@wendellresearch-ai/benchwarmers 0.2.0 → 0.2.1

package/metrics/scheduler.py

@@ -1,275 +1,275 @@
-"""The bench's own scheduler: the recurring checks, without Windows Task Scheduler.
-
-    python metrics/scheduler.py              # mode, every job, when each is next due
-    python metrics/scheduler.py --selftest
-
-RELEASE-PLAN.md Phase 3. The checks used to exist only as Windows scheduled
-tasks (scripts/schedule.ps1), which macOS, Linux and Docker do not have. serve.py
-starts this in a background thread; the jobs below are the same seven, at the
-same times, with the same arguments (schedule.ps1 explains each choice).
-
-MODE, from config/scheduler.json:
-  absent / "off"   nothing runs. A fresh install spends no model calls until
-                   someone opts in.
-  "log-only"       at each due time, write "would run X" to runs/scheduler.log
-                   and run nothing. The switch-over week: it runs beside the
-                   Windows tasks, which still do the work, so its timing can be
-                   checked against theirs without doubling a single call.
-  "run"            run the job (the repo's venv Python), one at a time per job,
-                   with the Windows tasks switched off in the same step.
-`"jobs": {"perf": false}` leaves a job out.
-
-A run missed while the dashboard was down is caught up once when it next
-starts (the Windows tasks' "start when available"). The last due time of each
-job is kept in runs/scheduler-state.json; a lock beside it means only one
-scheduler runs however many dashboards are started.
-"""
-
-from __future__ import annotations
-
-import datetime as dt
-import json
-import pathlib
-import subprocess
-import sys
-import threading
-import time
-
-ROOT = pathlib.Path(__file__).resolve().parent.parent
-from paths import DATA  # noqa: E402  -- the install's data folder (metrics/paths.py)
-CONFIG = DATA / "config" / "scheduler.json"
-LOG = DATA / "runs" / "scheduler.log"
-STATE = DATA / "runs" / "scheduler-state.json"
-TICK_S = 30
-RUN_LIMIT_S = 2 * 3600                             # schedule.ps1's ExecutionTimeLimit
-CATCH_UP_DAYS = 7                                  # older misses are not worth a burst of calls
-
-# name -> (script and arguments, slots). A slot is (weekday or None, "HH:MM");
-# weekday 0 is Monday. Mirrors scripts/schedule.ps1, which says why each runs when.
-JOBS = {
-    "availability": (["perf/run.py", "-n", "1", "--tiny-only", "--timeout", "90", "--client", "builtin"],
-                     [(None, "05:00")]),
-    "catalog": (["metrics/catalog.py"], [(None, "06:00")]),
-    "suite-free": (["suite/run.py", "--next-free"], [(None, "06:00")]),
-    "deprecations": (["scripts/check_deprecations.py", "--quiet"], [(None, "08:00")]),
-    "perf": (["perf/run.py", "-n", "3", "--timeout", "90", "--client", "builtin"], [(0, "07:00")]),
-    "capabilities": (["metrics/probe_caps.py", "--gap", "30"], [(6, "07:00")]),
-    "suite": (["suite/run.py", "--next"], [(None, t) for t in ("03:00", "09:00", "15:00", "21:00")]),
-}
-# Each mode as the Settings page names and explains it (the docstring above).
-MODES = {"off": ("Off", "Nothing runs. A fresh install spends no model calls until someone opts in."),
-         "log-only": ("Log only", "At each due time it writes \"would run\" to runs/scheduler.log and runs"
-                                  " nothing: for checking its times against another scheduler first."),
-         "run": ("Run", "Runs each job at its time, one at a time per job, making the calls the job makes.")}
-# What each job is for, in a sentence, for the Settings page.
