Keldron — GPU & Hardware Monitor Agent

---

name: keldron-agent description: GPU monitoring with risk intelligence. Local + cloud fleet monitoring, health tracking, proactive alerts, and AI-powered fleet analytics. version: 2.0.0 emoji: "🔥" homepage: https://github.com/keldron-ai/keldron-agent metadata: openclaw: requires: bins: - curl - jq anyBins: - docker primaryEnv: "KELDRON_CLOUD_API_KEY"

---

Keldron Agent — GPU Monitoring with Risk Intelligence

1. Overview

Keldron Agent is a vendor-neutral GPU monitoring agent that runs locally and exposes real-time telemetry and risk scores via a Prometheus endpoint. It supports Apple Silicon (M1–M5), NVIDIA consumer GPUs (RTX 3090/4090/5090), NVIDIA datacenter (H100/B200), AMD GPUs, and any Linux machine.

Dual mode: Use the local agent on localhost:9100 for fast, real-time, single-device queries (works offline). Use Keldron Cloud (https://api.keldron.ai) for fleet overview, historical telemetry, analytics, and proactive fleet monitoring when an API key is configured.

No sudo required on any platform for the agent binary. On Linux, Docker may require sudo or membership in the docker group — see Docker post-install or rootless Docker if you hit permission errors.

Use this skill when the user wants to:

Monitor GPU temperature, power, utilization, or memory

Get risk assessments for their GPU

Track power costs

Set up alerts or watch a fleet (via cloud polling)

Open the real dashboard (local or cloud)

Ask fleet, history, or analytics questions

Mode detection (run at the start of an interaction)

Field names differ by endpoint — see Cloud API field names before writing jq filters.

# Check 1: Is the local agent running?
LOCAL_AGENT=$(curl -sf localhost:9100/healthz 2>/dev/null | jq -r '.status' 2>/dev/null)
Check 2: Is cloud configured? (env → ~/.keldron/credentials from login → YAML)
CLOUD_KEY="${KELDRON_CLOUD_API_KEY:-}"
CLOUD_ENDPOINT=""
if [ -z "$CLOUD_KEY" ] && [ -f ~/.keldron/credentials ] && command -v jq &>/dev/null; then
  CLOUD_KEY=$(jq -r '.api_key // ""' ~/.keldron/credentials 2>/dev/null)
  CLOUD_ENDPOINT=$(jq -r '.endpoint // ""' ~/.keldron/credentials 2>/dev/null)
fi
if [ -z "$CLOUD_KEY" ]; then
  if command -v yq &>/dev/null; then
    CLOUD_KEY=$(yq '.cloud.api_key // ""' ~/.config/keldron/keldron-agent.yaml 2>/dev/null)
    CLOUD_ENDPOINT=$(yq '.cloud.endpoint // ""' ~/.config/keldron/keldron-agent.yaml 2>/dev/null)
  else
    CLOUD_KEY=$(grep -A3 'cloud:' ~/.config/keldron/keldron-agent.yaml 2>/dev/null \
      | grep 'api_key:' | awk '{print $2}' | tr -d "\"'" | xargs 2>/dev/null)
    CLOUD_ENDPOINT=$(grep -A3 'cloud:' ~/.config/keldron/keldron-agent.yaml 2>/dev/null \
      | grep 'endpoint:' | awk '{print $2}' | tr -d "\"'" | xargs 2>/dev/null)
  fi
fi
CLOUD_ENDPOINT="${CLOUD_ENDPOINT:-https://api.keldron.ai}"
if [ -z "$CLOUD_KEY" ]; then
  echo "No cloud API key found. Run keldron-agent login, set KELDRON_CLOUD_API_KEY (non-interactive login or streaming), or add cloud.api_key to ~/.config/keldron/keldron-agent.yaml. Sign up at https://app.keldron.ai"
fi
Check 3: Does cloud respond? (only if we have a key)
CLOUD_OK=""
if [ -n "$CLOUD_KEY" ]; then
  CLOUD_OK=$(curl -sf "${CLOUD_ENDPOINT}/health" 2>/dev/null | jq -r '.status' 2>/dev/null)
fi

Mode priority:

Both (healthy local + cloud) → Cloud for fleet, history, analytics, proactive fleet loops; local for instantaneous single-device metrics.

