[tirth8205]

What went wrong

CI on Python 3.11 and 3.12 once stopped at about 65% with a KeyboardInterrupt raised from unittest/mock.py.

Several watch tests ended the watch loop with patch("time.sleep", side_effect=KeyboardInterrupt). That patches sleep for the whole process. subprocess.run(timeout=...) also calls time.sleep while it waits for a child to exit. On a busy runner the child has not exited on the first poll, so a call in watch startup hit the patched sleep, the interrupt escaped watch(), and pytest took it as Ctrl+C and stopped the run.

Note: patch("code_review_graph.<module>.time.sleep") would not help. That path names the same shared time module, so it is still the global sleep.

Fix (tests only)

watch() already has a seam: stop_event. The loop waits on it once per tick and nothing else calls it, so no code under test can reach it.

• Startup tests pass an event that is already set, so the loop ends at its first tick.
• test_cli_watch_with_the_pair_keeps_running_on_a_never_embedded_graph drives cli.main(), which passes no event, so it wraps incremental.watch to add one.
• test_watch_robustness.py's tick driver is now a stop event whose wait() is one tick: it runs one scripted step, then raises KeyboardInterrupt like a real Ctrl+C. It had the same leak: a sleep inside a subprocess call could run a scripted step early.

Tests changed: test_postprocessing.py (2), test_incremental.py (2), test_watch_refresh_without_index.py (1), test_embedding_refresh_skip_notice.py (1), test_watch_robustness.py (tick driver plus 1). No production change. No raising patch of the global time.sleep is left in tests/. No CHANGELOG entry, as only tests change.

Proof

A small pytest plugin forces the subprocess wait path: it makes Popen._try_wait(WNOHANG) report "still running" once per process, which is what a loaded runner does by chance.

Before (on staging), with the plugin:

• test_postprocessing.py and test_embedding_refresh_skip_notice.py: !!! KeyboardInterrupt !!! at unittest/mock.py:1228, run stopped after 20 tests. Each module triggers it on its own.
• test_watch_robustness.py: test_normal_file_change_still_updates_the_graph fails 3 runs out of 3 (a subprocess sleep runs a tick early).

After, with the plugin:

• The 5 changed modules: 286 passed (Python 3.13, and Python 3.11).
• test_watch_robustness.py and test_postprocessing.py: 97 passed, 3 runs out of 3.

Gates: uv run pytest tests/ --tb=short -q: 6390 passed, 900 skipped, 3 xfailed, 4 xpassed. ruff check code_review_graph/ clean (changed test files clean too). mypy: no issues.

🤖 Generated with Claude Code

[github-actions]

code-review-graph review

Overall risk: 0.30 (LOW) — 24 changed function(s)/class(es), 0 affected flow(s), 0 test gap(s)

Risk-scored changes

Risk	Level	Symbol	Location	Tested
0.30	low	tests/test_embedding_refresh_skip_notice.py::TestWatchTellsTheUser.watch_one_tick	tests/test_embedding_refresh_skip_notice.py:473	(test)
0.30	low	tests/test_watch_robustness.py::_TickDriver.__init__	tests/test_watch_robustness.py:138	(test)
0.30	low	tests/test_watch_robustness.py::_TickDriver.wait	tests/test_watch_robustness.py:142	(test)
0.15	low	tests/test_watch_robustness.py::_TickDriver	tests/test_watch_robustness.py:126	(test)
0.15	low	tests/test_watch_robustness.py::TestWatchLoop._watch_with	tests/test_watch_robustness.py:955	(test)
0.05	low	tests/test_embedding_refresh_skip_notice.py::TestWatchTellsTheUser	tests/test_embedding_refresh_skip_notice.py:418	(test)
0.05	low	tests/test_embedding_refresh_skip_notice.py::TestWatchTellsTheUser.test_cli_watch_with_the_pair_keeps_runn...	tests/test_embedding_refresh_skip_notice.py:451	(test)
0.05	low	tests/test_incremental.py::TestWatchReconciliation	tests/test_incremental.py:1795	(test)
0.05	low	tests/test_incremental.py::TestWatchReconciliation.test_watch_reconciles_before_observer_startup	tests/test_incremental.py:1842	(test)
0.05	low	tests/test_incremental.py::TestWatchReconciliation.test_watch_startup_postprocess_warning_stops_observer	tests/test_incremental.py:1865	(test)

Token savings: this graph-backed report used ~56,437 fewer tokens (~94%) than reading every changed file in full (estimated, chars/4 approximation).

---

Powered by code-review-graph — local-first analysis; no code leaves the CI runner.