-ABOUT = {
-    "availability": "Is each model still answering? One one-word prompt per model every morning;"
-                    " feeds the status on At a glance and auto-retire.",
-    "catalog": "Refreshes what is known about models (context, output limit, price, free tier) and the"
-               " Discover list from public feeds. No model calls.",
-    "suite-free": "Scores one free model on the review suite: up to six tasks it has not answered yet,"
-                  " within the free tier's shared daily budget.",
-    "deprecations": "Checks which models each endpoint still serves, and retires models that are gone or"
-                    " have failed for days (auto-retire), after one confirming call.",
-    "perf": "Weekly speed test: a short and a long prompt, three samples per model, for typical and"
-            " slow response times and output speed. The trend is on Benchmarks.",
-    "capabilities": "Weekly check of what each model can do: tool calling and reasoning, two short calls"
-                    " per model, spaced out.",
-    "suite": "Scores one paid model on the review suite every six hours: up to six tasks it has not"
-             " answered yet. These scores are what the router ranks by.",
-}
-
-
-def settings() -> dict:
-    try:
-        data = json.loads(CONFIG.read_text(encoding="utf-8"))
-    except (OSError, ValueError):
-        return {"mode": "off", "jobs": {}}
-    return {"mode": data.get("mode", "off"), "jobs": data.get("jobs") or {}}
-
-
-def enabled(conf: dict) -> list[str]:
-    return [n for n in JOBS if conf["jobs"].get(n, True)]
-
-
-def due_between(slots, start: dt.datetime, end: dt.datetime) -> list[dt.datetime]:
-    """Every slot time in (start, end], oldest first."""
-    out = []
-    day = start.date()
-    while day <= end.date():
-        for weekday, hhmm in slots:
-            if weekday is not None and day.weekday() != weekday:
-                continue
-            h, m = map(int, hhmm.split(":"))
-            when = dt.datetime.combine(day, dt.time(h, m))
-            if start < when <= end:
-                out.append(when)
-        day += dt.timedelta(days=1)
-    return sorted(out)
-
-
-def next_due(slots, after: dt.datetime) -> dt.datetime:
-    return due_between(slots, after, after + dt.timedelta(days=8))[0]
-
-
-def _log(line: str) -> None:
-    LOG.parent.mkdir(parents=True, exist_ok=True)
-    with LOG.open("a", encoding="utf-8") as fh:
-        fh.write(f"{dt.datetime.now():%Y-%m-%d %H:%M:%S}  {line}\n")
-
-
-def _run(name: str, args: list[str], due: dt.datetime, running: dict) -> None:
-    """One job, in its own thread, recorded in the log with its exit code."""
-    sys.path.insert(0, str(ROOT / "suite"))
-    from run import bench_python
-    started = time.monotonic()
-    try:
-        done = subprocess.run([bench_python(), *args], cwd=ROOT, capture_output=True, text=True,
-                              encoding="utf-8", errors="replace", timeout=RUN_LIMIT_S)
-        result = f"exit {done.returncode}"
-    except subprocess.TimeoutExpired:
-        result = f"stopped at the {RUN_LIMIT_S // 3600} h limit"
-    except OSError as exc:
-        result = f"could not start ({exc})"
-    finally:
-        running.pop(name, None)
-    _log(f"ran {name} (due {due:%H:%M}): {result} in {time.monotonic() - started:.0f}s")
-
-
-def tick(now: dt.datetime, state: dict, conf: dict, running: dict, start_job=None) -> list[str]:
-    """Act on everything due since each job's last due time. Changes `state`
-    in place; returns what it logged (for the self-check)."""
