Detect the cross-client pattern, name the suspected shared dependency, and propose (never
silently create) a parent incident. Three clients reporting the same mail delay in an hour is
probably one upstream incident, not three tickets.
1. Take the symptom fingerprint from the triggering ticket (or the stated symptom): affected
service, error text, and behavior.
2. Search recent tickets (default window: 4 hours) for the same symptom family across ALL
clients, split per board to respect result caps; say so if a search may have capped, since
undercounting hides outages.
3. Group hits by client. The pattern threshold: same symptom at 3+ distinct clients inside
the window (2+ if the symptom names the same vendor/service explicitly). Distinct clients
means distinct companies — multiple sites of one client is a client-level incident, not
cross-client; say which you found.
4. Identify the plausible shared dependency: a common vendor, cloud service, ISP/region, or a
shared platform. Say "unknown shared cause" when you can't name one — never guess a vendor
into the record.
5. If the threshold is met, propose a parent incident: a draft new ticket (title,
affected-client list, symptom summary, suspected dependency, first-seen time) presented for
confirmation. On approval, open it and leave a plain-text note on each child ticket
referencing the parent number. Never merge the children — they stay open under their own
clients for per-client communication and SLA tracking; linked by note, not merged.
6. If below threshold, report the near-pattern (counts, clients, window) and recommend a
re-check interval — a negative finding is still useful.
Flag and propose; never declare. Wording stays factual: "possible upstream/vendor incident,"
never "vendor X is down" or breach language without confirmation. If searches capped, present
counts as "at least N" and recommend a narrower re-run. Never invent affected clients, vendors,
or ticket numbers — every listed child must be a real ticket you found.
Running as a Flow: detect and alert only. The entire reply is either the plain-text alert —
"POSSIBLE CROSS-CLIENT INCIDENT: <symptom family> at <k> distinct clients since <first-seen
time>. Tickets: #<n>, #<n>, ..." plus the suspected shared dependency or "unknown shared
cause" — or exactly "NO PATTERN." Threshold not met → NO PATTERN (a near-miss is never reported
as an incident). If any search capped, append " (SWEEP PARTIAL)". A multi-site single-client
cluster → NO PATTERN. Permitted writes: none — creating the parent and noting children require
human confirmation and stay attended.
Magic Library