GitHub answered the confirming run query with a weeks-old cached result
even with page=1, so alerts for fixed workflows stayed. A created filter
changes the query each day and reads current runs.
A new pr_open alert carries the head commit SHA of each open, non-draft, same-repository pull request. An action with auto enabled sends its webhook once for each pull request and head commit. An authors list limits which pull requests trigger it. The alert never changes repository health.
GitHub can answer one exact run-listing query from an old cached result
for weeks. The same query on a new connection returned the same old
failure, so a workflow whose latest runs passed still showed as failing.
The confirmation request now adds page=1. The changed query skips the
cached result and returns current runs. The test models a cache that
serves the plain query only.
GitHub sometimes answers a workflow run listing from a lagging replica. For one workflow the answer held 135 runs and ended three weeks ago, while the current listing held 414 runs. The client reuses one connection, so a stale replica could answer every request of a scan and the dashboard showed a failure that a newer run had already replaced.
A failing workflow is now requested again on a new connection that skips the response cache. The client keeps the union of both answers, so the newest run decides.
ghtracker authenticates as a GitHub App and scans a configured list of repositories on a schedule. It serves a dashboard, a JSON state endpoint, a health probe and Prometheus metrics.
The dashboard lists only items that need attention: failing workflows, pull requests that wait too long or fail their checks, stale issues and projects that failed to scan. It has search, dynamic filters and a button that rescans on demand. The Vue app is embedded in the Go binary.
The client protects the GitHub rate limit with conditional requests, a request reserve and Retry-After handling.
lukaszraczylo