-    logged = []
-    for name in enabled(conf):
-        args, slots = JOBS[name]
-        last = dt.datetime.fromisoformat(state[name]) if name in state else now
-        due = due_between(slots, max(last, now - dt.timedelta(days=CATCH_UP_DAYS)), now)
-        if not due:
-            state.setdefault(name, now.isoformat(timespec="seconds"))
-            continue
-        latest = due[-1]
-        missed = " (catching up a run missed while stopped)" if now - latest > dt.timedelta(minutes=5) else ""
-        state[name] = latest.isoformat(timespec="seconds")
-        if conf["mode"] == "log-only":
-            line = f"would run {name} (due {latest:%a %H:%M}){missed}: {' '.join(args)}"
-        elif name in running:
-            line = f"skipped {name} (due {latest:%H:%M}): the previous run is still going"
-        else:
-            line = f"starting {name} (due {latest:%H:%M}){missed}"
-            running[name] = True
-            (start_job or _start)(name, args, latest, running)
-        _log(line)
-        logged.append(line)
-    return logged
-
-
-def _start(name, args, due, running) -> None:
-    threading.Thread(target=_run, args=(name, args, due, running), daemon=True).start()
-
-
-_stop = threading.Event()
-
-
-def _loop() -> None:
-    import locked
-    # Its own lock file, not the state file's: update() below locks that one
-    # for each write, and holding it here as well would wait on itself.
-    try:
-        instance = locked.lock(DATA / "runs" / "scheduler", timeout=0)
-        instance.__enter__()                            # one scheduler per machine
-    except TimeoutError:
-        return                                          # another dashboard already runs it
-    try:
-        running: dict = {}
-        conf = settings()
-        _log(f"scheduler started, mode {conf['mode']}, jobs {', '.join(enabled(conf))}")
-        while True:
-            try:
-                conf = settings()                       # a mode change applies without a restart
-                if conf["mode"] in ("log-only", "run"):
-                    locked.update(STATE, lambda s: tick(dt.datetime.now().replace(microsecond=0),
-                                                        s, conf, running))
-            except Exception as exc:                    # noqa: BLE001 -- one bad tick must not end it
-                _log(f"tick failed: {type(exc).__name__}: {exc}")
-            if _stop.wait(TICK_S):
-                return
-    finally:
-        instance.__exit__(None, None, None)
-
-
-def start() -> threading.Thread:
-    """Called once by serve.py. Does nothing visible when the mode is off."""
-    _stop.clear()
-    thread = threading.Thread(target=_loop, name="scheduler", daemon=True)
-    thread.start()
-    return thread
-
-
-def _selftest() -> None:
-    mon = dt.datetime(2026, 10, 19, 4, 0)               # a Monday
-    slots = JOBS["suite"][1]
-    assert [d.hour for d in due_between(slots, mon, mon + dt.timedelta(days=1))] == [9, 15, 21, 3]
-    assert next_due(JOBS["perf"][1], mon) == dt.datetime(2026, 10, 19, 7, 0)
-    assert next_due(JOBS["perf"][1], mon.replace(hour=8)) == dt.datetime(2026, 10, 26, 7, 0)
-    assert next_due(JOBS["capabilities"][1], mon) == dt.datetime(2026, 10, 25, 7, 0)
-    conf = {"mode": "log-only", "jobs": {n: n == "availability" for n in JOBS}}
-    state: dict = {}
-    global _log
-    real_log, _log = _log, (lambda line: None)
-    try:
-        assert tick(mon, state, conf, {}) == [] and state == {"availability": mon.isoformat()}
-        lines = tick(mon.replace(hour=5, minute=1), state, conf, {})
-        assert len(lines) == 1 and lines[0].startswith("would run availability"), lines
-        assert tick(mon.replace(hour=5, minute=2), state, conf, {}) == []      # once per due time
-        # Stopped over two mornings: one catch-up, not two runs.
-        lines = tick(mon.replace(hour=5) + dt.timedelta(days=2, hours=3), state, conf, {})
-        assert len(lines) == 1 and "catching up" in lines[0], lines
-        started = []
-        run_conf = {"mode": "run", "jobs": conf["jobs"]}
-        lines = tick(mon.replace(hour=5) + dt.timedelta(days=3, minutes=1), state, run_conf,
-                     {}, start_job=lambda *a: started.append(a[0]))
-        assert started == ["availability"] and lines[0].startswith("starting availability"), lines
-        lines = tick(mon.replace(hour=5) + dt.timedelta(days=4, minutes=1), state, run_conf,
-                     {"availability": True}, start_job=lambda *a: started.append(a[0]))
-        assert started == ["availability"] and lines[0].startswith("skipped"), lines   # no overlap
-    finally:
-        _log = real_log
-    # The loop itself, briefly, on temp files: it must write its state while
-    # holding the one-per-machine lock (the first version waited on itself).