Cloud only → Use cloud for everything that needs the API; mention local agent if they want lower-latency realtime on that machine.

Local only → Use localhost:9100 for realtime; naturally mention cloud for history, fleet, and alerts: *"That needs historical data — connect at app.keldron.ai."*

Neither → Run the Auto-setup flow.

---

2. Installation

Prefer a GitHub release binary (keldron-agent). To build from source, clone the repo and run make build (requires Go and Node.js for the full dashboard).

Mac (Apple Silicon)

curl -sfL https://github.com/keldron-ai/keldron-agent/releases/latest/download/keldron-agent-darwin-arm64 -o keldron-agent
chmod +x keldron-agent

Linux (AMD64)

curl -sfL https://github.com/keldron-ai/keldron-agent/releases/latest/download/keldron-agent-linux-amd64 -o keldron-agent
chmod +x keldron-agent

Linux (ARM64)

curl -sfL https://github.com/keldron-ai/keldron-agent/releases/latest/download/keldron-agent-linux-arm64 -o keldron-agent
chmod +x keldron-agent

Linux (Docker)

docker rm -f keldron-agent 2>/dev/null || true
docker run -d --name keldron-agent --restart unless-stopped \
  -p 9100:9100 -p 9200:9200 -p 8081:8081 \
  -e KELDRON_OUTPUT_PROMETHEUS_HOST=0.0.0.0 \
  -e KELDRON_API_HOST=0.0.0.0 \
  -e KELDRON_HEALTH_BIND=0.0.0.0:8081 \
  ghcr.io/keldron-ai/keldron-agent:latest

Verify installation

./keldron-agent --version

---

3. Auto-setup flow

Trigger phrases: "monitor my hardware", "set up monitoring", "install keldron", "get started", "help me set up"

Step 1: Detect environment

OS=$(uname -s)
if [ "$OS" = "Darwin" ]; then
  CHIP=$(sysctl -n machdep.cpu.brand_string 2>/dev/null || echo "unknown")
  echo "Detected: macOS with $CHIP"
elif command -v nvidia-smi &>/dev/null; then
  GPU=$(nvidia-smi --query-gpu=name --format=csv,noheader 2>/dev/null | head -1)
  echo "Detected: Linux with NVIDIA $GPU"
else
  echo "Detected: Linux (generic thermal monitoring)"
fi

Step 2: Check if agent is running

if curl -sf localhost:9100/healthz | jq -e '.status == "healthy"' &>/dev/null; then
  echo "Keldron agent is already running. Skipping to Step 4."
  exit 0
fi

Step 3: Install agent

Download the release binary for the OS/arch into the current directory, then start in local mode. Example:

ARCH=$(uname -m)
if [ "$OS" = "Darwin" ]; then
  BINARY="keldron-agent-darwin-arm64"
elif [ "$ARCH" = "x86_64" ]; then
  BINARY="keldron-agent-linux-amd64"
else
  BINARY="keldron-agent-linux-arm64"
fi
if [ "$OS" = "Darwin" ]; then
  curl -sfL "https://github.com/keldron-ai/keldron-agent/releases/latest/download/${BINARY}" -o keldron-agent
  chmod +x keldron-agent
  ./keldron-agent --local &
  sleep 3
fi
if [ "$OS" = "Linux" ]; then
  if command -v docker &>/dev/null; then
    docker rm -f keldron-agent 2>/dev/null || true
    if ! docker run -d --name keldron-agent --restart unless-stopped \
      -p 9100:9100 -p 9200:9200 -p 8081:8081 \
      -e KELDRON_OUTPUT_PROMETHEUS_HOST=0.0.0.0 \
      -e KELDRON_API_HOST=0.0.0.0 \
      -e KELDRON_HEALTH_BIND=0.0.0.0:8081 \
      ghcr.io/keldron-ai/keldron-agent:latest; then
      echo "Error: Failed to start keldron-agent container. Check Docker permissions and network."
      exit 1
    fi
  else
    curl -sfL "https://github.com/keldron-ai/keldron-agent/releases/latest/download/${BINARY}" -o keldron-agent
    chmod +x keldron-agent
    ./keldron-agent --local &
  fi
  sleep 3
fi
curl -sf localhost:9100/healthz | jq -e '.status == "healthy"'

