When the graph-injector queue backed up two weeks ago, the post-mortem Jean had filed in April was already sitting in Notion, and I'd skimmed it the morning the alert paged. It still took me ten minutes to find: four candidate pages with "graph" or "connector" in the title, a wrong turn through the architecture diagram, the breadcrumb that pointed back to a doc I'd already ruled out. The page that turned out to be right didn't have "graph-injector" anywhere in its title.
Yesterday, when the Notion integration went GA, I asked Annie the same question I'd just spent ten minutes answering by hand.
Eighteen seconds.
She returned Jean's April post-mortem with the two-line root cause lifted verbatim: NUM_WORKERS=5 on xlarge under-provisioned the remediation deploy, restored 2026-04-30 09:24 UTC to NUM_WORKERS=20 on 3xlarge. The action-items section came back tagged: two checked off in Notion, two still open, page anchor for each. The open one we should actually do this week is "alert on NumberOfMessagesDeleted dropping to zero for >1h". It's a Linear ticket now.
Notion is the second-worst place to put runbooks
Confluence is worse. Pages drift and breadcrumbs lie, and search keeps returning the linked-from page instead of the page itself. The post-mortem I needed didn't have "graph-injector" anywhere in its title because I'd been searching the producer side when the doc was filed on the consumer. Annie picked it anyway, off the body text. I can ask her the question the way I'd ask a teammate, with a half-remembered detail and the wrong service name, and get the right page back.
The next time Notion ships an integration feature, Annie has it without us touching agent-gateway. We built the credential bridge once.
