ci: Coverage job fails on a random package's test:coverage task and hides the failing test #1909

TanStack AI version

main (8731d75)

Framework/Library version

CI only (.github/workflows/pr.yml, Coverage job)

Describe the bug and the steps to reproduce it

The PR Coverage job fails at random on PRs that change @tanstack/ai. No coverage metric drops. One package's test:coverage task fails, and it is a different package each time. A re-run passes.

Seen so far:

The same tests pass locally:

• packages/ai-harness: vitest run, 2469 passed.
• packages/ai-react: vitest run and nx run @tanstack/ai-react:test:coverage --skip-nx-cache, 251 passed.

Likely cause: a change to @tanstack/ai makes the job measure about 76 packages, twice (PR head and merge-base). nx run-many runs them in parallel on one runner. Tests that depend on timing then time out under the load.

Second problem: the job log does not show the output of the failed task. pr.yml:76 and pr.yml:105 run nx run-many --outputStyle=static, and the log only has the Nx summary line ✖ @tanstack/ai-harness:test:coverage. So nobody can name the flaky test from CI.

Fix

1. Make the Coverage job print the full output of a failed task, so the failing test name and error are in the job log.
2. Find the timing-sensitive tests in ai-harness and ai-react. To do this, run their coverage under CPU load locally, or use the log from step 1. Then fix them: remove real-time waits, use fake timers, or wait for events instead of a fixed time.
3. If step 2 does not remove all failures, cap the parallel tasks for the coverage step (--parallel=N). Do not add retries that hide failures.

Your Minimal, Reproducible Example - (Sandbox Highly Recommended)

CI runs listed above. No local reproduction yet.

Do you intend to try to help solve this bug with your own PR?

Yes, I am also opening a PR to fix this issue.