerted flag);
     clears on first successful scan so flapping doesn't spam.
   - AI_RECOVERED log line on first scan after a streak.

3. Sentinel ban path still substring-matched 192.168
   Same vulnerability class as admin_ban_ip had — only protected
   one /16. Replaced 4 sites with is_allowlisted(ip):
   - threat-list display filter (line 7638): now hides ALL
     allowlisted IPs from the panel
   - mass-ban API (line 8016): refuses ban for any allowlisted IP
   - sentinel analysis filter (line 12786): saves AI tokens by
     never sending allowlisted-IP traffic to the judge
   - sentinel ban verdict gate (line 12949): defense in depth —
     even if the AI says "ban" on an allowlisted IP, this catches it

   Combined with the layered defenses in b09b73c (track_violation,
   _auto_escalate, _nginx_ban, admin_ban_ip), there is now no
   code path that can ban an allowlisted IP. Operator self-ban
   is structurally impossible.

Privilege note: the systemd unit at /root/llm-team-ui/llm-team-ui.service
runs as User=root, so subprocess.run(["fail2ban-client", ...]) and
systemctl reload nginx have permission. The "errors in logs" J was
seeing weren't permission-denied; they were silent non-zero exits.
The new subprocess wrappers surface those.

If the operator later splits the app into a non-root tier
(Opus OB-3 architectural recommendation, deferred), this same
infrastructure still works — the wrappers will then surface
"PermissionError" with full path + uid context, telling the
operator exactly which command needs sudo NOPASSWD or
PolicyKit rule.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

.gitignore

@@ -2,3 +2,4 @@ __pycache__/
 *.pyc
 .env
 *.log
+.memory/

llm_team_ui.py

@@ -3,6 +3,7 @@
 
 import json
 import os
+import sys
 import time
 import threading
 import secrets
@@ -36,8 +37,23 @@ app.config["SESSION_COOKIE_HTTPONLY"] = True
 app.config["SESSION_COOKIE_SAMESITE"] = "Lax"
 
 # ─── SECURITY LOGGING ─────────────────────────────────────────
-# Dedicated security log for fail2ban and audit trail
-_sec_handler = logging.FileHandler("/var/log/llm-team-security.log")
+# Dedicated security log for fail2ban and audit trail.
+#
+# Cross-lineage scrum 2026-04-30 (Opus BLOCK OB-1): wrapped in
+# try/except. Pre-fix this raised PermissionError at import time
+# when the service user couldn't write /var/log/llm-team-security.log,
+# crashing the app before Flask started. Now falls back to stderr;
+# sec_log still works (ban events still land in journald via stderr),
+# but the app starts. Operator should still create the file with
+# proper perms on production deploy. Path is overridable via
+# LLM_TEAM_SECURITY_LOG env var.
+_LOG_PATH = os.environ.get("LLM_TEAM_SECURITY_LOG", "/var/log/llm-team-security.log")
+try:
+    _sec_handler = logging.FileHandler(_LOG_PATH)
+except (PermissionError, FileNotFoundError, OSError) as _log_err:
+    print(f"[security] WARNING: can't open {_LOG_PATH} ({_log_err}); "
+          f"falling back to stderr.", file=sys.stderr, flush=True)
+    _sec_handler = logging.StreamHandler(sys.stderr)
 _sec_handler.setFormatter(logging.Formatter("%(asctime)s %(message)s"))
 sec_log = logging.getLogger("security")
 sec_log.addHandler(_sec_handler)
@@ -93,8 +109,24 @@ _original_sentinel_interval = None  # stash the normal interval during high-aler
 
