GeezerVison

Problem

When a test fails in the macOS or Windows test legs (merge queue, main push, nightly) or in test-release, the only annotation is "Process completed with exit code 1". The panic sits thousands of lines above the end of the log, past the 5,000-line tail the log API returns, so the failure can't be diagnosed after the fact.

It happened in tonight's nightly, run 41009903891 at 5ca5623:

• test (macos-latest, 2) failed on -p socket-patch-cli --test e2e_golang_workspace_build, which took 157 s against 41 s or less for every other Go target. The panic is in the first ~1,500 lines of a 6,566-line log, and nothing that reads only the tail (log API, MCP log tool, annotations) can reach it. The suite passes on Linux (31/31 locally in 175 s). It is the first nightly to include the Go workspace changes from Fix Go replace written to unused go.work (#1810) #1812, Fix Go patches ignoring GOFLAGS=-modfile (#1799) #1800 and Fix Go replace-target patches reported applied (#1823) #1824, so this may be a macOS-only regression, but it can't be confirmed without the panic.
• test (windows-latest, 1) failed on crawler_maven_e2e::get_maven_repo_paths_home_dot_m2_fallback. Fix Maven home tests failing on Windows nightly #1852 already fixes it.

The nightly has been red 9 days running (since 2026-10-22). Each time, the failing test has had to be found by hand.

Fix

Tee each leg's output to $RUNNER_TEMP/tests.log. On failure, run scripts/ci-test-failures.py on it, then exit with the original status. coverage already does exactly this. The script prints one ::error annotation per failing test, titled with the target and test and carrying the panic, and writes the same to the job summary. On a passing leg the only change is the tee.

Proof

• Ran the new step body in bash on the saved Windows log from the nightly, with a fake shard that exits 1. It printed ::error title=target\debug\deps\crawler_maven_e2e get_maven_repo_paths_home_dot_m2_fallback::thread ... panicked at crates\socket-patch-core\tests\crawler_maven_e2e.rs:297:5 ... and exited 1.
• python3 -m unittest discover -s scripts/tests -p 'test_ci_*.py': 114 tests OK. These include test_ci_test_shard.py, which parses the test-release block of ci.yml.
• On ci.yml, actionlint and zizmor --offline report the same findings before and after the change (179 and 54, all pre-existing), with none in the edited lines. YAML parses.

Where tests run

No test is added, removed or moved. Job and check names are unchanged, and so is ci-ok.

Note

Workflow changes only take effect on a new head commit. The macOS legs don't run on PRs, and on this PR's head test and test-release were skipped too (ci-ok green: 32 checks passed, 19 skipped), so the merge queue is the first place the changed steps run. They change only how output is captured; the exit status is the shard's own.

https://claude.ai/code/session_011f9aeNzimDA4mfWzq1NoZa

Generated by Claude Code