Flaky Agent Host E2E: 'identical plugin republished under a different URI does not restart servers' ends 'stopped' (hit an official insider) #344208

Summary

Agent Host E2E — Copilot › runtime MCP: identical plugin republished under a different URI does not restart servers is flaky. The MCP server ends up stopped when the test expects ready, and there's no restart: processStarts: 1, restartStates: [].

AssertionError [ERR_ASSERTION]: Expected values to be strictly deep-equal:
  { processStarts: 1, restartStates: [], state: 'stopped' }   // actual
  { processStarts: 1, restartStates: [], state: 'ready' }     // expected

Hits

• Official build: 489097, the Mon Oct 26 05:00 UTC releasing insider, in macOS ARM64 Agent Host E2E Tests (Electron) (AcesShared 4). A failed test job blocks Publish until someone reruns it.
• GitHub PR CI: 1 on macOS / Electron (Oct 21, unrelated PR benibenj/agents/artifact-stack-03-ui). Also 16 in the Oct 21 Windows 16/32/64-core runner benchmark runs (agents/electron-integration-parallel-ben…, ci: move PR checks to GitHub-hosted 8-core runners #343094), so heavy parallel load makes it much more likely.
• def 111 main CI: no failures since Oct 10, apart from the official build above.

The failure is stopped with no restart recorded, which reads like a race between "republished under a different URI" handling and the readiness wait, rather than an unwanted restart. That's an inference, not verified. The test came from #342634.

Co-authored with Copilot