I

Diagnose and bound stalled Playwright project fixture creation

Found on 2026-07-17 while landing PR #2075. The PR's TUI spec passed on its first attempt, but an unrelated browser spec was reported as flaky inside an otherwise-green preview check:

  • Evidence check: Cloudflare Previews / deploy + e2e (Depot job).
  • Depot attempt: cgw3bdlr53; test-results artifact: 019f7090-8f07-707c-a531-488f39e9051b.
  • Test: specs/repl-examples.spec.ts → runs "discover-tree" through the project REPL.
  • First attempt: worker 7, started 2026-07-17T14:49:58.765Z, timed out after 90,037ms, with a completely blank screenshot and no stdout/stderr.
  • Playwright retry: worker 8, started 2026-07-17T14:51:29.392Z, passed in 10,800ms with the same example.

Artifact files:

  • playwright-results.json contains both attempts and the exact timings.
  • playwright-output/repl-examples-itx-REPL-cat-52d08-ee-through-the-project-REPL-chromium/test-failed-1.png is the blank first-attempt screenshot.
  • The corresponding error-context.md contains only the outer 90-second test timeout.
  • ...-retry1/trace.zip covers the successful retry only because playwright.config.ts currently uses trace: "on-first-retry".

Server-side evidence#

Cloudflare Workers observability was queried for account 376ef7ed81b0573f93524de763666c15, service os-preview-4, and the bounded window 2026-07-17T14:49:50Z–14:51:35Z.

A case-insensitive all-dataset search for repl-discover-tree returned only two events, both at 2026-07-17T14:51:30.687Z after the retry began:

  • [create-timing] {"slug":"repl-discover-tree-mrp2322f-f1f0ef19","step":"auth-register","ms":317,"ok":true}
  • the matching PROJECT_DIRECTORY KV write on trace d816008d61a5e794f53445a480d1bc39.

There is no repl-discover-tree slug event during the failed attempt. Project creation for the successful retry reached OS immediately, while the first attempt never completed the earliest logged auth-register create step. This fits the blank page: helpers.createFixture() runs before page.goto(). It does not implicate the REPL example or this PR's terminal work.

The relevant client path is:

  1. specs/test-support/test.ts → helpers.createFixture
  2. specs/test-support/forged-session.ts → createProjectFixture
  3. createAdminProject → await created.__describe()

packages/iterate/src/itx/itx-node-client.ts already configures a 15-second WebSocket handshake timeout. A local socket that accepted TCP but never answered the upgrade rejected an in-flight RPC after 15,011ms with WebSocket connection failed., so the basic client handshake bound works. The recurrence below identifies a server-side terminal failure instead of a client-handshake defect.

2026-07-23 recurrence: server-side terminal failure identified#

PR #2265's second formal preview run reproduced the outer 90-second fixture timeout in empty agent feeds distinguish waiting from filtered zero matches. This time the named timing and trace evidence identify the exact project and terminal state:

  • Failed project: prj_ddc2a50bba2941878bed307b207417ea (slug agent-initializing-mrwsg32j-f065698a).
  • Root trace: eb26b8dc5ae96c5da9373abef6e8606d on os-preview-11.
  • Project.create remained pending for 88.3 seconds until Playwright canceled the request. Project birth took 37.7 seconds; config-repo birth took 36.3 seconds.
  • Inside config-repo birth, artifact-get-or-create failed after 32.7 seconds. The config stream then recorded repos/create-failed with An internal error occurred. and cross-posted it to the root stream. It never recorded repos/created or terminal project/created, including after later processor wakes.
  • The test retry created prj_20587…, completed birth in 3.8 seconds, and passed. The test retry therefore hid a defective first project; this was not merely a slow browser or a lost readiness notification.

Cloudflare's binding contract classifies this as Artifacts INTERNAL_ERROR. Repo creation previously treated every Artifacts error except a still-materializing repo as a permanent domain failure. PR #2265 changes the processor to redeliver the already-idempotent creation obligation for INTERNAL_ERROR and UPSTREAM_UNAVAILABLE, while preserving terminal fail-closed behavior for invalid input and other domain errors. The focused recovery test proves that the first service error appends no create-failed, the next delivery succeeds, and exactly one repos/created fact is recorded.

Remaining proof#

  • Prove the fix with repeated retry-disabled preview runs and zero unexplained error/retry telemetry.

Was this page helpful?

Powered by Docsbook