-    import tempfile
-    global DATA, CONFIG, LOG, STATE, TICK_S
-    real = DATA, CONFIG, LOG, STATE, TICK_S
-    with tempfile.TemporaryDirectory() as tmp:
-        DATA = pathlib.Path(tmp)                        # the lock and state go here, not to the real data
-        CONFIG, LOG, STATE, TICK_S = DATA / "s.json", DATA / "runs" / "s.log", DATA / "runs" / "s.json", 0.05
-        CONFIG.write_text('{"mode": "log-only"}', encoding="utf-8")
-        thread = start()
-        try:
-            for _ in range(100):
-                if STATE.exists():
-                    break
-                time.sleep(0.05)
-            assert STATE.exists() and set(json.loads(STATE.read_text())) == set(JOBS), "loop wrote no state"
-        finally:
-            _stop.set()                                 # stopped, and its lock let go, before
-            thread.join(5)                              # the real paths come back
-            assert not thread.is_alive(), "the loop did not stop"
-            DATA, CONFIG, LOG, STATE, TICK_S = real
-    print("ok")
-
-
-def main() -> None:
-    sys.stdout.reconfigure(encoding="utf-8", errors="replace")
-    if "--selftest" in sys.argv:
-        _selftest()
-        return
-    conf, now = settings(), dt.datetime.now()
-    print(f"mode: {conf['mode']}  (config/scheduler.json)")
-    for name in JOBS:
-        args, slots = JOBS[name]
-        on = "on " if name in enabled(conf) else "off"
-        print(f"  {on} {name:<13} next {next_due(slots, now):%a %d %b %H:%M}   {' '.join(args)}")
-
-
-if __name__ == "__main__":
-    main()
+"""The bench's own scheduler: the recurring checks, without Windows Task Scheduler.
+
+    python metrics/scheduler.py              # mode, every job, when each is next due
+    python metrics/scheduler.py --selftest
+
+RELEASE-PLAN.md Phase 3. The checks used to exist only as Windows scheduled
+tasks (scripts/schedule.ps1), which macOS, Linux and Docker do not have. serve.py
+starts this in a background thread; the jobs below are the same seven, at the
+same times, with the same arguments (schedule.ps1 explains each choice).
+
+MODE, from config/scheduler.json:
+  absent / "off"   nothing runs. A fresh install spends no model calls until
+                   someone opts in.
+  "log-only"       at each due time, write "would run X" to runs/scheduler.log
+                   and run nothing. The switch-over week: it runs beside the
+                   Windows tasks, which still do the work, so its timing can be
+                   checked against theirs without doubling a single call.
+  "run"            run the job (the repo's venv Python), one at a time per job,
+                   with the Windows tasks switched off in the same step.
+`"jobs": {"perf": false}` leaves a job out.
+
+A run missed while the dashboard was down is caught up once when it next
+starts (the Windows tasks' "start when available"). The last due time of each
+job is kept in runs/scheduler-state.json; a lock beside it means only one
+scheduler runs however many dashboards are started.
+"""
+
+from __future__ import annotations
+
+import datetime as dt
+import json
+import pathlib
+import subprocess
+import sys
+import threading
+import time
+
+ROOT = pathlib.Path(__file__).resolve().parent.parent
+from paths import DATA  # noqa: E402  -- the install's data folder (metrics/paths.py)
+CONFIG = DATA / "config" / "scheduler.json"
+LOG = DATA / "runs" / "scheduler.log"
+STATE = DATA / "runs" / "scheduler-state.json"
+TICK_S = 30
+RUN_LIMIT_S = 2 * 3600                             # schedule.ps1's ExecutionTimeLimit
+CATCH_UP_DAYS = 7                                  # older misses are not worth a burst of calls
+
+# name -> (script and arguments, slots). A slot is (weekday or None, "HH:MM");
+# weekday 0 is Monday. Mirrors scripts/schedule.ps1, which says why each runs when.