Report initial readings (temperature, utilization, risk score) from Quick status queries once healthy.

Step 4: Offer cloud connection

Guide the user conversationally:

1. Account: Ask: *Do you have a Keldron Cloud account? You can sign up free at https://app.keldron.ai* 2. If yes: *Run keldron-agent login — you can use email/password or paste your API key (option 2 in the menu).* 3. If no: *Sign up at https://app.keldron.ai (GitHub login available), then run keldron-agent login.* 4. Verify: *Run keldron-agent whoami to confirm you're connected.* 5. Restart the agent so it picks up credentials and begins streaming.

If they are not yet interested, summarize value:

if [ -n "$CLOUD_KEY" ]; then
  echo "Cloud already connected."
else
  echo "Want to connect to Keldron Cloud?"
  echo "  • 180-day telemetry history"
  echo "  • Fleet analytics and device comparison"
  echo "  • Device health tracking"
  echo "  • Proactive fleet alerts"
  echo "  • Dashboard at app.keldron.ai"
fi

Step 5 optional: env or YAML without CLI login

Prefer Step 4 (keldron-agent login) for normal setup. Use this path when the user cannot run the interactive CLI (CI, containers without TTY) or explicitly wants config-file or env-only configuration. For non-interactive login, set KELDRON_CLOUD_API_KEY or pipe the key via stdin: echo "$KEY" | keldron-agent login.

When setting a key starting with kldn_, use a temporary environment variable — do not echo or paste the full key into commands or transcripts (see Rules).

# Store the user-provided key in a variable (do not inline the raw key)
export CLOUD_KEY=""
mkdir -p ~/.config/keldron
Add or update cloud config in agent YAML
if [ -f ~/.config/keldron/keldron-agent.yaml ]; then
  if ! grep -q 'cloud:' ~/.config/keldron/keldron-agent.yaml; then
    cat >> ~/.config/keldron/keldron-agent.yaml << EOF
cloud:
  enabled: true
  api_key: $CLOUD_KEY
EOF
  else
    # cloud: section exists — update or insert api_key under the cloud: block
    awk -v key="$CLOUD_KEY" '
      /^cloud:/ { in_cloud=1; found=0; print; next }
      in_cloud && /^[^ ]/ {
        if (!found) { print "  api_key: " key; found=1 }
        in_cloud=0
      }
      in_cloud && /^[[:space:]]+api_key:/ { $0="  api_key: " key; found=1 }
      { print }
      END { if (in_cloud && !found) print "  api_key: " key }
    ' ~/.config/keldron/keldron-agent.yaml > ~/.config/keldron/keldron-agent.yaml.tmp \
      && mv ~/.config/keldron/keldron-agent.yaml.tmp ~/.config/keldron/keldron-agent.yaml
  fi
else
  cat > ~/.config/keldron/keldron-agent.yaml << EOF
cloud:
  enabled: true
  api_key: $CLOUD_KEY
EOF
fi
Restart agent
pkill -f './keldron-agent'
sleep 2
./keldron-agent --local &
sleep 5
Verify cloud connection (optional — uses env var, not raw key)
curl -sf "${CLOUD_ENDPOINT}/v1/fleet/overview" \
  -H "X-API-Key: $CLOUD_KEY" | jq '.total_devices'

Tell the user: *Cloud connected. Your device is streaming to Keldron Cloud. Dashboard: https://app.keldron.ai*

---

4. Cloud connection

CLI (primary): keldron-agent login — stores credentials under ~/.keldron/credentials. For non-interactive use: set KELDRON_CLOUD_API_KEY or pipe via stdin.