 
 def _track_violation(ip, event_type="unknown"):
-    """Record a security violation. If velocity threshold exceeded, auto-escalate."""
+    """Record a security violation. If velocity threshold exceeded, auto-escalate.
+
+    Cross-lineage scrum 2026-04-30 follow-up: OB-4's path-bypass for
+    admins was incomplete — _track_violation is called from 3 sites
+    (exploit_scan, rate_limit, login_fail) and only the exploit one
+    had the bypass. An admin hitting rate-limit + login-typo + a
+    legit URL containing 'UNION' within 60s could still self-ban
+    via the OTHER paths. Now is_allowlisted bails early so allowlisted
+    IPs never accumulate violations from ANY path. Defense in depth —
+    _auto_escalate also re-checks below.
+
+    Eviction sweep when tracker grows >10K entries (same pattern as
+    _rate_limit; both had identical unbounded-dict WARNs)."""
+    if is_allowlisted(ip):
+        return False
     now = time.time()
+    if len(_violation_tracker) > 10000:
+        _evict_stale_violation_tracker(now)
     if ip not in _violation_tracker:
         _violation_tracker[ip] = []
     _violation_tracker[ip].append(now)
@@ -107,8 +139,26 @@ def _track_violation(ip, event_type="unknown"):
     return False
 
 
+def _evict_stale_violation_tracker(now):
+    """Drop _violation_tracker entries whose newest timestamp is past
+    the window. Called from _track_violation only when dict exceeds
+    10K — the hot path stays untouched for normal traffic."""
+    cutoff = now - VELOCITY_WINDOW
+    stale = [ip for ip, ts in _violation_tracker.items() if not ts or max(ts) < cutoff]
+    for ip in stale:
+        del _violation_tracker[ip]
+
+
 def _auto_escalate(ip, violation_count, event_type):
-    """Auto-ban IP and switch sentinel to high-alert mode."""
+    """Auto-ban IP and switch sentinel to high-alert mode.
+
+    Defense in depth (2026-04-30): _track_violation already short-
+    circuits for allowlisted IPs, but if a future code path calls
+    _auto_escalate directly we still want the allowlist guard. Bail
+    early; nothing escalates against a trusted IP."""
+    if is_allowlisted(ip):
+        sec_log.info("AUTO_ESCALATE_BLOCKED ip=%s — allowlisted, refused to ban", ip)
+        return
     global _original_sentinel_interval, SENTINEL_INTERVAL
     sec_log.warning("AUTO_ESCALATE ip=%s violations=%d/%ds type=%s", ip, violation_count, VELOCITY_WINDOW, event_type)
     _sentinel_log_entry(f"AUTO_ESCALATE ip={ip} violations={violation_count}/{VELOCITY_WINDOW}s type={event_type}")
@@ -150,8 +200,19 @@ def _check_high_alert_expiry():
 
 # IPs that never get rate-limited (your LAN, localhost)
 ALLOWLIST_IPS = {"127.0.0.1", "::1", "192.168.1.1"}
-# Demo mode state — toggled by admin at runtime
-_demo_mode = {"active": True, "started_by": "boot", "showcase": True}
+# Demo mode state — toggled by admin at runtime.
+#
+# Cross-lineage scrum 2026-04-30 (Opus BLOCK OB-5): pre-fix this
+# defaulted to active=True, meaning fresh installs shipped with
+# public unauthenticated access enabled — login_required let demo
+# users straight through. Combined with /api/run + /api/imagegen
+# proxies, that was an open LLM/compute abuse surface from first
+# boot. Now defaults to active=False; operators flip it on
+# explicitly via the admin UI or LLM_TEAM_DEMO_MODE=1 env override
+# (the env override exists for the demo systemd unit so the public
+# devop.live deployment doesn't need a manual toggle on every restart).
+_DEMO_DEFAULT = os.environ.get("LLM_TEAM_DEMO_MODE", "0") == "1"
+_demo_mode = {"active": _DEMO_DEFAULT, "started_by": "boot" if _DEMO_DEFAULT else "off", "showcase": _DEMO_DEFAULT}
 