+JOBS = {
+    "availability": (["perf/run.py", "-n", "1", "--tiny-only", "--timeout", "90", "--client", "builtin"],
+                     [(None, "05:00")]),
+    "catalog": (["metrics/catalog.py"], [(None, "06:00")]),
+    "suite-free": (["suite/run.py", "--next-free"], [(None, "06:00")]),
+    "deprecations": (["scripts/check_deprecations.py", "--quiet"], [(None, "08:00")]),
+    "perf": (["perf/run.py", "-n", "3", "--timeout", "90", "--client", "builtin"], [(0, "07:00")]),
+    "capabilities": (["metrics/probe_caps.py", "--gap", "30"], [(6, "07:00")]),
+    "suite": (["suite/run.py", "--next"], [(None, t) for t in ("03:00", "09:00", "15:00", "21:00")]),
+}
+# Each mode as the Settings page names and explains it (the docstring above).
+MODES = {"off": ("Off", "Nothing runs. A fresh install spends no model calls until someone opts in."),
+         "log-only": ("Log only", "At each due time it writes \"would run\" to runs/scheduler.log and runs"
+                                  " nothing: for checking its times against another scheduler first."),
+         "run": ("Run", "Runs each job at its time, one at a time per job, making the calls the job makes.")}
+# What each job is for, in a sentence, for the Settings page.
+ABOUT = {
+    "availability": "Is each model still answering? One one-word prompt per model every morning;"
+                    " feeds the status on At a glance and auto-retire.",
+    "catalog": "Refreshes what is known about models (context, output limit, price, free tier) and the"
+               " Discover list from public feeds. No model calls.",
+    "suite-free": "Scores one free model on the review suite: up to six tasks it has not answered yet,"
+                  " within the free tier's shared daily budget.",
+    "deprecations": "Checks which models each endpoint still serves, and retires models that are gone or"
+                    " have failed for days (auto-retire), after one confirming call.",
+    "perf": "Weekly speed test: a short and a long prompt, three samples per model, for typical and"
+            " slow response times and output speed. The trend is on Benchmarks.",
+    "capabilities": "Weekly check of what each model can do: tool calling and reasoning, two short calls"
+                    " per model, spaced out.",
+    "suite": "Scores one paid model on the review suite every six hours: up to six tasks it has not"
+             " answered yet. These scores are what the router ranks by.",
+}
+
+
+def settings() -> dict:
+    try:
+        data = json.loads(CONFIG.read_text(encoding="utf-8"))
+    except (OSError, ValueError):
+        return {"mode": "off", "jobs": {}}
+    return {"mode": data.get("mode", "off"), "jobs": data.get("jobs") or {}}
+
+
+def enabled(conf: dict) -> list[str]:
+    return [n for n in JOBS if conf["jobs"].get(n, True)]
+
+
+def due_between(slots, start: dt.datetime, end: dt.datetime) -> list[dt.datetime]:
+    """Every slot time in (start, end], oldest first."""
+    out = []
+    day = start.date()
+    while day <= end.date():
+        for weekday, hhmm in slots:
+            if weekday is not None and day.weekday() != weekday:
+                continue
+            h, m = map(int, hhmm.split(":"))
+            when = dt.datetime.combine(day, dt.time(h, m))
+            if start < when <= end:
+                out.append(when)
+        day += dt.timedelta(days=1)
+    return sorted(out)
+
+
+def next_due(slots, after: dt.datetime) -> dt.datetime:
+    return due_between(slots, after, after + dt.timedelta(days=8))[0]
+
+
+def _log(line: str) -> None:
+    LOG.parent.mkdir(parents=True, exist_ok=True)
+    with LOG.open("a", encoding="utf-8") as fh:
+        fh.write(f"{dt.datetime.now():%Y-%m-%d %H:%M:%S}  {line}\n")
+
+
+def _run(name: str, args: list[str], due: dt.datetime, running: dict) -> None:
+    """One job, in its own thread, recorded in the log with its exit code."""
