[claude] Hermes health monitor — system resources + model management (#1073) (#1133)

Co-authored-by: Claude (Opus 4.6) <claude@hermes.local>
Co-committed-by: Claude (Opus 4.6) <claude@hermes.local>

src/config.py

@@ -396,6 +396,16 @@ class Settings(BaseSettings):
     # Default timeout for git operations.
     hands_git_timeout: int = 60
 
+    # ── Hermes Health Monitor ─────────────────────────────────────────
+    # Enable the Hermes system health monitor (memory, disk, Ollama, processes, network).
+    hermes_enabled: bool = True
+    # How often Hermes runs a full health cycle (seconds). Default: 5 minutes.
+    hermes_interval_seconds: int = 300
+    # Alert threshold: free memory below this triggers model unloading / alert (GB).
+    hermes_memory_free_min_gb: float = 4.0
+    # Alert threshold: free disk below this triggers cleanup / alert (GB).
+    hermes_disk_free_min_gb: float = 10.0
+
     # ── Error Logging ─────────────────────────────────────────────────
     error_log_enabled: bool = True
     error_log_dir: str = "logs"

src/dashboard/app.py

@@ -33,6 +33,7 @@ from dashboard.routes.calm import router as calm_router
 from dashboard.routes.chat_api import router as chat_api_router
 from dashboard.routes.chat_api_v1 import router as chat_api_v1_router
 from dashboard.routes.daily_run import router as daily_run_router
+from dashboard.routes.hermes import router as hermes_router
 from dashboard.routes.db_explorer import router as db_explorer_router
 from dashboard.routes.discord import router as discord_router
 from dashboard.routes.experiments import router as experiments_router
@@ -180,6 +181,33 @@ async def _thinking_scheduler() -> None:
         await asyncio.sleep(settings.thinking_interval_seconds)
 
 
+async def _hermes_scheduler() -> None:
+    """Background task: Hermes system health monitor, runs every 5 minutes.
+
+    Checks memory, disk, Ollama, processes, and network.
+    Auto-resolves what it can; fires push notifications when human help is needed.
+    """
+    from infrastructure.hermes.monitor import hermes_monitor
+
+    await asyncio.sleep(20)  # Stagger after other schedulers
+
+    while True:
+        try:
+            if settings.hermes_enabled:
+                report = await hermes_monitor.run_cycle()
+                if report.has_issues:
+                    logger.warning(
+                        "Hermes health issues detected — overall: %s",
+                        report.overall.value,
+                    )
+        except asyncio.CancelledError:
+            raise
+        except Exception as exc:
+            logger.error("Hermes scheduler error: %s", exc)
+
+        await asyncio.sleep(settings.hermes_interval_seconds)
+
+
 async def _loop_qa_scheduler() -> None:
     """Background task: run capability self-tests on a separate timer.
 
@@ -381,6 +409,7 @@ def _startup_background_tasks() -> list[asyncio.Task]:
         asyncio.create_task(_loop_qa_scheduler()),
         asyncio.create_task(_presence_watcher()),
         asyncio.create_task(_start_chat_integrations_background()),
+        asyncio.create_task(_hermes_scheduler()),
     ]
     try:
         from timmy.paperclip import start_paperclip_poller
@@ -639,6 +668,7 @@ app.include_router(world_router)
 app.include_router(matrix_router)
 app.include_router(tower_router)
 app.include_router(daily_run_router)
+app.include_router(hermes_router)
 app.include_router(quests_router)
 app.include_router(scorecards_router)
 app.include_router(sovereignty_metrics_router)