Environment variable: KELDRON_CLOUD_API_KEY is also read when running the agent for cloud streaming (in addition to non-interactive login).

Config file (alternative): ~/.config/keldron/keldron-agent.yaml under cloud.api_key (see Step 5 optional for automation-only edits).

HTTP header for API calls: X-API-Key: .

Base URL: https://api.keldron.ai

Never store or paste a full API key into this skill file. When confirming configuration, show at most the first 8 characters, e.g. kldn_liv….

Interaction patterns (cloud)

| User says | Skill response | |-----------|------------------| | "connect to cloud" / "set up cloud" | Guide through keldron-agent login (email/password or paste API key); sign up at app.keldron.ai if needed. | | "am I connected to cloud?" | Have them run keldron-agent whoami (or use Check 2 for CLOUD_KEY). | | "log out of cloud" / "disconnect" | Run keldron-agent logout; note the agent falls back to local-only unless KELDRON_CLOUD_API_KEY or YAML still sets a key. | | "how do I get my API key?" | Sign in at app.keldron.ai — your API key is shown in the app. Or run keldron-agent login to authenticate without manually copying into YAML. |

---

5. Running the agent

Start the agent in local mode:

./keldron-agent --local

The agent auto-detects hardware. Basic use needs no config.

Verify it is running:

curl -sf localhost:9100/healthz | jq -e '.status == "healthy"'

A non-zero exit means the agent is not healthy or not running.

Metrics:

curl -s localhost:9100/metrics | grep keldron_

---

6. Endpoints

Local agent (localhost)

| Port | Path | Description | |------|------|-------------| | 9100 | /metrics | Prometheus metrics (all keldron_* gauges) | | 9100 | /healthz | Liveness check (JSON) | | 9200 | / | Local web dashboard (embedded UI) | | 9100 | /api/v1/status | Agent version, device name, active adapters |

Cloud API reference

All cloud routes use TLS, base URL $CLOUD_ENDPOINT (defaults to https://api.keldron.ai), header X-API-Key: $CLOUD_KEY.

| Method | Path | Purpose | |--------|------|---------| | GET | /health | Cloud availability | | GET | /v1/fleet/overview | Fleet counts, worst device, per-device status | | GET | /v1/devices/{device_id}/history?window=12h | Historical points (window: 12h, 24h, …) | | GET | /v1/devices/{device_id}/health?window=24 | Device health (window = integer hours) | | GET | /v1/analytics/fleet-health?window=7d | Fleet analytics + ai_brief (window: 7d, 24h, …) |

Dashboards

Cloud connected: Fleet: https://app.keldron.ai/fleet — Device: https://app.keldron.ai/device/{device_id} — Analytics: https://app.keldron.ai/analytics

Local only: http://localhost:9200 — For fleet analytics and history, connect at https://app.keldron.ai

When the user asks for a dashboard, link these URLs. Do not render ASCII art or fake terminal dashboards.