 # Routes that demo users CAN trigger (read-like POSTs — enrichment, self-analysis, team runs)
 DEMO_ALLOWED_POSTS = {
@@ -172,9 +233,21 @@ def is_allowlisted(ip):
 
 
 def rate_limited(ip, max_req=RATE_LIMIT_MAX):
+    """Rolling rate-limit check. Returns True when the IP has exceeded
+    max_req requests within RATE_LIMIT_WINDOW seconds.
+
+    Cross-lineage scrum 2026-04-30 (Opus WARN): _rate_limit was
+    unbounded per-worker, so an attacker rotating slowly through IPs
+    leaked memory forever. Fix: lazy eviction sweep when the dict
+    grows beyond 10K entries. Real production wants a Redis-backed
+    counter shared across workers; this is the in-process band-aid
+    that prevents runaway growth without changing the deploy shape.
+    """
     if is_allowlisted(ip):
         return False
     now = time.time()
+    if len(_rate_limit) > 10000:
+        _evict_stale_rate_limit(now)
     if ip not in _rate_limit or now - _rate_limit[ip][1] > RATE_LIMIT_WINDOW:
         _rate_limit[ip] = (1, now)
         return False
@@ -185,6 +258,16 @@ def rate_limited(ip, max_req=RATE_LIMIT_MAX):
     return False
 
 
+def _evict_stale_rate_limit(now):
+    """Drop _rate_limit entries older than 2× the window. Called from
+    rate_limited() only when dict growth exceeds 10K — keeps the cost
+    off the hot path for normal traffic."""
+    cutoff = now - (RATE_LIMIT_WINDOW * 2)
+    stale = [ip for ip, (_, start) in _rate_limit.items() if start < cutoff]
+    for ip in stale:
+        del _rate_limit[ip]
+
+
 def is_admin():
     return session.get("role") == "admin"
 
@@ -560,8 +643,24 @@ def security_checks():
     # Check high-alert expiry
     _check_high_alert_expiry()
 
-    # Exploit scanner detection — log, alert, track velocity, block
-    if EXPLOIT_PATTERNS.search(path) or EXPLOIT_PATTERNS.search(request.query_string.decode("utf-8", errors="ignore")):
+    # Exploit scanner detection — log, alert, track velocity, block.
+    #
+    # Cross-lineage scrum 2026-04-30 (Opus BLOCK OB-4): pre-fix the
+    # path regex matched on substrings like UNION, SELECT, ;-- and
+    # auto-banned after 3 hits. Admin URLs containing those keywords
+    # in query strings (e.g. an LLM team named "select-rebrand" or a
+    # docs link to /admin/select_a_mode) self-banned the admin's IP.
+    # Now: skip the path-based check for authenticated admins from
+    # an allowlisted IP. The user-agent + body checks (sentinel) still
+    # apply. Allowlisted-IP admins clicking weird URLs no longer
+    # lock themselves out.
+    _skip_exploit_check = False
+    if ip in ALLOWLIST_IPS and session.get("role") == "admin":
+        _skip_exploit_check = True
+    if not _skip_exploit_check and (
+        EXPLOIT_PATTERNS.search(path) or
+        EXPLOIT_PATTERNS.search(request.query_string.decode("utf-8", errors="ignore"))
+    ):
         sec_log.warning("EXPLOIT_SCAN ip=%s path=%s ua=%s", ip, path, ua)
         _track_violation(ip, "exploit_scan")
         send_security_alert(
@@ -836,7 +935,21 @@ def auth_login():
         with get_db() as conn:
             with conn.cursor(cursor_factory=psycopg2.extras.RealDictCursor) as cur:
                 if is_setup:
-                    # First-time setup: create admin
+                    # First-time setup: create admin.
+                    #
+                    # Cross-lineage scrum 2026-04-30 (Opus WARN): pre-fix
+                    # this was gated only by COUNT(*) FROM users == 0.
+                    # If an operator ever truncated/restored the users
+                    # table, the next external visitor (any IP) could
+                    # claim admin. Now also requires the source IP to
+                    # be in ALLOWLIST_IPS — typically loopback + LAN
+                    # gateway — so a remote attacker hitting the setup
+                    # endpoint after an empty-users state can't seize
+                    # the account. Local operator running setup from
+                    # the box itself still works.
+                    if ip not in ALLOWLIST_IPS:
+                        sec_log.warning("SETUP_DENIED ip=%s — first-time setup requires allowlisted IP", ip)
+                        return jsonify({"error": "setup must be initiated from an allowlisted IP (typically localhost or LAN gateway)"}), 403
                     cur.execute("SELECT COUNT(*) as c FROM users")
                     if cur.fetchone()["c"] > 0:
                         return jsonify({"error": "Setup already completed"}), 400
@@ -1349,6 +1462,22 @@ async function loadThreats() {
     });
     // Mass action buttons
     var spacer = document.createElement('div'); spacer.style.flex = '1'; toolbar.appendChild(spacer);
+    // Master "select all on this page" checkbox (2026-04-30 J UX request).
+    // Mirrors the per-row .ip-check style; toggles every visible row.
+    // Three-state: unchecked (none selected), checked (all selected),
+    // indeterminate (partial). updateSelCount keeps it in sync as
+    // individual rows are toggled.
+    var selAllWrap = document.createElement('label');
+    selAllWrap.style.cssText = 'display:flex;align-items:center;gap:6px;font-family:JetBrains Mono,monospace;font-size:9px;text-transform:uppercase;letter-spacing:0.5px;color:#7a7872;cursor:pointer';
+    selAllWrap.title = 'Toggle every IP on this page';
+    var selAll = document.createElement('input'); selAll.type = 'checkbox';
+    selAll.id = 'sel-all';
+    selAll.style.cssText = 'width:16px;height:16px;cursor:pointer;accent-color:#e2b55a';
+    selAll.onchange = function(){ toggleAllChecks(this.checked); };
+    selAllWrap.appendChild(selAll);
+    var selAllLabel = document.createElement('span'); selAllLabel.textContent = 'all';
+    selAllWrap.appendChild(selAllLabel);
+    toolbar.appendChild(selAllWrap);
     var selCount = document.createElement('span'); selCount.id = 'sel-count';
     selCount.style.cssText = 'font-family:JetBrains Mono,monospace;font-size:10px;color:#7a7872';
     toolbar.appendChild(selCount);
@@ -1472,9 +1601,33 @@ async function loadThreats() {
 var currentSort = 'hits';
 
