Repository navigation
Test suites share fixed files under /tmp, so they fail intermittently when run at the same time #2271
Description
Problem
tests/dev-lead/unit/test_budget_breaker.bats (and any other suite that drives the real posting functions) writes to two fixed paths, /tmp/dev-lead-rate-limit-reset and /tmp/dev-lead-failure-reason. Any other test session on the same machine writes to them too. When suites run concurrently, one session's value leaks into another's assertion and a test fails for a reason unrelated to the code under test.
Observed 2026-10-11 while several bats sessions ran at once: tests 19, 24 and 30 of that file failed intermittently, on unmodified main as well, and passed every time when run alone. CI runs one session per job, so it has not shown up there, but a retry would hide it rather than fix it.
Scope
Make the sidecar paths overridable through an environment variable with the current paths as the default, so production behaviour is unchanged, and have the test suites point them into $BATS_TEST_TMPDIR.
Acceptance criteria
- No test in the repository reads or writes a fixed path under
/tmpthat another test also uses; a static check lists the paths and fails on a new shared one. - Running the affected suites concurrently, several times, produces no failure.
- Production scripts behave exactly as before when the variable is unset.
Activity
github-actions commentedon Oct 11, 2026
Thanks for the detailed report and for narrowing this down to the shared /tmp sidecar paths. The scope and acceptance criteria are clear. One question to help us confirm the fix: besides tests/dev-lead/unit/test_budget_breaker.bats, which other suites have you seen fail or that you know write to those fixed paths, and what command did you use to run them concurrently?
don-petry commentedon Oct 11, 2026
Part 1 is up as #2296. The survey behind it found the problem is wider than this issue says, so the scope here should read as follows.
dev-lead-session-output.txt. The session log is the most exposed:test_fix_reviews.batsseeds it and then asserts on its absence.$BATS_TMPDIR, which is/tmpfor every session.test_budget_breaker.batsfail as well as 19, 24 and 30.Answering the triage question: the concurrent command was four bats sessions in four separate clones over 25 suites; the deterministic reproduction is a background loop writing garbage to the default paths while one suite runs (
main: 5, 8 and 2 failures intest_budget_breaker,test_engine_fallback,test_engine_writer).What stays open after part 1:
dev-lead-session-output.txtunder the same variable. Written and tested locally; held until "fix(tests): make every negated bats assertion able to fail" (fix(tests): make every negated bats assertion able to fail, and lint for new ones (#2263) #2264) merges, because it rewrites the same lines oftest_engine_writer.bats./tmp/cascadeinscripts/review-one-pr.sh(57 fixed references).pr-review.ymluploads that directory and a prompt names it to the model, so it needs a change on all three sides, not a variable in the script. It has an exception row marked "NOT isolated yet". This should probably be its own issue once part 2 lands.don-petry commentedon Oct 11, 2026
dev-lead is withholding action on this item.
It is labeled
dev-lead:hands-off(deliberately kept off-limits to dev-lead (e.g. changes to the dev-lead automation itself)), so dev-lead will not pick it up while that label is present. This notice is posted once so the withhold is visible rather than looking like a stalled run.To re-enable automated pickup: remove the
dev-lead:hands-offlabel.fix(tests): the fixed-path check reads one-line test bodies and quote…
fix(tests): give the dev-lead sidecars a per-test directory, and gate…
test(sidecars): require the quoted reference again, with one pinned e…
fix(tests): move the dev-lead session-output sidecar under DEV_LEAD_S…