Cloud API field names (jq reference)

Do not assume names match across endpoints.

| Concept | Fleet overview | History (points) | Device status (e.g. nested current) | Health | |---------|----------------|-------------------|--------------------------------------|--------| | Temperature | temperature_primary | temperature | temperature_primary | idle_temp_c, peak_temp_c | | Power | power_draw | power | power_draw | — | | Composite score | composite_risk_score | composite_score | composite_risk_score | — | | Severity | severity_band | severity | severity_band | — | | Thermal sub | — | thermal_sub | thermal_sub_score | — | | Efficiency | — | — | — | perf_per_watt.value (not perf_per_watt.perf_per_watt) |

Window formats

History: string with unit → ?window=12h, ?window=24h

Health: integer hours → ?window=24

Analytics: string with unit → ?window=7d, ?window=24h

Health endpoint: thermal_stability may be omitted when unavailable (not null). Use jq that tolerates a missing key. Use idle_temp_c / peak_temp_c — not idle_baseline / peak_temp.

---

7. Key metrics reference

| Metric | Description | |--------|-------------| | keldron_gpu_temperature_celsius | GPU temperature in Celsius | | keldron_risk_severity | 0=normal, 1=warning, 2=critical | | keldron_risk_composite | Composite risk score (0–100) | | keldron_risk_thermal | Thermal risk score | | keldron_risk_power | Power risk score | | keldron_risk_volatility | Volatility risk score | | keldron_risk_memory | Memory-related risk score | | keldron_power_cost_monthly | Estimated power cost per month ($) | | keldron_power_cost_daily | Estimated power cost per day ($) | | keldron_power_cost_hourly | Estimated power cost per hour ($) | | keldron_gpu_power_watts | GPU power draw in watts | | keldron_gpu_utilization_ratio | GPU utilization 0–1 | | keldron_gpu_memory_used_bytes | GPU memory in use | | keldron_gpu_memory_total_bytes | GPU memory total | | keldron_gpu_memory_pressure_ratio | Memory pressure 0–1 | | keldron_gpu_throttle_active | 1 if throttled, 0 otherwise | | keldron_system_swap_used_bytes | System swap in use | | keldron_agent_info | Agent metadata (device_model, device_name labels) |

---

8. Quick status queries (local, real-time)

"What's my GPU temperature?"

curl -s localhost:9100/metrics | grep 'keldron_gpu_temperature_celsius{' | awk '{print $2}'

Extract the device_model label from the metric line. Report as: *Your {device_model} is at {value}°C.*

"Is my GPU at risk?"

curl -s localhost:9100/metrics | grep -E 'keldron_risk_(composite|severity|thermal|power|volatility|memory)' | grep -v '^#'

Parse keldron_risk_composite (0–100) and keldron_risk_severity (0=normal, 1=warning, 2=critical). Report composite, severity, and which sub-score is highest.

Assessment thresholds:

| Score | Assessment | |-------|------------| | <30 | Looking good | | 30–60 | Moderate — keep an eye on it | | 60–80 | Warning — consider reducing load | | >80 | Critical — take action now |

"Give me a quick status"

curl -s localhost:9100/metrics | grep -E 'keldron_(gpu_temperature|gpu_utilization|risk_composite|risk_severity|power_cost_monthly|gpu_memory_pressure)' | grep -v '^#'

Format:

🌡️ Temperature: XX°C
⚡ Utilization: XX%
🎯 Risk Score: XX/100 (severity)
💰 Monthly cost: $X.XX
🧠 Memory pressure: XX%

"What GPU do I have?"

curl -s localhost:9100/metrics | grep 'keldron_agent_info'

Extract device_model and device_name from labels. Report: *You're running a {device_model} ({device_name}).*

"How much is my GPU costing me?"

curl -s localhost:9100/metrics | grep 'keldron_power_cost' | grep -v '^#'

Report hourly, daily, and monthly from keldron_power_cost_*.

"How's my memory?"

curl -s localhost:9100/metrics | grep -E 'keldron_gpu_memory|keldron_system_swap' | grep -v '^#'

Derive pressure from keldron_gpu_memory_used_bytes / keldron_gpu_memory_total_bytes. On Apple Silicon, high swap means the workload likely exceeds unified memory — suggest a smaller or quantized model.

