Correlation measurement can't reach a multi-day live window: debug/correlation caps at 100 incidents, GET /incidents times out #197

New Issue

2026-10-04T01:15:39-04:00

logan commented

2026-10-04 01:15:39 -04:00

Found 2026-10-04 while pulling the 09-27→09-29 live-AI window for #170.

GET /admin/debug/correlation?limit=100&orphan_hours=168: limit is capped at 100 incidents sorted by updated_at. At live rates (10–19 incidents/hr) that only reaches back about 5–10h. On 10-04 it only got to 2026-09-30 04:11Z. Orphans hit orphan_scan_cap 3000 inside 2.5h of untranscribed calls.
GET /incidents has no time filter or limit (collection_list over the whole org). From the browser it never returned (Failed to fetch after about 2 min).
Workaround that did work: /calls/search in 1-hour date_from/date_to windows (about 3s each), then /incidents/{id} per id. A single 3.5-day /calls/search request also failed.

Also: Chrome Blob download into ~/Downloads still doesn't work, so only a slim dump (incident rows + aggregates) made it to disk (corr_dump_2026-10-04.json).

Ask: give debug/correlation a date_from/date_to (it rides the same started_at index as /calls/search) and add a bounded/time-ranged incidents list. Related: #57 (standing eval procedure), #170.

Found 2026-10-04 while pulling the 09-27→09-29 live-AI window for #170. 1. `GET /admin/debug/correlation?limit=100&orphan_hours=168`: `limit` is capped at 100 incidents sorted by updated_at. At live rates (10–19 incidents/hr) that only reaches back about 5–10h. On 10-04 it only got to 2026-09-30 04:11Z. Orphans hit `orphan_scan_cap` 3000 inside 2.5h of untranscribed calls. 2. `GET /incidents` has no time filter or limit (`collection_list` over the whole org). From the browser it never returned (`Failed to fetch` after about 2 min). 3. Workaround that did work: `/calls/search` in 1-hour `date_from/date_to` windows (about 3s each), then `/incidents/{id}` per id. A single 3.5-day `/calls/search` request also failed. Also: Chrome Blob download into ~/Downloads still doesn't work, so only a slim dump (incident rows + aggregates) made it to disk (`corr_dump_2026-10-04.json`). Ask: give debug/correlation a `date_from/date_to` (it rides the same started_at index as /calls/search) and add a bounded/time-ranged incidents list. Related: #57 (standing eval procedure), #170.

logan referenced this issue

2026-10-04 01:15:58 -04:00

Admin Replay: re-run the pipeline over past calls in a sandbox, so correlation changes are measured without live AI windows #170

logan referenced this issue

2026-10-04 18:01:07 -04:00

Long-lived Major incident becomes a sink: 'Rollover Entrapment at TSP' absorbed 40 calls / 2h34m / 37 km path on Ch1 #189

logan added the area/correlation