+    sys.path.insert(0, str(ROOT / "suite"))
+    from run import bench_python
+    started = time.monotonic()
+    try:
+        done = subprocess.run([bench_python(), *args], cwd=ROOT, capture_output=True, text=True,
+                              encoding="utf-8", errors="replace", timeout=RUN_LIMIT_S)
+        result = f"exit {done.returncode}"
+    except subprocess.TimeoutExpired:
+        result = f"stopped at the {RUN_LIMIT_S // 3600} h limit"
+    except OSError as exc:
+        result = f"could not start ({exc})"
+    finally:
+        running.pop(name, None)
+    _log(f"ran {name} (due {due:%H:%M}): {result} in {time.monotonic() - started:.0f}s")
+
+
+def tick(now: dt.datetime, state: dict, conf: dict, running: dict, start_job=None) -> list[str]:
+    """Act on everything due since each job's last due time. Changes `state`
+    in place; returns what it logged (for the self-check)."""
+    logged = []
+    for name in enabled(conf):
+        args, slots = JOBS[name]
+        last = dt.datetime.fromisoformat(state[name]) if name in state else now
+        due = due_between(slots, max(last, now - dt.timedelta(days=CATCH_UP_DAYS)), now)
+        if not due:
+            state.setdefault(name, now.isoformat(timespec="seconds"))
+            continue
+        latest = due[-1]
+        missed = " (catching up a run missed while stopped)" if now - latest > dt.timedelta(minutes=5) else ""
+        state[name] = latest.isoformat(timespec="seconds")
+        if conf["mode"] == "log-only":
+            line = f"would run {name} (due {latest:%a %H:%M}){missed}: {' '.join(args)}"
+        elif name in running:
+            line = f"skipped {name} (due {latest:%H:%M}): the previous run is still going"
+        else:
+            line = f"starting {name} (due {latest:%H:%M}){missed}"
+            running[name] = True
+            (start_job or _start)(name, args, latest, running)
+        _log(line)
+        logged.append(line)
+    return logged
+
+
+def _start(name, args, due, running) -> None:
+    threading.Thread(target=_run, args=(name, args, due, running), daemon=True).start()
+
+
+_stop = threading.Event()
+
+
+def _loop() -> None:
+    import locked
+    # Its own lock file, not the state file's: update() below locks that one
+    # for each write, and holding it here as well would wait on itself.
+    try:
+        instance = locked.lock(DATA / "runs" / "scheduler", timeout=0)
+        instance.__enter__()                            # one scheduler per machine
+    except TimeoutError:
+        return                                          # another dashboard already runs it
+    try:
+        running: dict = {}
+        conf = settings()
+        _log(f"scheduler started, mode {conf['mode']}, jobs {', '.join(enabled(conf))}")
+        while True:
+            try:
+                conf = settings()                       # a mode change applies without a restart
+                if conf["mode"] in ("log-only", "run"):
+                    locked.update(STATE, lambda s: tick(dt.datetime.now().replace(microsecond=0),
+                                                        s, conf, running))
+            except Exception as exc:                    # noqa: BLE001 -- one bad tick must not end it
+                _log(f"tick failed: {type(exc).__name__}: {exc}")
+            if _stop.wait(TICK_S):
+                return
+    finally:
+        instance.__exit__(None, None, None)
+
+
+def start() -> threading.Thread:
+    """Called once by serve.py. Does nothing visible when the mode is off."""
