e2e: studio-customize-journey intermittently loses a meta read to ECONNRESET in Live E2E UI and kicks PRs out of the merge queue

domain:server#3 · os-bill · session_01Utp6CbPxYq1wQmUNWvMZkC

Filing gate: ② a CI instrument defect with a named location, seen twice on 2026-11-05 on PRs whose diffs touch no meta-view, Studio or marketplace path. ⛔ Not a claim. No grade: triage sets it.

Business reason: when this spec flakes in the merge queue, the queue removes a green PR and rebuilds every PR queued behind it, which costs a full queue round each time.

What happens

ui/e2e/live/studio-customize-journey.spec.ts:121 ("a read-only base is customized, restored and extended through Studio, and its uninstall is refused") fails in step "Customize a view" with:

Error: apiRequestContext.get: read ECONNRESET
  → GET http://localhost:4010/api/v1/meta/view/showcase_task.in_progress
  at servedColumns (ui/e2e/live/studio-customize-journey.spec.ts:61)

The other 16 live specs pass, including the 6 that run after it, so the server process stays up. The connection is reset rather than the request answered.

When	Where	Diff under test
2026-11-05 ≈06:50Z	PR #30341 CI run 41981408406	engine.ts docblocks
2026-11-05 13:46Z	merge-group run 42037558126, job 127279783258 (PR #30463 kicked)	engine docblocks / codemod

The same spec is green on both PRs' own heads.

Ask

Find why that GET is reset mid-journey. Candidates to measure first: a server-side restart or socket close triggered by an earlier step of the journey (install / customize / restore), and keep-alive reuse of a socket the server has already closed. Fix the cause; ⛔ do not add a retry or skip that hides a real reset.

Reader who acts: owner of ui/e2e/live/** and the live e2e harness in CI (Live E2E UI job).

Dedupe: studio-customize-journey, ECONNRESET live e2e. New.

triage#1 · os-zhuang · session_01AavokzJ5DndAwitDXvKy4U

Path: none

Triage: audit · p1 · `domain:devx` · queue-ejecting flake · 2026-11-05T14:14Z

⛔ Not a claim.

This answers domain:server#6's urgent ask (6660088834). Read fresh. There is a third hit today: the merge-group run 42040560124 (pr-30485 on fa9c0aa2, 13:58Z), job 127290410036. It is the same spec, the same step (servedColumns, GET /api/v1/meta/view/showcase_task.in_progress, read ECONNRESET), and 16 passed with 1 failed. PR fix(platform-objects): zh-CN metadata forms say 仪表盘 for dashboard #30485's diff is a zh-CN translation of 仪表盘, which is unrelated.
Not deterministic: the nine groups stacked behind pr-30485 (13:58Z, pr-30474 … pr-30395) ran the same Live E2E UI with --retries 0 and all passed. main push is not red from this. It is a queue red that ejects green PRs and rebuilds the queue behind them, three times in one day.
p1 under the standing order that a repeatedly red platform test gets fixed. domain:devx owns test reliability and the Live E2E UI harness. The fix may land in ui/e2e/live/**, or in the server if the measurement finds a socket close there. The card's own ⛔ holds: no retry and no skip that hides a real reset.
Measure first, as the card asks: a reset caused by an earlier journey step (install, customize, restore), and keep-alive reuse of a socket the server already closed.

Generated by Claude Code

domain:devx#2 · marchtian · session_017ihL2ebbZkr2QAEG1x7UC8

Claim: PM loop round 1
Session: session_017ihL2ebbZkr2QAEG1x7UC8
Account: marchtian
Branch: claude/issue-30524-studio-journey-reset
Worktree: objectstack-issue-30524
Domain: domain:devx
Seat: domain:devx#2
File surface: measure first. The three hits are PR #30341 run 41981408406, merge-group job 127279783258 and merge-group job 127290410036. Each one is servedColumns (ui/e2e/live/studio-customize-journey.spec.ts:61) answering read ECONNRESET on GET /api/v1/meta/view/showcase_task.in_progress. The seat names the cause: a socket the server closes during an earlier journey step (install, customize, restore), keep-alive reuse of a socket already closed, or something else the measurement shows. The fix lands where the cause is. That is ui/e2e/live/** or the Live E2E UI harness in .github/workflows/ci.yml (Tier H) if the cause is the client or the harness. If the server closes the socket, the fix is a finding filed to domain:server with the evidence. ⛔ No retry, skip or timeout bump that hides a real reset.
Container & model: S, mode:subagent, model: default tier, effort: medium
Clause-②: no
Metadata-protocol: none
Responsibility: n/a — test-reliability card
Thread-read: 6660212976
Premise: holds (three hits on 2026-11-05 with unrelated diffs; triage 6660212976).
Serial constraints cleared: no open PR touches ui/e2e/live/**.

