Build a reliability issue tracker for the workflows in my tenant — a place where problems with workflows get filed, triaged, worked, and closed. It should accept issues from three directions: a person reporting one, another workflow reporting one about itself, and an automated scan that catches failing runs on its own. Use Tines Cases as the system of record (or another issue tracker if you prefer), so nothing about an issue is stored inside this workflow itself. Every issue is a case filed into one team, tagged so the tracker can tell its own cases apart from everything else in that team. Keep structured context — which workflow, which step, which run, who reported it, a dedupe key, occurrence count, and a deep link back to the workflow in 3B — in case metadata, and lean on the tracker's native status, priority, assignee, and comment fields for everything else. Triggers and pieces: A web board at its own route (space members only) as the entry point: columns per status, cards per issue, filters that live in the query string so a filtered view is shareable, a detail drawer with comments, and a form for filing a new issue. Render it as a static bundle that fetches its data client-side so it never blocks on the tracker API. A small JSON API behind the board for listing issues, fetching one issue with its comments, creating an issue, updating one (status, priority, assignment, claiming, commenting), and a trend endpoint that aggregates issues per day and per workflow for charting. An ingest endpoint other workflows can POST to, authenticated as a workflow-backed connector, deduping on a caller-supplied key so repeat reports collapse into one issue. A scheduled scanner (roughly every 15 minutes) that walks the 3B API for errored runs on published branches only, keeps a watermark and the ids it has already seen on a persistent volume so nothing is double-reported, and excludes this workflow from its own sweep. A step that turns detected failures into cases, grouping by workflow and step so a burst of retries becomes one issue with an occurrence count and a comment, rather than a pile of duplicates. Prioritize by how often the step failed in the batch. An AI triage step that reads the failing step's logs and metadata and posts a root-cause hypothesis, the suspect line, and a suggested fix as a comment on newly opened cases. Use Claude via Anthropic — or whichever model provider you prefer. A scheduled auto-close step that checks whether auto-detected issues have recovered (a few consecutive clean runs since the last failure) and closes them with a comment explaining why. A weekday-morning digest posted to Slack — or your team's chat tool — summarizing open issues by priority and calling out ones that are unassigned or have gone quiet, with links back to the board. Keep all human-facing routes private to the space, and make the board the one link on the Links page. Tools used: Tines Cases, 3B, Anthropic Claude, Slack, React, TypeScript