 function updateSelCount() {
-  var checks = document.querySelectorAll('.ip-check:checked');
+  var all = document.querySelectorAll('.ip-check');
+  var checked = document.querySelectorAll('.ip-check:checked');
   var el = document.getElementById('sel-count');
-  if (el) el.textContent = checks.length ? checks.length + ' selected' : '';
+  if (el) el.textContent = checked.length ? checked.length + ' selected' : '';
+  // Sync the master "all" checkbox to reflect the page's actual state.
+  // Three states: none → unchecked, all → checked, partial → indeterminate.
+  // Indeterminate is the visual "half-tick" most browsers render — gives
+  // operators a clear "you've got some but not all selected" hint.
+  var master = document.getElementById('sel-all');
+  if (master) {
+    if (checked.length === 0) {
+      master.checked = false; master.indeterminate = false;
+    } else if (checked.length === all.length) {
+      master.checked = true; master.indeterminate = false;
+    } else {
+      master.indeterminate = true;
+    }
+  }
+}
+
+function toggleAllChecks(checked) {
+  // Master "select all" handler — flips every per-row checkbox on the
+  // page to match the master's state. Used by the toolbar's `[ ] all`
+  // checkbox so operators don't have to click each threat individually
+  // before hitting Ban Selected. (2026-04-30 J UX request.)
+  document.querySelectorAll('.ip-check').forEach(function(cb){ cb.checked = checked; });
+  updateSelCount();
 }
 
 async function massAction(action) {
@@ -1888,7 +2041,18 @@ def get_api_key(provider_name):
     env_map = {"openrouter": "OPENROUTER_API_KEY", "openai": "OPENAI_API_KEY", "anthropic": "ANTHROPIC_API_KEY", "ollama_cloud": "OLLAMA_CLOUD_API_KEY"}
     return os.environ.get(env_map.get(provider_name, ""), "")
 