src/dashboard/routes/hermes.py

@@ -0,0 +1,45 @@
+"""Hermes health monitor routes.
+
+Exposes the Hermes health monitor via REST API so the dashboard
+and external tools can query system status and trigger checks.
+
+Refs: #1073
+"""
+
+import logging
+
+from fastapi import APIRouter
+
+from infrastructure.hermes.monitor import hermes_monitor
+
+logger = logging.getLogger(__name__)
+
+router = APIRouter(prefix="/hermes", tags=["hermes"])
+
+
+@router.get("/status")
+async def hermes_status():
+    """Return the most recent Hermes health report.
+
+    Returns the cached result from the last background cycle — does not
+    trigger a new check.  Use POST /hermes/check to run an immediate check.
+    """
+    report = hermes_monitor.last_report
+    if report is None:
+        return {
+            "status": "no_data",
+            "message": "No health report yet — first cycle pending",
+            "seconds_since_last_run": hermes_monitor.seconds_since_last_run,
+        }
+    return report.to_dict()
+
+
+@router.post("/check")
+async def hermes_check():
+    """Trigger an immediate Hermes health check cycle.
+
+    Runs all monitors synchronously and returns the full report.
+    Use sparingly — this blocks until all checks complete (~5 seconds).
+    """
+    report = await hermes_monitor.run_cycle()
+    return report.to_dict()

src/infrastructure/hermes/__init__.py

@@ -0,0 +1,9 @@
+"""Hermes health monitor — system resources + model management.
+
+Monitors the local machine (Hermes/M3 Max) for memory pressure, disk usage,
+Ollama model health, zombie processes, and network connectivity.
+"""
+
+from infrastructure.hermes.monitor import HermesMonitor, HealthLevel, HealthReport, hermes_monitor
+
+__all__ = ["HermesMonitor", "HealthLevel", "HealthReport", "hermes_monitor"]