+    _stop.clear()
+    thread = threading.Thread(target=_loop, name="scheduler", daemon=True)
+    thread.start()
+    return thread
+
+
+def _selftest() -> None:
+    mon = dt.datetime(2026, 10, 19, 4, 0)               # a Monday
+    slots = JOBS["suite"][1]
+    assert [d.hour for d in due_between(slots, mon, mon + dt.timedelta(days=1))] == [9, 15, 21, 3]
+    assert next_due(JOBS["perf"][1], mon) == dt.datetime(2026, 10, 19, 7, 0)
+    assert next_due(JOBS["perf"][1], mon.replace(hour=8)) == dt.datetime(2026, 10, 26, 7, 0)
+    assert next_due(JOBS["capabilities"][1], mon) == dt.datetime(2026, 10, 25, 7, 0)
+    conf = {"mode": "log-only", "jobs": {n: n == "availability" for n in JOBS}}
+    state: dict = {}
+    global _log
+    real_log, _log = _log, (lambda line: None)
+    try:
+        assert tick(mon, state, conf, {}) == [] and state == {"availability": mon.isoformat()}
+        lines = tick(mon.replace(hour=5, minute=1), state, conf, {})
+        assert len(lines) == 1 and lines[0].startswith("would run availability"), lines
+        assert tick(mon.replace(hour=5, minute=2), state, conf, {}) == []      # once per due time
+        # Stopped over two mornings: one catch-up, not two runs.
+        lines = tick(mon.replace(hour=5) + dt.timedelta(days=2, hours=3), state, conf, {})
+        assert len(lines) == 1 and "catching up" in lines[0], lines
+        started = []
+        run_conf = {"mode": "run", "jobs": conf["jobs"]}
+        lines = tick(mon.replace(hour=5) + dt.timedelta(days=3, minutes=1), state, run_conf,
+                     {}, start_job=lambda *a: started.append(a[0]))
+        assert started == ["availability"] and lines[0].startswith("starting availability"), lines
+        lines = tick(mon.replace(hour=5) + dt.timedelta(days=4, minutes=1), state, run_conf,
+                     {"availability": True}, start_job=lambda *a: started.append(a[0]))
+        assert started == ["availability"] and lines[0].startswith("skipped"), lines   # no overlap
+    finally:
+        _log = real_log
+    # The loop itself, briefly, on temp files: it must write its state while
+    # holding the one-per-machine lock (the first version waited on itself).
+    import tempfile
+    global DATA, CONFIG, LOG, STATE, TICK_S
+    real = DATA, CONFIG, LOG, STATE, TICK_S
+    with tempfile.TemporaryDirectory() as tmp:
+        DATA = pathlib.Path(tmp)                        # the lock and state go here, not to the real data
+        CONFIG, LOG, STATE, TICK_S = DATA / "s.json", DATA / "runs" / "s.log", DATA / "runs" / "s.json", 0.05
+        CONFIG.write_text('{"mode": "log-only"}', encoding="utf-8")
+        thread = start()
+        try:
+            for _ in range(100):
+                if STATE.exists():
+                    break
+                time.sleep(0.05)
+            assert STATE.exists() and set(json.loads(STATE.read_text())) == set(JOBS), "loop wrote no state"
+        finally:
+            _stop.set()                                 # stopped, and its lock let go, before
+            thread.join(5)                              # the real paths come back
+            assert not thread.is_alive(), "the loop did not stop"
+            DATA, CONFIG, LOG, STATE, TICK_S = real
+    print("ok")
+
+
+def main() -> None:
+    sys.stdout.reconfigure(encoding="utf-8", errors="replace")
+    if "--selftest" in sys.argv:
+        _selftest()
+        return
+    conf, now = settings(), dt.datetime.now()
+    print(f"mode: {conf['mode']}  (config/scheduler.json)")
+    for name in JOBS:
+        args, slots = JOBS[name]
+        on = "on " if name in enabled(conf) else "off"
+        print(f"  {on} {name:<13} next {next_due(slots, now):%a %d %b %H:%M}   {' '.join(args)}")
+
+
+if __name__ == "__main__":
+    main()