-DB_DSN = "dbname=knowledge_base user=kbuser password=IPbLBA0EQI8u4TeM2YZrbm1OAy5nSwqC host=localhost"
+# Cross-lineage scrum 2026-04-30 (Opus BLOCK OB-2 + harness LLM
+# convergent finding): DB_DSN previously had the password hardcoded
+# in source. Same `kbuser`/`knowledge_base` DSN was leaked in
+# voice-ai's audiosocket_bridge.py + sales_assistant.py — confirmed
+# canonical leak by 3 independent reviewers across 2 sessions. Now
+# sourced from env (set via systemd EnvironmentFile=/etc/llm-team-ui.env).
+# No silent fallback to the leaked literal — fail loud. The leaked
+# password is in git history regardless; rotate it in Postgres.
+DB_DSN = os.environ.get("LLM_TEAM_DB_DSN", "")
+if not DB_DSN:
+    print("[llm-team-ui] WARNING: LLM_TEAM_DB_DSN not set — DB ops will fail. "
+          "Set in systemd EnvironmentFile or shell env.", file=sys.stderr, flush=True)
 
 def get_db():
     return psycopg2.connect(DB_DSN)
@@ -7471,7 +7635,10 @@ def admin_security_data():
     sort_by = request.args.get("sort", "hits")
     result = []
     for ip, d in ips.items():
-        if ip.startswith("192.168."):
+        # 2026-04-30: was substring "192.168." — replaced with the
+        # canonical allowlist so 10.x, IPv6 ::1, and operator-added
+        # entries also stay out of the threat panel.
+        if is_allowlisted(ip):
             continue
         result.append({
             "ip": ip, "hits": d["hits"], "exploit_scans": d["exploit_scans"],
@@ -7510,21 +7677,52 @@ def _kill_connections(ip):
         pass
 
 def _nginx_ban(ip):
-    """Add IP to nginx deny list and reload."""
+    """Add IP to nginx deny list and reload.
+
+    Defense in depth (2026-04-30): refuse to write allowlisted IPs
+    to the deny list under ANY circumstance — even a buggy caller
+    that bypassed _track_violation's allowlist check. The deny list
+    is the last write before nginx reload; this is the last place
+    we can stop a bad ban."""
+    if is_allowlisted(ip):
+        sec_log.info("NGINX_BAN_BLOCKED ip=%s — allowlisted, refused to write deny rule", ip)
+        return
     import subprocess
+    line = f"deny {ip};\n"
+    # Each step has its own try/except so we know WHICH step failed.
+    # Pre-2026-04-30 a single bare `except: pass` swallowed every
+    # error including PermissionError on the conf file write and
+    # CalledProcessError from systemctl. Sentinel + auto-escalate
+    # logged "BAN" but the request actually never landed in nginx.
+    # Now each failure mode hits sec_log so the operator sees why.
     try:
-        line = f"deny {ip};\n"
         try:
             with open(_NGINX_BAN_FILE) as f:
                 if line in f.read():
                     return
         except FileNotFoundError:
             pass
+    except PermissionError as e:
+        sec_log.warning("NGINX_BAN_READ_DENIED file=%s err=%s — won't dedup, attempting append anyway", _NGINX_BAN_FILE, e)
+    try:
         with open(_NGINX_BAN_FILE, "a") as f:
             f.write(line)
-        subprocess.run(["systemctl", "reload", "nginx"], capture_output=True, timeout=5)
-    except Exception:
-        pass
+    except PermissionError as e:
+        sec_log.error("NGINX_BAN_WRITE_DENIED ip=%s file=%s err=%s — ban NOT effective at nginx layer", ip, _NGINX_BAN_FILE, e)
+        return
+    except Exception as e:
+        sec_log.error("NGINX_BAN_WRITE_ERROR ip=%s err=%s", ip, e)
+        return
+    try:
+        result = subprocess.run(["systemctl", "reload", "nginx"], capture_output=True, text=True, timeout=5)
+        if result.returncode != 0:
+            sec_log.error("NGINX_RELOAD_FAILED ip=%s rc=%d stderr=%s", ip, result.returncode, result.stderr.strip())
+    except subprocess.TimeoutExpired:
+        sec_log.error("NGINX_RELOAD_TIMEOUT ip=%s — systemctl reload nginx didn't finish in 5s", ip)
+    except FileNotFoundError:
+        sec_log.error("NGINX_RELOAD_NO_SYSTEMCTL ip=%s — systemctl not in PATH for service user", ip)
+    except Exception as e:
+        sec_log.error("NGINX_RELOAD_ERROR ip=%s err=%s", ip, e)
 