src/infrastructure/hermes/monitor.py

@@ -0,0 +1,668 @@
+"""Hermes health monitor — system resources + model management.
+
+Monitors the local machine (Hermes/M3 Max) and keeps it running smoothly.
+Runs every 5 minutes, auto-resolves issues where possible, alerts when
+human intervention is needed.
+
+Monitors:
+  1. Memory pressure — unified memory, alert if <4GB free, unload models
+  2. Disk usage — alert if <10GB free, clean temp files
+  3. Ollama status — verify reachable, restart if crashed, manage loaded models
+  4. Process health — detect zombie processes
+  5. Network — verify Gitea connectivity
+
+Refs: #1073
+"""
+
+import asyncio
+import json
+import logging
+import shutil
+import subprocess
+import time
+import urllib.request
+from dataclasses import dataclass, field
+from datetime import UTC, datetime
+from enum import Enum
+from typing import Any
+
+from config import settings
+
+logger = logging.getLogger(__name__)
+
+
+class HealthLevel(str, Enum):
+    """Severity level for a health check result."""
+
+    OK = "ok"
+    WARNING = "warning"
+    CRITICAL = "critical"
+    UNKNOWN = "unknown"
+
+
+@dataclass
+class CheckResult:
+    """Result of a single health check."""
+
+    name: str
+    level: HealthLevel
+    message: str
+    details: dict[str, Any] = field(default_factory=dict)
+    auto_resolved: bool = False
+    needs_human: bool = False
+
+    def to_dict(self) -> dict[str, Any]:
+        return {
+            "name": self.name,
+            "level": self.level.value,
+            "message": self.message,
+            "details": self.details,
+            "auto_resolved": self.auto_resolved,
+            "needs_human": self.needs_human,
+        }
+
+
+@dataclass
+class HealthReport:
+    """Full health report from a single monitor cycle."""
+
+    timestamp: str
+    checks: list[CheckResult]
+    overall: HealthLevel
+
+    @property
+    def has_issues(self) -> bool:
+        return any(c.level != HealthLevel.OK for c in self.checks)
+
+    def to_dict(self) -> dict[str, Any]:
+        return {
+            "timestamp": self.timestamp,
+            "overall": self.overall.value,
+            "has_issues": self.has_issues,
+            "checks": [c.to_dict() for c in self.checks],
+        }
+
+
+class HermesMonitor:
+    """System health monitor for Hermes (local M3 Max machine).
+
+    All blocking I/O (subprocess, HTTP) is wrapped in asyncio.to_thread()
+    so it never blocks the event loop.  Results are cached so the dashboard
+    can read the last report without triggering a new cycle.
+    """
+
+    OLLAMA_REQUEST_TIMEOUT = 5
+    NETWORK_REQUEST_TIMEOUT = 5
+
+    def __init__(self) -> None:
+        self._last_report: HealthReport | None = None
+        self._last_run_ts: float = 0.0
+
+    @property
+    def last_report(self) -> HealthReport | None:
+        """Most recent health report, or None if no cycle has run yet."""
+        return self._last_report
+
+    @property
+    def seconds_since_last_run(self) -> float:
+        if self._last_run_ts == 0.0:
+            return float("inf")
+        return time.monotonic() - self._last_run_ts
+
+    async def run_cycle(self) -> HealthReport:
+        """Run a full health check cycle and return the report."""
+        self._last_run_ts = time.monotonic()
+        logger.info("Hermes health cycle starting")
+
+        check_fns = [
+            self._check_memory(),
+            self._check_disk(),
+            self._check_ollama(),
+            self._check_processes(),
+            self._check_network(),
+        ]
+
+        raw_results = await asyncio.gather(*check_fns, return_exceptions=True)
+
+        checks: list[CheckResult] = []
+        for i, r in enumerate(raw_results):
+            if isinstance(r, Exception):
+                name = ["memory", "disk", "ollama", "processes", "network"][i]
+                logger.warning("Hermes check '%s' raised: %s", name, r)
+                checks.append(
+                    CheckResult(
+                        name=name,
+                        level=HealthLevel.UNKNOWN,
+                        message=f"Check error: {r}",
+                    )
+                )
+            else:
+                checks.append(r)
+
+        # Compute overall level
+        levels = {c.level for c in checks}
+        if HealthLevel.CRITICAL in levels:
+            overall = HealthLevel.CRITICAL
+        elif HealthLevel.WARNING in levels:
+            overall = HealthLevel.WARNING
+        elif HealthLevel.UNKNOWN in levels:
+            overall = HealthLevel.UNKNOWN
+        else:
+            overall = HealthLevel.OK
+
+        report = HealthReport(
+            timestamp=datetime.now(UTC).isoformat(),
+            checks=checks,
+            overall=overall,
+        )
+        self._last_report = report
+
+        await self._handle_alerts(report)
+
+        logger.info("Hermes health cycle complete — overall: %s", overall.value)