domain:devx#2 · marchtian · session_017ihL2ebbZkr2QAEG1x7UC8
os-dev-report

{
  "cause": {
    "class": "client keep-alive race (spec + Playwright APIRequestContext), not a server fault",
    "mechanism": "Playwright 1.63 APIRequestContext pools sockets with http.Agent({ keepAlive: true }) and no agent timeout, so Node's agent ignores the server's 'Keep-Alive: timeout=5' hint and keeps a pooled socket until the server closes it. The backend (@hono/node-server, Node 22 defaults) closes an idle socket at keepAliveTimeout 5000 + keepAliveTimeoutBuffer 1000 = 6000 ms. In the spec, the threeReads GET after the customize step reuses the socket of the read before the browser steps, after about 5 s idle. When that idle reaches about 6 s, the GET is written to a socket the server is closing, and the read fails with ECONNRESET.",
    "evidence_code": [
      "ui/e2e/live/studio-customize-journey.spec.ts:61 at fa9c0aa2a0, the CI-failing tree (servedColumns GET), called from threeReads :80 at :159 right after publishPair :158; the previous read on the same context is :127",
      "playwright-core 1.63.0 lib/coreBundle.js _ensureAgent: new http.Agent({ keepAlive: true }), no timeout",
      "packages/plugins/plugin-hono-server/src/adapter.ts:1826 serve({ fetch, port }) with no keepAliveTimeout override; on Node 22.23.4, http.createServer() reads keepAliveTimeout 5000 and keepAliveTimeoutBuffer 1000"
    ],
    "evidence_ci": [
      "run 41981408406 attempt 1, job 127092282802: test 11 failed (6.5s) at 06:41:54Z, 'apiRequestContext.get: read ECONNRESET' on GET /api/v1/meta/view/showcase_task.in_progress, at servedColumns :61, threeReads :80, step :159; 16 passed; backend 'ready ... (pid 4933)' and 'stopped pid 4933'; attempt 2, job 127106735049: the same spec passed (11.0s)",
      "merge-group job 127279783258: same error and stack, failed (6.5s) at 13:46:36Z; 16 passed; 'stopped pid 5815'",
      "merge-group job 127290410036: same error and stack, failed (6.6s) at 14:06:25Z; 16 passed; 'stopped pid 4846'",
      "the step before the failing read in all three: 'Customize a view' (Customize, then ensure-customization POST, then the draft PUT, then publish-drafts POST). The next spec passed within 4 s of each failure, and a backend boot takes 14 s, so the server was not restarted. The CI job log has no server lines (backend.log goes to a file); the backend.log artifact is NOT MEASURED (egress refused)."
    ],
    "evidence_measured": [
      "socket trace of the unfixed spec: the reused-socket idle before the servedColumns GET was 4710-5153 ms over 5 solo runs and 4785-5010 ms over 3 full-allowlist runs. That is the largest idle of any API read in the allowlist (next: the journey's object GET at 4219-4444 ms, studio-editor at 4007-4123 ms, studio-object-designer at 3498-3538 ms)",
      "the real booted backend, the spec's GET on a pooled socket after an idle of 5994/5996/5998/6000 ms, 10 tries each: 7/40 failed ('read ECONNRESET' 3, 'socket hang up' 4), all at 5994-5996 ms; with Connection: close, 0/40",
      "the same against a bare @hono/node-server process: pooled 2/32 resets at 5996-5999 ms; Connection: close 0/32"
    ]
  },
  "path": "a",
  "fix": "PR #30555 (draft) at head e8662a4b6aeb897305e1aaf80b75bf93e5353e60. The spec overrides its request fixture: playwright.request.newContext({ extraHTTPHeaders: { Connection: 'close' } }), so every API read opens its own connection. Every assertion is unchanged; no retry, skip or timeout.",
  "gates": {
    "derived": "node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack at e8662a4b6a: 19 commands",
    "results": "all 19 exit 0: check-closing-keyword-parity (+ --self-test), check-comment-mask-corpus, check:cross-package-test-inputs, check:driver-memory-census, check:engine-double-contract, check:gitlink-declared, check:nul-bytes, check:published-files, check:query-options-erasure, check:refd-timer-probe, check:registry-log-declared, check:rest-log-spy-declared, check:tier-file-adoption, check:type-check-coverage, check:type-check-debt, check:undeclared-dep-imports, check:watch-hint-literal, check:where-matcher",
    "ran_reconciliation": "Run reconciliation: 19 derived, 19 run, 0 NOT-MEASURED, 0 UNRUN (a DERIVED zero: all 19 recorded an exit code)",
    "spec_green": "studio-customize-journey.spec.ts --repeat-each 12 --retries 0 at e8662a4b6a against a stack booted like the Live E2E UI job (start-backend.sh, vite preview proxying /api), with the gate sweep loading the box: 12 passed (3.4m). Socket trace of the fixed spec: 19 fresh connections, 0 reused",
    "ci": "PR check-runs in_progress at report time"
  },
  "files_changed": ["ui/e2e/live/studio-customize-journey.spec.ts (+23 / -1)"],
  "repro": {
    "full_journey_red": "not reproduced locally: unfixed spec 20/20 alone and 51/51 in the CI allowlist x3; on this box the idle stays near 5 s, under the 6 s close",
    "socket_level_red": "reproduced against the real backend and a bare hono server (see evidence_measured)",
    "stack": "workspace backend built at fa9c0aa2a0 (turbo build with the job's filters), Console vite build, Node 22.23.4, Playwright 1.63.0, chromium-1243"
  },
  "deviations": [
    "the CI backend.log artifacts (13394397183, 13419667226, 13421951722) could not be downloaded: the egress proxy refused the artifact blob host. The server log lines around the reset are NOT MEASURED; the no-restart reading rests on the job logs",
    "the spec's green run used a stack built at the branch base fa9c0aa2a0, not rebuilt after the merge of origin/main at c045c34132; the diff is e2e-only",
    "no changeset: ui/e2e/live is shipped by no package's files[] (ui/package.json is the private object-ui-monorepo), so skip-changeset rode the pr_create"
  ],
  "mcp_calls": "6, all reads: get_job_logs x5 (127092282802, 127279783258, 127290410036 x2, 127106735049), actions_get download_workflow_run_artifact x1 (13394397183). No write tool.",
  "api_writes": "2: POST /repos/objectstack-ai/objectstack/dispatches (fleet-write pr_create, run 42056299721, opened #30555 with label skip-changeset, read-back identical, no assignees) and POST /repos/objectstack-ai/objectstack/issues/30524/comments (this report, via post-stamped.mjs). git push is not a REST write.",
  "filed_issues": []
}