 def _nginx_unban(ip):
     """Remove IP from nginx deny list and reload."""
@@ -7552,8 +7750,13 @@ def admin_ban_ip():
     action = data.get("action", "ban")
     if not ip:
         return jsonify({"error": "IP required"}), 400
-    if ip.startswith("192.168."):
-        return jsonify({"error": "Cannot ban LAN addresses"}), 400
+    # Defense in depth (2026-04-30): use the canonical ALLOWLIST_IPS
+    # check rather than a substring on "192.168." which would let
+    # 10.0.0.0/8 LANs and IPv6 loopback ::1 through. Same allowlist
+    # the auto-ban paths now respect — operator can't accidentally
+    # cut off their own LAN gateway.
+    if is_allowlisted(ip):
+        return jsonify({"error": f"refusing to ban allowlisted IP {ip} (in ALLOWLIST_IPS)"}), 400
     try:
         if action == "ban":
             subprocess.run(["fail2ban-client", "set", "llm-team-exploit", "banip", ip],
@@ -7813,7 +8016,10 @@ def admin_mass_ban():
     results = {"success": 0, "failed": 0, "skipped": 0}
     for ip in ip_list:
         ip = ip.strip()
-        if not ip or ip.startswith("192.168."):
+        # 2026-04-30: substring "192.168." → is_allowlisted so all
+        # trusted networks (LAN gateways, IPv6 loopback, custom
+        # entries) are skipped, not just one /16.
+        if not ip or is_allowlisted(ip):
             results["skipped"] += 1
             continue
         try:
@@ -12535,7 +12741,17 @@ SENTINEL_MODEL = "qwen2.5:latest"
 SENTINEL_INTERVAL = 300  # 5 minutes
 _sentinel_last_pos = 0
 _sentinel_results = []  # last 50 analyses
-_sentinel_stats = {"scans": 0, "bans": 0, "last_run": None, "last_error": None, "next_scan_ts": 0}
+_sentinel_stats = {
+    "scans": 0, "bans": 0, "last_run": None, "last_error": None, "next_scan_ts": 0,
+    # 2026-04-30 J: track consecutive AI-query failures so we can
+    # fire a callback (email alert) when Ollama is sustainedly busy
+    # or unreachable. Pre-fix a model-busy state preserved log
+    # position + skipped the scan with no operator notification.
+    "consecutive_ai_failures": 0,
+    "ai_busy_alerted": False,  # one alert per outage; clears on first success
+}
+SENTINEL_AI_FAILURE_ALERT_THRESHOLD = 3  # consecutive failures before email
+SENTINEL_AI_RETRY_DELAY_SECS = 30        # wait before retry inside same scan
 
 def _sentinel_log_entry(msg):
     """Write to sentinel log file."""
@@ -12583,7 +12799,10 @@ def _sentinel_scan():
             if token.startswith("ip="):
                 ip = token[3:]
                 break
-        if ip and not ip.startswith("192.168."):
+        # 2026-04-30: was substring "192.168." — sentinel now skips
+        # ALL allowlisted IPs from analysis (saves tokens + prevents
+        # the AI judge from getting confused by legitimate admin traffic).
+        if ip and not is_allowlisted(ip):
             ip_activity[ip].append(line)
 
     if not ip_activity:
@@ -12648,21 +12867,85 @@ def _sentinel_scan():
     for ip, summary, _ in analysis_items[:15]:  # max 15 IPs per scan
         prompt += summary + "\n"
 