+        return report
+
+    # ── Memory ───────────────────────────────────────────────────────────────
+
+    async def _check_memory(self) -> CheckResult:
+        """Check unified memory usage (macOS vm_stat)."""
+        memory_free_min_gb = getattr(settings, "hermes_memory_free_min_gb", 4.0)
+        try:
+            info = await asyncio.to_thread(self._get_memory_info)
+            free_gb = info.get("free_gb", 0.0)
+            total_gb = info.get("total_gb", 0.0)
+            details: dict[str, Any] = {
+                "free_gb": round(free_gb, 2),
+                "total_gb": round(total_gb, 2),
+            }
+
+            if free_gb < memory_free_min_gb:
+                # Attempt auto-remediation: unload Ollama models
+                unloaded = await self._unload_ollama_models()
+                if unloaded:
+                    return CheckResult(
+                        name="memory",
+                        level=HealthLevel.WARNING,
+                        message=(
+                            f"Low memory ({free_gb:.1f}GB free) — "
+                            f"unloaded {unloaded} Ollama model(s)"
+                        ),
+                        details={**details, "models_unloaded": unloaded},
+                        auto_resolved=True,
+                    )
+                return CheckResult(
+                    name="memory",
+                    level=HealthLevel.CRITICAL,
+                    message=(
+                        f"Critical: only {free_gb:.1f}GB free "
+                        f"(threshold: {memory_free_min_gb}GB)"
+                    ),
+                    details=details,
+                    needs_human=True,
+                )
+
+            return CheckResult(
+                name="memory",
+                level=HealthLevel.OK,
+                message=f"Memory OK — {free_gb:.1f}GB free of {total_gb:.1f}GB",
+                details=details,
+            )
+        except Exception as exc:
+            logger.warning("Memory check failed: %s", exc)
+            return CheckResult(
+                name="memory",
+                level=HealthLevel.UNKNOWN,
+                message=f"Memory check unavailable: {exc}",
+            )
+
+    def _get_memory_info(self) -> dict[str, float]:
+        """Get memory stats via macOS sysctl + vm_stat.
+
+        Falls back gracefully on non-macOS systems.
+        """
+        gb = 1024**3
+        total_bytes = 0.0
+        free_bytes = 0.0
+
+        # Total memory via sysctl
+        try:
+            result = subprocess.run(
+                ["sysctl", "-n", "hw.memsize"],
+                capture_output=True,
+                text=True,
+                timeout=3,
+            )
+            total_bytes = float(result.stdout.strip())
+        except Exception:
+            pass
+
+        # Free + inactive pages via vm_stat (macOS)
+        try:
+            result = subprocess.run(
+                ["vm_stat"],
+                capture_output=True,
+                text=True,
+                timeout=3,
+            )
+            page_size = 16384  # 16 KB default on Apple Silicon
+            for line in result.stdout.splitlines():
+                if "page size of" in line:
+                    parts = line.split()
+                    for i, part in enumerate(parts):
+                        if part == "of" and i + 1 < len(parts):
+                            try:
+                                page_size = int(parts[i + 1])
+                            except ValueError:
+                                pass
+                elif "Pages free:" in line:
+                    pages = int(line.split(":")[1].strip().rstrip("."))
+                    free_bytes += pages * page_size
+                elif "Pages inactive:" in line:
+                    pages = int(line.split(":")[1].strip().rstrip("."))
+                    free_bytes += pages * page_size
+        except Exception:
+            pass
+
+        return {
+            "total_gb": total_bytes / gb if total_bytes else 0.0,
+            "free_gb": free_bytes / gb if free_bytes else 0.0,
+        }
+
+    # ── Disk ─────────────────────────────────────────────────────────────────
+
+    async def _check_disk(self) -> CheckResult:
+        """Check disk usage via shutil.disk_usage."""
+        disk_free_min_gb = getattr(settings, "hermes_disk_free_min_gb", 10.0)
+        try:
+            usage = await asyncio.to_thread(shutil.disk_usage, "/")
+            free_gb = usage.free / (1024**3)
+            total_gb = usage.total / (1024**3)
+            used_pct = (usage.used / usage.total) * 100
+
+            details: dict[str, Any] = {
+                "free_gb": round(free_gb, 2),
+                "total_gb": round(total_gb, 2),
+                "used_pct": round(used_pct, 1),
+            }
+
+            if free_gb < disk_free_min_gb:
+                cleaned_gb = await self._cleanup_temp_files()
