Fix remaining Outerloop failures in DCP, CLI, and Dashboard tests by radical · Pull Request #20466 · microsoft/aspire

Ankit Jain (radical) · 2026-09-25T01:13:34Z

Four repeat failures remained in the Outerloop workflow:

The SDK 'Aspire.AppHost.Sdk/14.0.0-ci' specified could not be found.

Watch task over Kubernetes ContainerExec resources terminated unexpectedly.
Polly.Timeout.TimeoutRejectedException: The operation didn't complete within the allowed timeout of '00:01:00'.

Expected rows not found: apiservice, webfrontend, cache

ScrollButtons_SmoothScrollHidesImmediately_AndReachesNewBottom
timed out after reaching the limit of 30000ms.

Root causes

LocalArchive test installs must preserve the package build identity. PR packages use a pr-N hive that matches the CLI's baked channel. Non-PR CI packages such as 14.0.0-ci use the local hive, but their CLI archive retains a baked daily identity. User configuration does not override the environment, install-sidecar, or assembly identity, so generated AppHosts selected daily feeds and could not resolve the CI SDK from the local hive.

DCP watch connections can stall while waiting for their initial HTTP response. Each connection attempt needs the API timeout so a stalled request is canceled and retried, while an established stream must remain active until its longer periodic restart. Factory or enumeration cancellation unrelated to that periodic restart must propagate instead of being silently retried.

The template tests matched .main-grid .resource-row. Fluent data-grid copies the configured row class onto optimized plain cells, so the locator returned both tr rows and td cells. A matched cell has no descendant grid cells, causing resource extraction to fail even though the dashboard displayed healthy resources.

The smooth-scroll test required observing an intermediate native animation frame. Windows headless Chromium can skip that frame under load even though the scroll request and scrollend behavior are correct.

Fix

LocalArchive installs now let the installer derive pr-N or local from the package version. Non-PR archives explicitly override ASPIRE_CLI_CHANNEL, ASPIRE_CLI_VERSION, and ASPIRE_CLI_PACKAGES so generated projects consume the local package directory. PR archives retain their automatically selected pr-N identity. Focused tests cover both package identities, both install command paths, and the exact local identity override.

DCP watches now apply the API timeout to each connection attempt and retry timeout failures until the periodic restart token is canceled. Once connected, the watch stream remains governed by the periodic restart token. PeriodicRestartAsyncEnumerable recreates factories canceled by that restart token, while cancellation from any other source propagates without retrying. Focused regressions require a second connection attempt after a deliberately stalled watch and cover both periodic-enumerable overloads.

Template resource checks now restrict the locator to .main-grid tr.resource-row and keep grid-cell lookups scoped to each matched row.

The Playwright test captures the requested smooth-scroll destination and dispatches scrollend explicitly after growing the content, preserving the product contract without depending on compositor timing.

Validation

CliInstallStrategyTests: 61 passed
Aspire.Deployment.EndToEnd.Tests build: succeeded
Aspire.Templates.Tests build: succeeded
KubernetesServiceTests and PeriodicRestartAsyncEnumerableTests: 14 passed
Stalled-watch retry regression: 10/10 repeated runs passed
Independent cancellation regressions: 20/20 repeated runs passed
DashboardInteractionsTests: 13 passed
Smooth-scroll regression: 10/10 repeated runs passed

Surprises

The intermittent terminal WebSocket failures are intentionally unchanged. They moved between operating systems and test cases across reruns, and no reproducible root cause was established.

github-actions · 2026-09-25T01:26:50Z

🚀 Dogfood this PR with:

⚠️ WARNING: Do not do this without first carefully reviewing the code of this PR to satisfy yourself it is safe.

curl -fsSL https://raw.githubusercontent.com/microsoft/aspire/main/eng/scripts/get-aspire-cli-pr.sh | bash -s -- 20466

Or

Run remotely in PowerShell:

iex "& { $(irm https://raw.githubusercontent.com/microsoft/aspire/main/eng/scripts/get-aspire-cli-pr.ps1) } 20466"

Karol Zadora-Przylecki (karolz-ms) · 2026-09-25T02:24:05Z

KubernetesService and DCP watcher changes look good. CLI tests seem unhappy though.

github-actions · 2026-09-25T05:05:55Z

Retrying the failed CI jobs for this pull request from the CI run attempt. The rerun is being tracked in the rerun attempt.

github-actions · 2026-09-25T06:44:47Z

Retrying the failed CI jobs for this pull request from the CI run attempt. The rerun is being tracked in the rerun attempt.

github-actions · 2026-09-25T17:51:24Z

Retrying the failed CI jobs for this pull request from the CI run attempt. The rerun is being tracked in the rerun attempt.

github-actions · 2026-09-25T18:09:43Z

Retrying the failed CI jobs for this pull request from the CI run attempt. The rerun is being tracked in the rerun attempt.

LocalArchive E2E installs copied built packages into the local hive but left the CLI on its baked channel, so aspire add could not resolve the 14.0.0-ci AppHost SDK. DCP watches also inherited the one-minute non-streaming API timeout even though the Kubernetes client buffers their long-lived response body. The dashboard smooth-scroll test separately depended on observing an intermediate compositor frame that Windows headless Chromium can skip. Select the local hive after LocalArchive installs, keep watch requests alive until their periodic restart, and drive the scrollend contract deterministically in the Playwright test. Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>

LocalArchive installs can auto-detect a PR package suffix and place packages in a pr-N hive, which conflicts with configuring the installed CLI to use the local hive. Long-lived watch factories can also be canceled by their periodic restart token before returning an enumerable. That cancellation currently escapes and terminates the outer watch instead of recreating it. Force LocalArchive installs into the local hive and restart factories whose restart interval expires. Cover both LocalArchive command paths and both periodic-enumerable overloads. Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>

The periodic factory restart handler currently treats every cancellation as a restart while the outer token is active. A factory canceled for another reason would be silently retried instead of propagating its failure. Handle cancellation only when the periodic restart token fired, and verify that unrelated cancellation propagates for both enumerable overloads. Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>

PR archive tests install pre-downloaded packages through LocalArchive. Forcing those packages into the local hive breaks the CLI's baked pr-N channel, so generated AppHosts cannot resolve the PR AppHost SDK. Allow the installer to infer the hive from the package version and only configure the global local channel for non-PR archives. Cover both PR and CI package identities with focused tests. Periodic restart handling also retried independent cancellation in the struct factory and class enumerator paths. Require the periodic token to be canceled before restarting, and verify unrelated cancellation is propagated without retrying the factory. Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>

Ankit Jain (radical) added this pull request to stack #20456 September 25, 2026 01:13