-    # Query local AI
-    try:
-        cfg = load_config()
-        base = cfg["providers"]["ollama"].get("base_url", "http://localhost:11434")
-        resp = requests.post(f"{base}/api/generate", json={
-            "model": SENTINEL_MODEL, "prompt": prompt, "stream": False,
-            "options": {"num_ctx": 4096, "temperature": 0.1}
-        }, timeout=60)
-        resp.raise_for_status()
-        ai_response = resp.json()["response"]
-    except Exception as e:
-        _sentinel_stats["last_error"] = f"AI query failed: {e}"
-        _sentinel_log_entry(f"AI_ERROR error={e}")
+    # Query local AI. 2026-04-30 J fix: retry once on model-busy /
+    # connection / timeout, and fire an operator callback when the
+    # AI is sustainedly unreachable. Pre-fix a single Ollama hiccup
+    # silently dropped the scan with no notification — operator only
+    # discovered the gap by checking sentinel-status manually.
+    cfg = load_config()
+    base = cfg["providers"]["ollama"].get("base_url", "http://localhost:11434")
+    body = {
+        "model": SENTINEL_MODEL, "prompt": prompt, "stream": False,
+        "options": {"num_ctx": 4096, "temperature": 0.1},
+    }
+    ai_response = None
+    last_err = None
+    for attempt in range(2):  # original try + 1 retry
+        try:
+            resp = requests.post(f"{base}/api/generate", json=body, timeout=60)
+            resp.raise_for_status()
+            ai_response = resp.json()["response"]
+            break
+        except (requests.exceptions.ConnectionError,
+                requests.exceptions.Timeout,
+                requests.exceptions.ReadTimeout) as e:
+            last_err = f"connection/timeout: {e}"
+            if attempt == 0:
+                _sentinel_log_entry(f"AI_BUSY_RETRY attempt=1 err={str(e)[:80]} sleeping={SENTINEL_AI_RETRY_DELAY_SECS}s")
+                time.sleep(SENTINEL_AI_RETRY_DELAY_SECS)
+                continue
+        except requests.exceptions.HTTPError as e:
+            # 503 Service Unavailable + 429 Too Many = busy; retry.
+            # Other HTTP errors (404 model missing, 400 bad prompt) won't
+            # recover from a retry, so fail fast.
+            sc = getattr(e.response, "status_code", 0)
+            last_err = f"HTTP {sc}: {e}"
+            if sc in (429, 503) and attempt == 0:
+                _sentinel_log_entry(f"AI_BUSY_RETRY attempt=1 status={sc} sleeping={SENTINEL_AI_RETRY_DELAY_SECS}s")
+                time.sleep(SENTINEL_AI_RETRY_DELAY_SECS)
+                continue
+            break
+        except Exception as e:
+            last_err = f"unexpected: {e}"
+            break
+
+    if ai_response is None:
+        _sentinel_stats["consecutive_ai_failures"] += 1
+        _sentinel_stats["last_error"] = f"AI query failed: {last_err}"
+        _sentinel_log_entry(
+            f"AI_ERROR error={last_err} consecutive={_sentinel_stats['consecutive_ai_failures']}"
+        )
+        # Operator callback: fire a security alert email when the AI
+        # has been down for ≥N consecutive scans. One alert per outage —
+        # cleared on next successful scan so a flapping AI doesn't
+        # spam the inbox.
+        if (_sentinel_stats["consecutive_ai_failures"] >= SENTINEL_AI_FAILURE_ALERT_THRESHOLD
+                and not _sentinel_stats["ai_busy_alerted"]):
+            _sentinel_stats["ai_busy_alerted"] = True
+            try:
+                send_security_alert(
+                    f"Sentinel AI unreachable ({_sentinel_stats['consecutive_ai_failures']} consecutive failures)",
+                    f"The sentinel auto-scanner has been unable to reach the LLM judge for "
+                    f"{_sentinel_stats['consecutive_ai_failures']} consecutive scans.\n\n"
+                    f"Last error: {last_err}\n"
+                    f"Model: {SENTINEL_MODEL}\n"
+                    f"Endpoint: {base}\n\n"
+                    f"Threats are being logged and surfaced in the threat-intel UI but "
+                    f"NOT auto-banned during this outage. Manual review recommended.",
+                )
+            except Exception as alert_err:
+                sec_log.error("SENTINEL_ALERT_SEND_FAILED err=%s", alert_err)
         return
 