+                if cleaned_gb > 0.01:
+                    return CheckResult(
+                        name="disk",
+                        level=HealthLevel.WARNING,
+                        message=(
+                            f"Low disk ({free_gb:.1f}GB free) — "
+                            f"cleaned {cleaned_gb:.2f}GB from /tmp"
+                        ),
+                        details={**details, "cleaned_gb": round(cleaned_gb, 2)},
+                        auto_resolved=True,
+                    )
+                return CheckResult(
+                    name="disk",
+                    level=HealthLevel.CRITICAL,
+                    message=(
+                        f"Critical: only {free_gb:.1f}GB free "
+                        f"(threshold: {disk_free_min_gb}GB)"
+                    ),
+                    details=details,
+                    needs_human=True,
+                )
+
+            return CheckResult(
+                name="disk",
+                level=HealthLevel.OK,
+                message=f"Disk OK — {free_gb:.1f}GB free ({used_pct:.0f}% used)",
+                details=details,
+            )
+        except Exception as exc:
+            logger.warning("Disk check failed: %s", exc)
+            return CheckResult(
+                name="disk",
+                level=HealthLevel.UNKNOWN,
+                message=f"Disk check unavailable: {exc}",
+            )
+
+    async def _cleanup_temp_files(self) -> float:
+        """Remove /tmp files older than 24 hours. Returns GB freed."""
+        return await asyncio.to_thread(self._cleanup_temp_files_sync)
+
+    def _cleanup_temp_files_sync(self) -> float:
+        """Synchronous /tmp cleanup — only touches files older than 24 hours."""
+        from pathlib import Path
+
+        freed_bytes = 0
+        cutoff = time.time() - 86400  # 24 hours ago
+
+        try:
+            tmp = Path("/tmp")
+            for item in tmp.iterdir():
+                try:
+                    stat = item.stat()
+                    if stat.st_mtime >= cutoff:
+                        continue
+                    if item.is_file():
+                        freed_bytes += stat.st_size
+                        item.unlink(missing_ok=True)
+                    elif item.is_dir():
+                        dir_size = sum(
+                            f.stat().st_size
+                            for f in item.rglob("*")
+                            if f.is_file()
+                        )
+                        freed_bytes += dir_size
+                        shutil.rmtree(str(item), ignore_errors=True)
+                except (PermissionError, OSError):
+                    pass  # Skip files we can't touch
+        except Exception as exc:
+            logger.warning("Temp cleanup error: %s", exc)
+
+        freed_gb = freed_bytes / (1024**3)
+        if freed_gb > 0.001:
+            logger.info("Hermes disk cleanup: freed %.2fGB from /tmp", freed_gb)
+        return freed_gb
+
+    # ── Ollama ───────────────────────────────────────────────────────────────
+
+    async def _check_ollama(self) -> CheckResult:
+        """Check Ollama status and loaded models."""
+        try:
+            status = await asyncio.to_thread(self._get_ollama_status)
+
+            if not status.get("reachable"):
+                restarted = await self._restart_ollama()
+                if restarted:
+                    return CheckResult(
+                        name="ollama",
+                        level=HealthLevel.WARNING,
+                        message="Ollama was unreachable — restart initiated",
+                        details={"restart_attempted": True},
+                        auto_resolved=True,
+                    )
+                return CheckResult(
+                    name="ollama",
+                    level=HealthLevel.CRITICAL,
+                    message="Ollama unreachable and restart failed",
+                    details={"reachable": False},
+                    needs_human=True,
+                )
+
+            models = status.get("models", [])
+            loaded = status.get("loaded_models", [])
+            return CheckResult(
+                name="ollama",
+                level=HealthLevel.OK,
+                message=(
+                    f"Ollama OK — {len(models)} model(s) available, "
+                    f"{len(loaded)} loaded"
+                ),
+                details={
+                    "reachable": True,
+                    "model_count": len(models),
+                    "loaded_count": len(loaded),
+                    "loaded_models": [m.get("name", "") for m in loaded],
+                },
+            )
+        except Exception as exc:
+            logger.warning("Ollama check failed: %s", exc)
+            return CheckResult(
+                name="ollama",
+                level=HealthLevel.UNKNOWN,
