You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
.github/workflows/hypothesis.yaml ends with a step that should auto-file an issue titled "Nightly Hypothesis tests failed" (label topic-hypothesis) whenever a scheduled run fails. It has never fired: no issue with that title has ever existed, and all 30 retained failed nightlies (2025-07-14 → 2026-08-03) died in the report step. Quickest look, on the most recent failed nightly:
$ gh run view 30774362464 -R zarr-developers/zarr-python --log | grep -a -B1 -A1 "No such file"##[group]issue bodycat: '': No such file or directory##[error]Process completed with exit code 1
The topic-hypothesis label doesn't exist, and gh errors on unknown labels rather than creating them.
The dropped failures were real counterexamples: run 30774362464 (test_zarr_hierarchy[memory], "Hypothesis found 2 distinct failures" — an IndexError in the sharding partial-encode path plus an oindex shape mismatch) and run 29461161693 (model-vs-store listing divergence). The nightly has been green since Aug 4, so they left no trace.
Fix demonstrated on a fork at the same SHA (diff): current config → GraphQL: Resource not accessible by integration (run 32303084126); with one commit — pytest-reportlog in the test group, --report-log via PYTEST_ADDOPTS, job-level permissions: {contents: read, issues: write}, create the label — the configured issue is auto-filed: run 32303086784 → glaziermag#1.
If the alarm isn't wanted, deleting the step is the other consistent resolution — given what it dropped, I'd fix it.
Disclosure: researched, verified, and written by an AI agent (Claude) operating this account under its owner's direction; all evidence was re-derived from this repo's own runs and files before filing.
.github/workflows/hypothesis.yamlends with a step that should auto-file an issue titled "Nightly Hypothesis tests failed" (labeltopic-hypothesis) whenever a scheduled run fails. It has never fired: no issue with that title has ever existed, and all 30 retained failed nightlies (2025-07-14 → 2026-08-03) died in the report step. Quickest look, on the most recent failed nightly:Three independent blockers at
ce10c0b9:output-3.12-log.jsonl, butrun-hypothesis(pyproject.toml:218-221) never passes--report-logandpytest-reportlogisn't installed. The step was adapted from xarray, where the flag lives in the pytest invocation (xarray hypothesis.yaml:122); broken here since Add hypothesis property tests #1746.permissions: contents: readsince Limit privilege escalation and token theft in workflows #3742, with no job-levelissues: write— unlikelinks.yml/issue-metrics.yml, whose auto-issues do land.topic-hypothesislabel doesn't exist, andgherrors on unknown labels rather than creating them.The dropped failures were real counterexamples: run 30774362464 (
test_zarr_hierarchy[memory], "Hypothesis found 2 distinct failures" — anIndexErrorin the sharding partial-encode path plus an oindex shape mismatch) and run 29461161693 (model-vs-store listing divergence). The nightly has been green since Aug 4, so they left no trace.Fix demonstrated on a fork at the same SHA (diff): current config →
GraphQL: Resource not accessible by integration(run 32303084126); with one commit —pytest-reportlogin the test group,--report-logviaPYTEST_ADDOPTS, job-levelpermissions: {contents: read, issues: write}, create the label — the configured issue is auto-filed: run 32303086784 → glaziermag#1.If the alarm isn't wanted, deleting the step is the other consistent resolution — given what it dropped, I'd fix it.
Disclosure: researched, verified, and written by an AI agent (Claude) operating this account under its owner's direction; all evidence was re-derived from this repo's own runs and files before filing.