+    # AI succeeded. Reset the failure counter + clear the alerted flag
+    # so the next outage gets its own notification.
+    if _sentinel_stats["consecutive_ai_failures"] > 0:
+        _sentinel_log_entry(
+            f"AI_RECOVERED after_failures={_sentinel_stats['consecutive_ai_failures']}"
+        )
+    _sentinel_stats["consecutive_ai_failures"] = 0
+    _sentinel_stats["ai_busy_alerted"] = False
+
     # Parse AI response
     try:
         # Extract JSON from response (handle markdown code blocks)
@@ -12688,9 +12971,41 @@ def _sentinel_scan():
     ban_futures = []
 
     def _execute_ban(ip, threat, reason, attack_type):
-        """Execute a single ban — fail2ban + nginx + kill connections."""
-        subprocess.run(["fail2ban-client", "set", "llm-team-exploit", "banip", ip],
-                       capture_output=True, text=True, timeout=5)
+        """Execute a single ban — fail2ban + nginx + kill connections.
+
+        2026-04-30 J fix: actually examine the fail2ban-client result.
+        Pre-fix capture_output=True was set but the result thrown away,
+        so a non-zero exit (jail not configured, IP already banned, IPv6
+        format quirk) silently said "AI_BAN" in the log while the
+        attacker walked through unimpeded. Now logs returncode + stderr
+        on failure so the operator sees WHY the ban didn't stick."""
+        try:
+            result = subprocess.run(
+                ["fail2ban-client", "set", "llm-team-exploit", "banip", ip],
+                capture_output=True, text=True, timeout=5,
+            )
+            if result.returncode != 0:
+                sec_log.error(
+                    "FAIL2BAN_BAN_FAILED ip=%s rc=%d stdout=%s stderr=%s",
+                    ip, result.returncode,
+                    result.stdout.strip()[:200],
+                    result.stderr.strip()[:200],
+                )
+                _sentinel_log_entry(
+                    f"FAIL2BAN_FAILED ip={ip} rc={result.returncode} "
+                    f"err={result.stderr.strip()[:120]}"
+                )
+                # Continue anyway — nginx layer is independent and may
+                # still take effect.
+        except subprocess.TimeoutExpired:
+            sec_log.error("FAIL2BAN_TIMEOUT ip=%s — client didn't return in 5s", ip)
+            _sentinel_log_entry(f"FAIL2BAN_TIMEOUT ip={ip}")
+        except FileNotFoundError:
+            sec_log.error("FAIL2BAN_NOT_INSTALLED ip=%s — fail2ban-client not in PATH", ip)
+            _sentinel_log_entry(f"FAIL2BAN_NOT_INSTALLED ip={ip}")
+        except Exception as e:
+            sec_log.error("FAIL2BAN_ERROR ip=%s err=%s", ip, e)
+            _sentinel_log_entry(f"FAIL2BAN_ERROR ip={ip} err={e}")
         _nginx_ban(ip)
         _kill_connections(ip)
         sec_log.warning("AI_BAN ip=%s threat=%s reason=%s attack=%s", ip, threat, reason, attack_type)
@@ -12714,7 +13029,11 @@ def _sentinel_scan():
             if len(_sentinel_results) > 50:
                 _sentinel_results.pop(0)
 
-            if action == "ban" and ip and not ip.startswith("192.168."):
+            # 2026-04-30: was substring "192.168." — replaced with
+            # canonical is_allowlisted so the sentinel's AI verdict
+            # can't accidentally ban any allowlisted IP that slipped
+            # past the analysis filter (defense in depth).
+            if action == "ban" and ip and not is_allowlisted(ip):
                 ban_futures.append(executor.submit(_execute_ban, ip, threat, reason, attack_type))
             else:
                 _sentinel_log_entry(f"AI_VERDICT ip={ip} threat={threat} action={action} reason={reason} attack_type={attack_type}")