+                message=f"Ollama check failed: {exc}",
+            )
+
+    def _get_ollama_status(self) -> dict[str, Any]:
+        """Synchronous Ollama status — checks /api/tags and /api/ps."""
+        url = settings.normalized_ollama_url
+
+        try:
+            req = urllib.request.Request(
+                f"{url}/api/tags",
+                method="GET",
+                headers={"Accept": "application/json"},
+            )
+            with urllib.request.urlopen(req, timeout=self.OLLAMA_REQUEST_TIMEOUT) as resp:
+                data = json.loads(resp.read().decode())
+                models = data.get("models", [])
+        except Exception:
+            return {"reachable": False, "models": [], "loaded_models": []}
+
+        # /api/ps lists currently loaded (in-memory) models — Ollama >=0.2
+        loaded: list[dict] = []
+        try:
+            req = urllib.request.Request(
+                f"{url}/api/ps",
+                method="GET",
+                headers={"Accept": "application/json"},
+            )
+            with urllib.request.urlopen(req, timeout=self.OLLAMA_REQUEST_TIMEOUT) as resp:
+                ps_data = json.loads(resp.read().decode())
+                loaded = ps_data.get("models", [])
+        except Exception:
+            pass  # /api/ps absent on older Ollama — non-fatal
+
+        return {"reachable": True, "models": models, "loaded_models": loaded}
+
+    async def _unload_ollama_models(self) -> int:
+        """Unload in-memory Ollama models to free unified memory.
+
+        Uses the keep_alive=0 trick: POSTing to /api/generate with
+        keep_alive=0 causes Ollama to immediately evict the model.
+        Returns the number of models successfully unloaded.
+        """
+        return await asyncio.to_thread(self._unload_ollama_models_sync)
+
+    def _unload_ollama_models_sync(self) -> int:
+        """Synchronous model unload implementation."""
+        url = settings.normalized_ollama_url
+        unloaded = 0
+
+        try:
+            req = urllib.request.Request(
+                f"{url}/api/ps",
+                method="GET",
+                headers={"Accept": "application/json"},
+            )
+            with urllib.request.urlopen(req, timeout=self.OLLAMA_REQUEST_TIMEOUT) as resp:
+                ps_data = json.loads(resp.read().decode())
+                loaded = ps_data.get("models", [])
+        except Exception:
+            return 0
+
+        for model in loaded:
+            name = model.get("name", "")
+            if not name:
+                continue
+            try:
+                payload = json.dumps({"model": name, "keep_alive": 0}).encode()
+                req = urllib.request.Request(
+                    f"{url}/api/generate",
+                    data=payload,
+                    method="POST",
+                    headers={"Content-Type": "application/json"},
+                )
+                with urllib.request.urlopen(req, timeout=10) as _:
+                    pass
+                logger.info("Hermes: unloaded Ollama model %s", name)
+                unloaded += 1
+            except Exception as exc:
+                logger.warning("Hermes: failed to unload model %s: %s", name, exc)
+
+        return unloaded
+
+    async def _restart_ollama(self) -> bool:
+        """Attempt to restart the Ollama service via launchctl or brew."""
+        return await asyncio.to_thread(self._restart_ollama_sync)
+
+    def _restart_ollama_sync(self) -> bool:
+        """Try launchctl first, then brew services."""
+        # macOS launchctl (installed via official Ollama installer)
+        try:
+            result = subprocess.run(
+                ["launchctl", "stop", "com.ollama.ollama"],
+                capture_output=True,
+                timeout=10,
+            )
+            if result.returncode == 0:
+                time.sleep(2)
+                subprocess.run(
+                    ["launchctl", "start", "com.ollama.ollama"],
+                    capture_output=True,
+                    timeout=10,
+                )
+                logger.info("Hermes: Ollama restarted via launchctl")
+                return True
+        except Exception:
+            pass
+
+        # Homebrew fallback
+        try:
+            result = subprocess.run(
+                ["brew", "services", "restart", "ollama"],
+                capture_output=True,
+                timeout=20,
+            )
+            if result.returncode == 0:
+                logger.info("Hermes: Ollama restarted via brew services")
+                return True
+        except Exception:
+            pass
+
+        logger.warning("Hermes: Ollama restart failed — manual intervention needed")
+        return False
+
+    # ── Processes ────────────────────────────────────────────────────────────
+
+    async def _check_processes(self) -> CheckResult:
+        """Check for zombie processes via ps aux."""
+        try:
