Summary

• deflake the dashboard pub/events websocket broadcast regression test
• split the previous two-websocket TestClient scenario into deterministic contract checks:
  • /api/pub forwards publisher frames to _broadcast_event
  • _broadcast_event fans out a frame to all registered channel subscribers

Why

The old test opened /api/events and /api/pub websockets at the same time through Starlette's TestClient, then waited for the subscriber client to receive the forwarded frame. Under load this can time out even though the production split points are simple: /api/pub receives a frame and calls _broadcast_event, while _broadcast_event sends the payload to each registered subscriber.

The replacement keeps those two behavior contracts covered without depending on a racy multi-websocket TestClient receive path.

Scope

Changed file:

• tests/hermes_cli/test_web_server.py

No production code changes.

Verification

• original test reproduced locally as flaky: failed once during repeated runs
• new /api/pub forwarding test: passed 30 repeated runs
• new _broadcast_event fanout test: passed 30 repeated runs
• python -m pytest tests/hermes_cli/test_web_server.py::TestPtyWebSocket -q -o 'addopts='
  • 13 passed
• python -m py_compile tests/hermes_cli/test_web_server.py
• git diff --check
• independent review: approved with no blockers

Related

Related to the failing test (2) CI shard observed on #37350.