| name | run-tests |
| description | Run this repo's pytest suite safely. Use for ANY pytest invocation here — an uncapped run once grew to 30GB and OOM-killed the machine. Covers the memory cap, orphan cleanup, and the mock-stream failure mode behind it. |
Running tests in chatgpt-bots-dev
Every pytest invocation gets a virtual-memory cap. No exceptions, including single-test runs.
(ulimit -v 4194304; timeout 600 python3 -m pytest tests/unit -q -p no:cacheprovider --tb=no -p no:logging)
make test (coverage) and make test-fast are fine to reach for, but wrap them the same way.
Why the cap
An uncapped python3 -m pytest tests/unit reached 30 GB RSS and the kernel OOM-killed it,
taking WSL down with it — repeatedly. The cap costs nothing and converts a runaway into a
MemoryError traceback that names the culprit test instead of a dead machine.
--tb=no is not sufficient on its own: assertion rewriting builds the failure explanation
inside the test, before any reporting happens. The ulimit is the real guard.
Orphan discipline
Interrupting the agent turn does not kill the process it spawned. A stopped turn left an
orphaned pytest eating the box for another 400 seconds.
pgrep -af pytest before and after any run.- Kill orphans explicitly, and run
pkill alone — inside a compound command it kills the
calling shell (exit 144).
- Never run the suite in the background, and never leave one running.
The failure mode it guards against
A mock stream that never terminates. output_text += <MagicMock> does not raise — it silently
turns output_text into a mock, and a stale side_effect once left an async iterator with no end.
Any test mocking a response stream must terminate and must yield real strings. More in
Docs/TOOL_SUBSYSTEMS.md.