+            result = await asyncio.to_thread(self._get_zombie_processes)
+            zombies = result.get("zombies", [])
+
+            if zombies:
+                return CheckResult(
+                    name="processes",
+                    level=HealthLevel.WARNING,
+                    message=f"Found {len(zombies)} zombie process(es)",
+                    details={"zombies": zombies[:5]},
+                    needs_human=len(zombies) > 3,
+                )
+
+            return CheckResult(
+                name="processes",
+                level=HealthLevel.OK,
+                message="Processes OK — no zombies detected",
+                details={"zombie_count": 0},
+            )
+        except Exception as exc:
+            logger.warning("Process check failed: %s", exc)
+            return CheckResult(
+                name="processes",
+                level=HealthLevel.UNKNOWN,
+                message=f"Process check unavailable: {exc}",
+            )
+
+    def _get_zombie_processes(self) -> dict[str, Any]:
+        """Detect zombie processes (state 'Z') via ps aux."""
+        result = subprocess.run(
+            ["ps", "aux"],
+            capture_output=True,
+            text=True,
+            timeout=5,
+        )
+        zombies = []
+        for line in result.stdout.splitlines()[1:]:  # Skip header row
+            parts = line.split(None, 10)
+            if len(parts) >= 8 and parts[7] == "Z":
+                zombies.append(
+                    {
+                        "pid": parts[1],
+                        "command": parts[10][:80] if len(parts) > 10 else "",
+                    }
+                )
+        return {"zombies": zombies}
+
+    # ── Network ──────────────────────────────────────────────────────────────
+
+    async def _check_network(self) -> CheckResult:
+        """Check Gitea connectivity."""
+        try:
+            result = await asyncio.to_thread(self._check_gitea_connectivity)
+            reachable = result.get("reachable", False)
+            latency_ms = result.get("latency_ms", -1.0)
+
+            if not reachable:
+                return CheckResult(
+                    name="network",
+                    level=HealthLevel.WARNING,
+                    message=f"Gitea unreachable: {result.get('error', 'unknown')}",
+                    details=result,
+                    needs_human=True,
+                )
+
+            return CheckResult(
+                name="network",
+                level=HealthLevel.OK,
+                message=f"Network OK — Gitea reachable ({latency_ms:.0f}ms)",
+                details=result,
+            )
+        except Exception as exc:
+            logger.warning("Network check failed: %s", exc)
+            return CheckResult(
+                name="network",
+                level=HealthLevel.UNKNOWN,
+                message=f"Network check unavailable: {exc}",
+            )
+
+    def _check_gitea_connectivity(self) -> dict[str, Any]:
+        """Synchronous Gitea reachability check."""
+        url = settings.gitea_url
+        start = time.monotonic()
+        try:
+            req = urllib.request.Request(
+                f"{url}/api/v1/version",
+                method="GET",
+                headers={"Accept": "application/json"},
+            )
+            with urllib.request.urlopen(req, timeout=self.NETWORK_REQUEST_TIMEOUT) as resp:
+                latency_ms = (time.monotonic() - start) * 1000
+                return {
+                    "reachable": resp.status == 200,
+                    "latency_ms": round(latency_ms, 1),
+                    "url": url,
+                }
+        except Exception as exc:
+            return {
+                "reachable": False,
+                "error": str(exc),
+                "url": url,
+                "latency_ms": -1.0,
+            }
+
+    # ── Alerts ───────────────────────────────────────────────────────────────
+
+    async def _handle_alerts(self, report: HealthReport) -> None:
+        """Send push notifications for issues that need attention."""
+        try:
+            from infrastructure.notifications.push import notifier
+        except Exception:
+            return
+
+        for check in report.checks:
+            if check.level == HealthLevel.CRITICAL or check.needs_human:
+                notifier.notify(
+                    title=f"Hermes Alert: {check.name}",
+                    message=check.message,
+                    category="system",
+                    native=check.level == HealthLevel.CRITICAL,
+                )
+            elif check.level == HealthLevel.WARNING and check.auto_resolved:
+                notifier.notify(
+                    title=f"Hermes: {check.name} auto-fixed",
+                    message=check.message,
+                    category="system",
+                )
+
+
+# Module-level singleton
+hermes_monitor = HermesMonitor()