---

9. Cloud queries

If cloud is not configured or unreachable, say: *Fleet monitoring requires Keldron Cloud. Sign up at app.keldron.ai to see your whole fleet from anywhere.* For history-only questions without cloud: *I can only see real-time data locally. Connect to Keldron Cloud for historical queries — sign up at app.keldron.ai.*

"How are my machines doing?" / Fleet overview

Use fleet overview field names (composite_risk_score, severity_band, temperature_primary, …).

FLEET=$(curl -s "${CLOUD_ENDPOINT}/v1/fleet/overview" \
  -H "X-API-Key: $CLOUD_KEY")
TOTAL=$(echo "$FLEET" | jq '.total_devices')
NORMAL=$(echo "$FLEET" | jq '.devices_normal')
WARNING=$(echo "$FLEET" | jq '.devices_warning')
CRITICAL=$(echo "$FLEET" | jq '.devices_critical')
WORST=$(echo "$FLEET" | jq -r '.worst_device_id')
WORST_SCORE=$(echo "$FLEET" | jq '.worst_score | floor')

Response style:

All healthy: *All {total} devices healthy. Highest risk is {worst} at {score}. Fleet looks good.*

Any warning: *Heads up — {device} is at Warning (score {score}). {Details}. Other {healthy_count} devices are normal.* Include dashboard: https://app.keldron.ai/fleet

Any critical: *{device} is Critical (score {score}). Immediate attention needed.* Link https://app.keldron.ai/fleet and the device URL if known.

"Show me a dashboard"

Cloud: *Your fleet dashboard: https://app.keldron.ai/fleet — Analytics: https://app.keldron.ai/analytics — Device detail: https://app.keldron.ai/device/{device_id}*

Local only: *Your local dashboard: http://localhost:9200 — Connect to Keldron Cloud at app.keldron.ai for fleet analytics and history.*

"What happened overnight?" / "What happened while I was away?"

History points use temperature, composite_score, severity (not the fleet overview names).

curl -s "${CLOUD_ENDPOINT}/v1/devices/${DEVICE_ID}/history?window=12h" \
  -H "X-API-Key: $CLOUD_KEY" | jq '{
  points: .points | length,
  max_temp: [.points[].temperature] | max,
  min_temp: [.points[].temperature] | min,
  max_score: [.points[].composite_score] | max,
  any_warning: [.points[] | select(.severity != "normal")] | length
}'

Interpret: quiet night vs events; cite min/max temp and whether any non-normal severities occurred.

"Watch for an hour and report" (history)

curl -s "${CLOUD_ENDPOINT}/v1/devices/${DEVICE_ID}/history?window=1h" \
  -H "X-API-Key: $CLOUD_KEY" | jq '{
  points: .points | length,
  max_temp: [.points[].temperature] | max,
  min_temp: [.points[].temperature] | min,
  avg_temp: (if (.points | length) > 0 then ([.points[].temperature] | add / length) else null end),
  max_score: [.points[].composite_score] | max,
  warnings: [.points[] | select(.severity != "normal")] | length
}'

"Compare my devices" / "Which device is the worst?"

curl -s "${CLOUD_ENDPOINT}/v1/analytics/fleet-health?window=7d" \
  -H "X-API-Key: $CLOUD_KEY" | jq '{
  flagged: .total_flagged,
  ai_brief: .ai_brief,
  devices: [.devices[] | {
    name: .hostname,
    score: .current.composite_risk_score,
    recovery: .current.recovery_seconds,
    temp: .current.temperature_primary,
    flags: [.flags[].message]
  }]
}'

Report ai_brief when present — it is a natural-language summary. Link https://app.keldron.ai/analytics.

"How's {device} trending?" / Device health

window is integer hours. Use idle_temp_c, peak_temp_c, perf_per_watt.value. Handle missing thermal_stability.

