[egagli]

Why

The scheduled live tier (the weekly CI run's test-live job) was red four weeks running: Sep 29, Oct 6, 13 and 20, 2026. On Oct 20 it caught the Yukon AquaCache API breaking: 29 failed test_yukon_client.py tests, with /timeseries returning HTTP 500 mid-deploy. But a NASA CMR outage had already made the run red (HLS, MODIS, VIIRS, UCLA SR and Sentinel-1 failures), so nobody looked. The archive went 12 days without Yukon data before anyone noticed.

What

1. scripts/report_live.py and a step in the live job.

• pytest now writes --junitxml=live-results.xml.
• A new if: always() step groups the results by data source (the test file: test_yukon_client.py → yukon).
• It files one issue per failing source under the existing data-source label, using the same pattern as report_health.py:
  • The issue body lists the errors with identical messages counted together, the failing tests, the run link, and a command to reproduce locally.
  • A source still failing a week later has its issue updated and commented, so it surfaces again.
  • A source whose tests pass again has its issue closed.
  • A source whose tests all skipped (no credentials) is left alone.
• Each run also writes a per-source table to the job summary.
• Only scheduled runs file issues; manual runs only write the summary. The job gets issues: write.
• It reuses report_health.GitHub under its own esd-live marker, so it never touches the health check's issues.

2. Watch Yukon's source repos (WATCHLIST.toml). Path-filtered commit feeds for YGwater's API (inst/plumber on dev.WRB) and AquaCache's schema patches (inst/patches on dev.g). The publicly_visible → default_show rename ("patch_65 live", 2026-10-15) was public there a week before the client broke. Those branches matched the live API on 2026-10-26. The entry's comment says what to do when a branch is renamed.

3. Fix a dead link from #79. The catalog's OGL-Yukon reference (open.yukon.ca/node/15612) returns 404. It now uses the licence URL that the Open Yukon dataset pages link to.

Testing

• tests/test_live_reporting.py (new, offline): grouping by file, failure versus error, skip-only sources, identical-error counting, open/update+comment/close against a fake GitHub, and no cross-talk with esd-health issues. Passes along with test_health_reporting.py (40 tests).
• Real input: ran the Yukon and CDEC live tests locally with --junitxml (107 passed) and fed the report to the script in dry-run mode. The output was a correct per-source table and two "close" actions.
• test_watch.py and test_catalog.py pass. One test, test_catalog.py::...cmr_probe..., fails locally on the untouched tree too and passes in CI.
• The workflow YAML and WATCHLIST.toml both parse; ruff check / ruff format --check are clean.

Not verified

• The two GitHub commit feeds: this session's network proxy returns 403 for every github.com .atom URL, including the geoBoundaries feed that already works in CI. If GitHub doesn't serve these path-filtered feeds, the first weekly watch run will report them as unreachable rather than failing.
• The issue-filing step itself can't run on a PR (the live job is schedule/manual only). The first scheduled run, Tuesday 08:00 UTC, will be its first real test. A manual run with run_live writes the summary without filing issues.

---

Generated by Claude Code