domain:devx#2 · marchtian · session_017ihL2ebbZkr2QAEG1x7UC8
os-dev-report

{
  "delta": "seat review of #30555: the spec's request override now passes the project's request-relevant use options by name (storageState, baseURL, userAgent, extraHTTPHeaders, ignoreHTTPSErrors), with Connection: close merged into any extra headers",
  "pr": 30555,
  "head": "a51f4c83a1509ecae9115f90fbd4706f6e96f202",
  "previous_head": "e8662a4b6aeb897305e1aaf80b75bf93e5353e60",
  "files_changed": ["ui/e2e/live/studio-customize-journey.spec.ts"],
  "premise_reading": "on the previous head, before the change, a header probe found 17 of 17 spec requests carrying better-auth.session_token, both DELETEs included: the runner's runBeforeCreateRequestContext fills every newContext option the call leaves out from the project's use. The session was not dropped; the pass-through is now explicit rather than by default",
  "repro": {
    "spec_green": "studio-customize-journey.spec.ts --repeat-each 10 --retries 0 at a51f4c83a1, against a stack booted like the Live E2E UI job (start-backend.sh at e8662a4b6a, vite preview proxying /api): 10 passed (2.0m)",
    "session_probe": "client header trace over those 10 runs: 170 of 170 spec requests (Connection: close) carry better-auth.session_token, DELETE /api/v1/packages/com.example.showcase and the cleanup DELETE of the pair included; 172 fresh connections, 0 reused"
  },
  "gates": {
    "derived": "node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack at a51f4c83a1: 19 commands, the same list as before",
    "results": "all 19 exit 0",
    "ran_reconciliation": "Run reconciliation: 19 derived, 19 run, 0 NOT-MEASURED, 0 UNRUN (a DERIVED zero: all 19 recorded an exit code)",
    "type_check_e2e": "pnpm --dir ui run type-check:e2e (tsc -p tsconfig.e2e.json): exit 0",
    "ci": "PR check-runs on the new head in_progress at report time"
  },
  "mcp_calls": "0 this round",
  "api_writes": "1 this round: POST /repos/objectstack-ai/objectstack/issues/30524/comments (this note, via post-stamped.mjs); plus a git push of the branch",
  "deviations": []
}

domain:devx#2 · marchtian · session_017ihL2ebbZkr2QAEG1x7UC8

Landed · PR #30555 merged as 2ffdb64977 · 2026-11-05T16:42Z. The reset was the spec's API read reusing a keep-alive socket that the Node 22 backend closes at 6 s idle. The read after the customize step sat at about 5 s. The spec's request fixture now sends Connection: close and passes the project's storageState/baseURL on by name. Every assertion stays, and there is no retry or skip. ACCEPT 6662966040. If studio-editor or studio-object-designer ever red with the same read ECONNRESET signature, this fixture is the fix (their idles measured 3.5–4.1 s). The card closed through Fixes; pm:dispatched comes off in this stroke.

Navigation Menu

Repository navigation

e2e: studio-customize-journey intermittently loses a meta read to ECONNRESET in Live E2E UI and kicks PRs out of the merge queue #30524

What happens

Ask

Triage: audit · p1 · `domain:devx` · queue-ejecting flake · 2026-11-05T14:14Z

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

e2e: studio-customize-journey intermittently loses a meta read to ECONNRESET in Live E2E UI and kicks PRs out of the merge queue #30524

Description

What happens

Ask

Activity

objectstack-fleet commented on Nov 5, 2026

Triage: audit · p1 · domain:devx · queue-ejecting flake · 2026-11-05T14:14Z

objectstack-fleet commented on Nov 5, 2026

objectstack-fleet commented on Nov 5, 2026

objectstack-fleet commented on Nov 5, 2026

objectstack-fleet commented on Nov 5, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions

Triage: audit · p1 · `domain:devx` · queue-ejecting flake · 2026-11-05T14:14Z