DEVICE_ENCODED=$(echo "$DEVICE_ID" | jq -sRr @uri)
curl -s "${CLOUD_ENDPOINT}/v1/devices/${DEVICE_ENCODED}/health?window=24" \
  -H "X-API-Key: $CLOUD_KEY" | jq '{
  idle_temp_c: .idle_temp_c,
  peak_temp_c: .peak_temp_c,
  perf_per_watt: .perf_per_watt.value,
  thermal_stability: (if has("thermal_stability") then .thermal_stability else null end)
}'

Report in plain language using idle/peak temps and efficiency. If cloud is down, fall back to local realtime metrics and avoid dumping raw errors.

---

10. Proactive fleet monitoring

Use cloud polling — not localhost:9100 loops — for "watch my fleet" / "alert me if anything changes".

"Watch my fleet" / "Alert me if anything changes"

Set MAX_CHECKS to limit the total number of iterations, including failed API calls (e.g., MAX_CHECKS=60 for ~1 hour). Leave unset or 0 for unlimited.

echo "Fleet monitoring active. Checking every 60 seconds via Keldron Cloud."
echo "Press Ctrl+C to stop."
trap 'echo ""; echo "Fleet monitoring stopped by user."; exit 0' INT
PREV_WORST=""
PREV_WORST_SCORE=0
CHECK_COUNT=0
MAX_CHECKS="${MAX_CHECKS:-0}"
while true; do
  CHECK_COUNT=$((CHECK_COUNT + 1))
  FLEET=$(curl -s "${CLOUD_ENDPOINT}/v1/fleet/overview" \
    -H "X-API-Key: $CLOUD_KEY" 2>/dev/null)
  if [ -z "$FLEET" ]; then
    if [ "$MAX_CHECKS" -gt 0 ] && [ "$CHECK_COUNT" -ge "$MAX_CHECKS" ]; then
      echo "Reached $MAX_CHECKS checks. Fleet monitoring stopped by timeout."
      break
    fi
    sleep 60
    continue
  fi
  WARNING=$(echo "$FLEET" | jq '.devices_warning // 0')
  CRITICAL=$(echo "$FLEET" | jq '.devices_critical // 0')
  WORST=$(echo "$FLEET" | jq -r '.worst_device_id')
  WORST_SCORE=$(echo "$FLEET" | jq '.worst_score // 0 | floor')
  if [ "${CRITICAL:-0}" -gt 0 ]; then
    DEVICE_INFO=$(echo "$FLEET" | jq -r '.devices[] | select(.severity_band == "critical") | "\(.hostname): score \((.composite_risk_score // 0) | floor), temp \((.temperature_primary // 0) | floor)°C"')
    echo "CRITICAL: $DEVICE_INFO"
    break
  fi
  if [ "${WARNING:-0}" -gt 0 ]; then
    DEVICE_INFO=$(echo "$FLEET" | jq -r '.devices[] | select(.severity_band == "warning") | "\(.hostname): score \((.composite_risk_score // 0) | floor), temp \((.temperature_primary // 0) | floor)°C"')
    echo "WARNING: $DEVICE_INFO"
  fi
  if [ -n "$PREV_WORST" ] && [ "$WORST" = "$PREV_WORST" ]; then
    SCORE_DELTA=$((WORST_SCORE - PREV_WORST_SCORE))
    if [ "$SCORE_DELTA" -gt 20 ]; then
      echo "SPIKE: $WORST jumped from $PREV_WORST_SCORE to $WORST_SCORE"
    fi
  fi
  PREV_WORST="$WORST"
  PREV_WORST_SCORE=$WORST_SCORE
  if [ "$MAX_CHECKS" -gt 0 ] && [ "$CHECK_COUNT" -ge "$MAX_CHECKS" ]; then
    echo "Reached $MAX_CHECKS checks. Fleet monitoring stopped by timeout."
    break
  fi
  sleep 60
done

"Give me a morning report"

