model-warmth-keeper

model-warmth-keeper

Why this exists

2026-04-28 NIGHT-1 run logged: Ollama qwen3.6/qwen3.5 hit llm-idle-timeout watchdog. Models eviction-time out during idle, next dispatch waits 30+s for warmup, occasionally that warmup itself OOMs because two models race for memory. This skill keeps the primary specialists warm with cheap 1-token pings.

Models pinged

Procedure

Step 1 — Time-of-day guard

import datetime
hour = datetime.datetime.now().hour
# Skip pings during 00:00–05:59 CT (Mac is set to America/Chicago).
# That's "off hours" — pings can be skipped to save battery / power.
if hour < 6:
    print("[warmth-keeper] off-hours, skipping ping cycle")
    raise SystemExit(0)

Step 2 — Ping each model

import urllib.request, json, time, pathlib
LOG = pathlib.Path.home() / ".openclaw/logs/warmth-keeper.log"
LOG.parent.mkdir(parents=True, exist_ok=True)

def log(msg):
    with open(LOG, "a") as f:
        f.write(f"{datetime.datetime.utcnow().isoformat()}Z {msg}\n")

OLLAMA_MODELS = ["qwen3.6:latest", "qwen3.5:latest", "qwen3.5:27B"]
LMSTUDIO_MODELS = ["qwen/qwen3-coder-next"]

def ping_ollama(model):
    body = json.dumps({"model": model, "prompt": "OK", "stream": False, "options": {"num_predict": 1}}).encode()
    req = urllib.request.Request("http://127.0.0.1:11434/api/generate", data=body,
                                 headers={"Content-Type": "application/json"}, method="POST")
    t0 = time.time()
    try:
        with urllib.request.urlopen(req, timeout=15) as r:
            r.read()
        return ("ok", round(time.time() - t0, 3), None)
    except Exception as e:
        return ("fail", round(time.time() - t0, 3), str(e)[:120])

def ping_lmstudio(model):
    body = json.dumps({
        "model": model,
        "messages": [{"role": "user", "content": "OK"}],
        "max_tokens": 1,
        "stream": False
    }).encode()
    req = urllib.request.Request("http://127.0.0.1:1234/v1/chat/completions", data=body,
                                 headers={"Content-Type": "application/json"}, method="POST")
    t0 = time.time()
    try:
        with urllib.request.urlopen(req, timeout=15) as r:
            r.read()
        return ("ok", round(time.time() - t0, 3), None)
    except Exception as e:
        return ("fail", round(time.time() - t0, 3), str(e)[:120])

failures = []
for m in OLLAMA_MODELS:
    status, dt, err = ping_ollama(m)
    log(f"ollama {m} {status} {dt}s {err or ''}")
    if status == "fail":
        failures.append(("ollama", m, err))

for m in LMSTUDIO_MODELS:
    status, dt, err = ping_lmstudio(m)
    log(f"lmstudio {m} {status} {dt}s {err or ''}")
    if status == "fail":
        failures.append(("lmstudio", m, err))