2026-10-22: The Null That Paged

A blank reminder stopped Ki-Adi-Mundi for seven hours.

Affinity returned "content": null on a due reminder for Pong (Vira Labs). dict.get("content", "") keeps a present null, and mundi-watch then sliced it. The run died before it saved state, so the same batch came back every half hour from 08:14. The failure clock paged once it passed an hour. sanctum-crm 7568f35 stores a blank body as an empty string. The 15:46 run exited clean, cleared the clock, and sent the one Signal page that reminder was owed. The other fresh mail was held.

The beat that went quiet on purpose

mbp:qwen-haus-dragon is the haus-dragon supervisor’s hourly proof that a cycle finished. Its limit is three hours. On 20 October at 17:32 EDT the vault SSH lane to manoir had been down for half an hour, so the supervisor withheld the beat. That was amendment A4: if a page cannot reach Bert, stop the heartbeat and let the Cloudflare eye email. The laptop then slept. At 12:07 UTC on 22 October the eye reported the beat 39 hours old.

The lane came back at 11:37 EDT and the beats resumed. The withhold was the wrong signal. A down lane means the pager is stuck, not that the supervisor has stopped. a4d3c16 on feat/haus-dragon keeps beating through a lane outage and withholds only when a dropped page has stayed undeliverable for 30 minutes. While the chain is alive, caffeinate -is -w runs in its own session so a dark display does not idle-sleep the Mac. Closing the lid on battery still sleeps it.

One slow call failed a whole job

The dataroom Function App paged four times in five days. Each page was one transient, and each one failed the entire run.

When (UTC)	What failed	What the next run did
18 Oct 04:18	Graph search for Anomaly Games in Guillaume’s mailbox	Later runs succeeded
19 Oct 04:55	No Friday-brief heartbeat for 26 hours	A run landed by 06:57
19 Oct 05:02	Credential watch, Graph 403 on app registrations	Application.Read.All is on the job identity
21 Oct 14:18	Drive listing returned HTTP 500	The next listing succeeded
22 Oct 04:17	Graph timed out searching joe@ for Tokidos, at the 300 second mark	Every later filing that day was ok

The credential-watch grant was already in place by the time of the later runs. The other three were retries the code did not take. Graph GETs now try a timeout, a dropped connection, and HTTP 429/500/502/503/504 three times. Drive GETs retry HTTP 500/502/503/504 the same way they already retried quota. The Friday-brief timer also fires at 08:07 UTC, an idle heartbeat, so a missed 20:07–22:07 cluster cannot empty the 26-hour dead-man.

That package is 8afe56a on the dataroom Function App, merged to main as ef8ae0a. The 19:45 UTC dataroom tick after the deploy was ok.

A one-line Prettier miss on the tearsheet repo’s CLAUDE.md was the other red check that afternoon. 1fffeec is on main, and the format check is green. It deploys nothing. The fleet work from the same day is The Sentinel That Patches the Fleet.

The eye, the filing job, and the watch now fail when the work has stopped, not when a field is empty or a single call is slow.