From abe6d2c04511ef9731cfe576675ed4cfdbb7a984 Mon Sep 17 00:00:00 2001
From: Tim Ledbetter <tim.ledbetter@ladybird.org>
Date: Sun, 7 Jun 2026 08:52:33 +0100
Subject: [PATCH] CI: Add a PR CI step to check new and modified tests for
 flakiness

---
 .github/workflows/lagom-template.yml | 14 ++++++++++++++
 Meta/check-test-flakiness.py         | 25 +++++++++++++++++++++++--
 2 files changed, 37 insertions(+), 2 deletions(-)

diff --git a/.github/workflows/lagom-template.yml b/.github/workflows/lagom-template.yml
index c88fe6cdce..3f04a76b5e 100644
--- a/.github/workflows/lagom-template.yml
+++ b/.github/workflows/lagom-template.yml
@@ -201,6 +201,20 @@ jobs:
             --load="module-native-protocol-unix auth-anonymous=1 socket=/tmp/pulse-native" \
             --load="module-null-sink sink_name=virtual_sink"
 
+      - name: Check new tests for flakiness
+        if: ${{ github.event_name == 'pull_request' && inputs.build_preset == 'Sanitizer' && (inputs.os_name == 'Linux' || inputs.os_name == 'macOS') }}
+        timeout-minutes: 30
+        working-directory: ${{ github.workspace }}
+        shell: bash
+        run: |
+          git fetch --no-tags --depth=1 origin ${{ github.event.pull_request.base.sha }}
+          "${{ env.pythonLocation }}/bin/python" Meta/check-test-flakiness.py \
+            --test-web-binary Build/bin/test-web \
+            --base-ref ${{ github.event.pull_request.base.sha }} \
+            --deadline-seconds 1200 \
+            --python-executable "${{ env.pythonLocation }}/bin/python" \
+            | tee -a "$GITHUB_STEP_SUMMARY"
+
       - name: Test
         if: ${{ inputs.build_preset == 'Sanitizer' }}
         working-directory: ${{ github.workspace }}
diff --git a/Meta/check-test-flakiness.py b/Meta/check-test-flakiness.py
index 6de777e78a..260faab818 100755
--- a/Meta/check-test-flakiness.py
+++ b/Meta/check-test-flakiness.py
@@ -82,6 +82,23 @@ def parse_dry_run_output(output):
     return tests
 
 
+def changed_files_in_test_root(repo_root, base_ref):
+    output = run_git(repo_root, "diff", "--name-only", base_ref, "--", TEST_ROOT_RELATIVE.as_posix())
+    prefix = TEST_ROOT_RELATIVE.as_posix() + "/"
+    return [line[len(prefix) :] for line in output.splitlines() if line.startswith(prefix)]
+
+
+def find_modified_tests(repo_root, base_ref, pr_tests):
+    tests_by_input_path = {}
+    for test in pr_tests:
+        tests_by_input_path.setdefault(test.partition("?")[0], set()).add(test)
+
+    modified_tests = set()
+    for relative_path in changed_files_in_test_root(repo_root, base_ref):
+        modified_tests.update(tests_by_input_path.get(relative_path, ()))
+    return modified_tests
+
+
 def dry_run_tests(test_web_binary, test_root, python_executable, results_dir):
     command = [
         str(test_web_binary),
@@ -115,7 +132,11 @@ def discover_candidates(args, repo_root, test_web_binary, base_worktree, scratch
     base_tests = dry_run_tests(test_web_binary, base_test_root, args.python_executable, scratch_dir / "dry-run-base")
     log(f"  {len(base_tests)} tests discovered.")
 
-    return sorted(pr_tests - base_tests)
+    new_tests = pr_tests - base_tests
+    modified_tests = find_modified_tests(repo_root, args.base_ref, pr_tests) - new_tests
+    log(f"  {len(new_tests)} new and {len(modified_tests)} modified test(s) found.")
+
+    return sorted(new_tests) + sorted(modified_tests)
 
 
 def parse_concurrency_levels(spec):
@@ -288,7 +309,7 @@ def main(argv):
         cleanup_worktree(repo_root, base_worktree_path, scratch_dir)
 
     if not candidates:
-        log("No new tests: nothing to check.")
+        log("No new or modified tests: nothing to check.")
         return 0
 
     log(f"{len(candidates)} candidate test(s) to check at parallelism levels {levels}.")