ANALYTICS=$(curl -s "${CLOUD_ENDPOINT}/v1/analytics/fleet-health?window=24h" \
  -H "X-API-Key: $CLOUD_KEY")
FLEET=$(curl -s "${CLOUD_ENDPOINT}/v1/fleet/overview" \
  -H "X-API-Key: $CLOUD_KEY")
TOTAL=$(echo "$FLEET" | jq '.total_devices')
NORMAL=$(echo "$FLEET" | jq '.devices_normal')
AI_BRIEF=$(echo "$ANALYTICS" | jq -r '.ai_brief')
FLAGGED=$(echo "$ANALYTICS" | jq '.total_flagged')
echo "Morning Fleet Report"
echo "--------------------"
echo "Fleet: $TOTAL devices, $NORMAL healthy"
echo ""
echo "$AI_BRIEF"
echo ""
if [ "$FLAGGED" -gt 0 ]; then
  echo "$FLAGGED device(s) flagged — https://app.keldron.ai/analytics"
else
  echo "All devices healthy."
fi
echo ""
echo "Dashboard: https://app.keldron.ai"

---

11. Configuration & management

Changing config

If the user needs to change settings (e.g. electricity rate), tell them: Edit ~/.config/keldron/keldron-agent.yaml — the agent picks up changes on restart. Avoid complex sed one-liners for YAML edits; prefer manual edits, YAML-aware tools like yq, or scoped awk scripts.

Multi-device / fleet

All devices that stream to the same cloud account appear in the fleet automatically.

Stop monitoring

pkill -f './keldron-agent'

Confirm: *Agent stopped. GPU monitoring is off.*

Restart the agent

pkill -f './keldron-agent'
sleep 2
./keldron-agent --local &
sleep 3
curl -s localhost:9100/healthz

Report the healthz response to confirm it is up.

---

12. Proactive behavior

After setup, suggest cloud and dashboards: *Your agent is running. Ask how your fleet is doing, open https://app.keldron.ai/fleet, or use local metrics on this machine.*

For alert / watch on a fleet, run the cloud polling loop in §10; for single-machine realtime questions, use §8.

---

13. Rules

Always check agent health first. Before any local query, verify: curl -sf localhost:9100/healthz | jq -e '.status == "healthy"'. Non-zero exit = agent down. Offer to start it or guide setup.

Auto-detect mode. Check for cloud API key: KELDRON_CLOUD_API_KEY, then ~/.keldron/credentials (from keldron-agent login), then cloud.api_key in ~/.config/keldron/keldron-agent.yaml. Optionally confirm with keldron-agent whoami. Use cloud for fleet, history, analytics; local for realtime single-device.

Guide setup, don't just explain. When the agent or cloud is not configured, walk through steps and run commands.

If metrics return 0 for temperature, the agent may still be warming up. Wait 30 seconds and retry once.

Always include severity assessment. Report severity (normal/warning/critical) alongside numeric scores where applicable.

Cloud upsell is helpful, not pushy. When a question needs cloud (history, fleet comparison), mention it naturally: *That needs historical data — connect at app.keldron.ai.*

Proactive monitoring uses the cloud API. Background poll loops use api.keldron.ai, not localhost metrics.

Link to dashboards. Include relevant links: app.keldron.ai/fleet, app.keldron.ai/analytics, or http://localhost:9200 for local.

API key security. Never echo the full API key. Show only the first 8 characters: *Configured with key kldn_liv…*

Fallback gracefully. Cloud unreachable → fall back to local. Local unreachable → guide setup. Never show raw errors to the user.

On Apple Silicon, high swap = model too large. If swap is high, suggest a smaller or quantized model.

The agent never requires sudo. Runs unprivileged on all platforms. Docker on Linux may need sudo or docker-group.

Use metric labels. Extract device_model and device_name from Prometheus labels for personalized responses.

When the user says "alert me" or "watch" (fleet), set up the cloud polling loop and execute it.

When the user says "dashboard", link to the real dashboard. Do not render ASCII art.