| name | watch-and-monitor-with-poller |
| description | Use when monitoring Crewrift evaluations, rounds, or rankings across turns with a persistent poller. |
Watch and monitor with a poller
The crewrift watch loops (crux-battery polling, "re-sweep every new round", "tell me when daveey hits
#1 in Wood") all share one failure mode and one fix. Read this before you launch any standing watch.
The rule (and the one exception)
- Multi-cycle / long-lived poll → use the
Monitor tool, arming it with the poller command
directly. Monitor is harness-managed: it streams the poller's stdout back as events and keeps the
command alive for the whole session. Each printed line becomes a notification.
- A bare
bash run_in_background with an in-shell for ...; do <work>; sleep 900; done DIES after
the first cycle. The harness keeps a launched command alive, but a long in-shell sleep inside a
detached wrapper script gets reaped before the next iteration. Observed twice: the crux-scout poller
(tmp/crux/scout/poll_scout.py) and a recon-sweep watch_loop.sh both logged cycle 1, hit the
sleep, and vanished from ps with a 0-byte output file — arms completed server-side but no verdicts
were harvested and no error surfaced. (session-derived, unverified)
- Exception — a single "tell me when X finishes" wait that exits in seconds is fine as
bash run_in_background with until grep -q DONE <log>; do sleep 30; done. The reaping only bites
the long-lived multi-cycle loop, not a short wait that terminates on its own. (session-derived,
unverified)
Reference tool
universal/tools/monitor.py is the canonical poller shape: a while True: loop that snapshots the
Observatory league/leaderboard via httpx, prints a line only on a meaningful change, and
time.sleep(900) between cycles. It is the thing you arm Monitor with. Copy its structure for a new
watch; only the snapshot body and the change-detection differ.
Recipe: launch a standing watch
1. Write a poller that is idempotent and emits one line per genuinely-new event
The poller must (a) hold its own "already seen" state and (b) print only on change, so every Monitor
event is real signal and not a per-tick repeat.
For a league/leaderboard watch, the snapshot is an authenticated Observatory call. The auth + base-URL
pattern (from monitor.py):
import time, json, sys, httpx
from softmax.auth import get_api_server, load_current_cogames_token
def snapshot():
server = get_api_server()
token = load_current_cogames_token(api_server=server)
H = {"X-Auth-Token": token}
base = f"{server.rstrip('/')}/observatory"
...
return state
def main():
prev = None
while True:
try:
cur = snapshot()
if cur != prev:
print(f"... {cur} ...", flush=True)
prev = cur
except Exception as e:
sys.stderr.write(f"poll error: {e}\n")
time.sleep(900)
if __name__ == "__main__":
main()
Constants you will need (from monitor.py, current as of 2026-06-12 — re-verify the ids if the league
rolled):
- Wood league id:
league_605ff338-0a2e-4e62-aeda-559df9a9198f
- Wood division id:
div_c2be3343-f046-4c21-8674-267b5797a059
For a crux/experience-request battery the snapshot is instead
GET /v2/experience-requests/{xid} per arm, checking episode statuses; gate the print on the NEW-count
(emit only when an arm reaches a terminal completed/failed status you have not yet reported). Keep a
seen-set on disk (e.g. tmp/crux/recon/seen_rounds.json) so a re-run does not re-announce old events.
(session-derived, unverified)
The one allowed try/except is the poll-error guard around the snapshot: a transient API hiccup
must not kill a multi-hour watch. This is the single deliberate exception to "let it crash" — keep it
narrow (wrap only the network call, log to stderr, continue the loop). Do not wrap anything else.
For the stdout filter when arming Monitor, use a grep that surfaces both verdicts and crashes, e.g.
grep -E "TEST |BATTERY DONE|Traceback|Error|rank=|NEW ". (session-derived, unverified)
2. Validate the poller runs before arming Monitor
The harness _harness / poller code runs with unittest, not pytest (pytest is not a project
dependency). If you wrote or touched a _harness test for the poller, run it with:
uv run python -m unittest discover -s _harness -p "test_*.py"
PYTHONPATH=_harness uv run python -m unittest <bare_module_name>
Do a one-shot smoke of the poller itself (let it print one cycle, then Ctrl-C) to confirm auth resolves
and the snapshot parses — before committing it to a long Monitor run.
3. Arm Monitor with the poller as its command
Use the Monitor tool (search its schema with ToolSearch select:Monitor if it is not yet loaded).
Arm it with the poller invocation as the command and the grep filter above, with an until/while
condition matching your terminal signal (e.g. until the battery prints BATTERY DONE, or run open-ended
for a rank watch). Monitor streams each printed line back to you as an event and survives the session.
Gotchas
- Do NOT poll a long watch with
Bash run_in_background + foreground sleep. It is reaped after
cycle 1, silently. This is the whole reason this skill exists. (session-derived, unverified)
- If the Monitor goes quiet longer than ~20 min, sanity-check the arms directly. For a battery, hit
GET /v2/experience-requests/{xid} and read episode statuses; the arms may have reached terminal
server-side while the poller exited early. The poller writes *_results.json incrementally, so if it
did exit you can re-run it once arms are terminal and harvest verdicts from disk. (session-derived,
unverified)
- Re-verify league/division/policy ids and live roster right before launch. The league rolls policy
versions and may roll league/division ids mid-session; a hardcoded id from a prior run can be stale.
- Dashboard staleness is a separate trap, not a poller problem. If you are also watching the local
_harness/dashboard.py dev server (port 8765) and it shows missing/stale runs: the run index
(league_runs.json) is per-checkout — an instance launched from an old worktree knows nothing
about runs submitted elsewhere. Check which checkout the live instance was launched from, kill stale
instances (including another agent's port-squatter), and restart from the current checkout. The
single-file HTML/JS is rendered once at server boot (render.py), so any frontend change needs a server
restart before the browser reflects it. (session-derived, unverified)
Success check
- The Monitor is running and has emitted at least one event (or you have confirmed via a one-shot smoke
that the poller prints and the auth resolves).
- Each emitted line corresponds to a genuinely-new occurrence, not a per-tick repeat (the change gate
works).
- The watch is still alive after the first cadence interval has elapsed — i.e. it did NOT die after
cycle 1 (the symptom of the run_in_background+sleep failure mode you are avoiding).
- When the terminal condition fires (battery
completed, rank reached), you can cross-check it against a
direct GET /v2/experience-requests/{xid} or GET /v2/division-leaderboards/{DV} and the numbers
agree.