Opsgenie's schedule export drops every rotation without the expand=rotation flag

The flag is the part of this that fails quietly. A GET on `/v2/schedules` does not paginate, so one call returns every schedule in the account ^[14]. That same call returns them with the rotations stripped unless you add `expand=rotation` ^[15]. The file parses, the schedule names are all present, and the layers are simply absent.

Rotation layers are where the export turns into real work, not just a backup. A schedule can carry several layers, and the API hands them back in list order without computing what happens where they overlap, because the last rotation in the list wins ^[16]. The payload leaves the effective on-call order unstated, and a team is never resolved into people ^[17]. Who gets paged at 03:00 on a Sunday is therefore a convention you write down per schedule, before the account goes away.

Two more fields carry logic that reads like data. `length` is a hand-off multiplier, so a weekly rotation with length 2 hands off every two weeks ^[18]. Participants can be users, teams, escalations, or none ^[19], which means a layer can be present and enabled while paging nobody.

The 403 and the 401 in this export have unrelated causes. Keys carry four rights, and the schedules, escalations, users and integrations endpoints all require Configuration Access; a Read-only key gets HTTP 403 from every one of them ^[10]. A key aimed at the wrong regional host gets 401 instead, since US accounts use `api.opsgenie.com` and EU accounts `api.eu.opsgenie.com` ^[11].

Scope matters more than rights on the integrations pull. A team API key only returns that team's objects, and a key lifted from inside an "API" integration is often restricted to alert data, so the guide by Sandeep Sidhu calls for the global key and notes that integrations created by people who have since left can sit in a team nobody looks at ^[20]. The user export has the same shape of trap: `blocked=true` means the account cannot log in, not that the name has been removed from a live schedule ^[13].

Atlassian's end of sales on 4 June 2025 and the deletion date of 5 April 2027 sit 670 days apart ^[23]. Against that, the export is about ten minutes of GETs ^[9], and four calls for an organisation small enough to fit inside a single 100-record page of users ^[24].

The checklist lists the six casualties but skips a failure mode for each. Rotation layers get a mechanism, and escalation delays are at least legible in the raw rules structure ^[26]. Time restrictions, overrides, heartbeats and email senders are named as things that do not survive intact, including under the automated Jira Service Management path ^[3][7], without the detail that would let you predict what breaks. Read the list as a map of where to look by hand, not as a description of what you will find.

That is also the argument for treating the JSON as an artefact rather than a download. Those files are the only record of the current configuration ^[21], which is a reason to store them beside the runbooks and not in the personal downloads folder of whoever happened to run curl.