E2E Shard Failures – Run 21377510901 (PR 550)

Issue: CI shards are failing/flaking against Docker environment (localhost:8080) while local runs pass. Need root-cause plan without re-enabling Vite/coverage. Status: 🔴 ACTIVE – Planning Priority: 🔴 CRITICAL – CI blocked Created: 2026-01-27

🔍 CI vs Local Findings

Shard 1 (passed but flaky): tests/core/access-lists-crud.spec.ts intermittently misses toast / ACL visibility assertion.
Shard 2 (hard fail): emergency-server/*.spec.ts and tier2-validation.spec.ts hit ECONNREFUSED ::1:2019/2020; access list creation returns "Blocked by access control list".
Shard 3 (fail): tests/core/account-settings.spec.ts certificate email validation – error message not visible after retries.
Shard 4 (fail):
- tests/core/system-settings.spec.ts success toast not observed.
- tests/core/user-management.spec.ts invite/resend flows fail with strict mode locator collisions (multiple matching buttons).
Container logs (shard 2 artifact): Emergency server disabled (CHARON_EMERGENCY_SERVER_ENABLED=false) and emergency bypass called. Tier-2 server (port 2020) never starts → explains connection refusals. Security ACL reported as disabled post emergency reset but initial access-list calls still 401/blocked until login.
Environment parity: Local likely starts emergency server (or uses 127.0.0.1), CI disables it via env; CI uses IPv6 loopback (::1) causing refusals when service is off.
Architecture: Vite/coverage already removed; tests target Docker app only.

🧭 Hypotheses

Emergency server/tier2 disabled in CI → all shard-2 tests fail; local enables by default. Root cause: env var CHARON_EMERGENCY_SERVER_ENABLED is false in e2e compose or workflow.
ACL bypass timing → initial emergency reset happens, but ACL state may still block access-list creation; needs deterministic disable hook.
UI assertion drift → account-settings/system-settings/user-management expectations mismatch current UI text/roles; strict-mode locator ambiguity for invite buttons.
Toast race / network latency → success toasts not awaited with retryable locator; CI slower than local.

🎯 Action Plan (phased)

Phase 1 – Environment parity (CI vs local)

Enable emergency server in CI Docker stack: set CHARON_EMERGENCY_SERVER_ENABLED=true, expose admin port 2019 and tier-2 port 2020, and ensure services bind for both IPv4/IPv6 (CI uses ::1).
Explicitly set emergency token for tier-2 if required; document its source (redacted) in test env.
Add startup assertion in global-setup to poll http://localhost:2019/config/ and http://localhost:2020/health (skip if disabled) with short timeout to fail fast.
Capture env snapshot in CI logs for emergency-related vars (redact secrets) and note resolved base URL (IPv4 vs IPv6).

Phase 2 – Deterministic security disable

After login/setup, call emergency reset and then verify ACL/rate-limit flags via /api/v1/security/config before continuing tests; make this idempotent and fail fast before any data creation.
If ACL still blocks create, call /api/v1/access-lists/templates to assert 200; otherwise retry emergency reset once and fail with clear error.
Add small utility in TestDataManager to assert ACL is disabled before creating ACL-dependent resources; short-circuit with actionable error.

Phase 3 – Shard-specific fixes

Shard 2: Once emergency server enabled, rerun to confirm. Add health check for tier-2 server; fail early if down.
Shard 1: Wrap ACL toast assertions with expect.poll/toHaveText on role-based toast locator; ensure list refresh after create. Add a shared toast helper (role-based with short retries) to reuse across specs.
Shard 3: Update certificate email validation assertion to target the visible validation message role/text; avoid brittle getByText timeouts.
Shard 4:
- System settings toast: use role-based toast locator with retry; ensure the form submit awaits network idle before assert.
- User management invite/resend: replace ambiguous button locators with role+name scoped to each row (e.g., row locator then getByRole('button', { name: /resend invite/i })); add a row-scoped locator helper to avoid strict-mode collisions.

Phase 4 – Observability and flake defense

Add Playwright trace/video for shard 1–4 in CI (already default? confirm); keep artifacts for failing shards only to save time.
Log emergency server state (enabled/disabled), ACL status, and resolved base URL (IPv4 vs IPv6) at start of each project.
Add short retries (max 2) for toast assertions using auto-retrying expect.

Phase 5 – Validation loop

Rerun shards 1–4 in CI after env toggle; compare to local.
If shard 2 passes but others fail, prioritize locator/UX updates in phases 3–4.
Keep Vite/coverage off until all shards green; plan separate coverage job later.

📄 Files/Areas to touch

Workflow/compose env: ensure CHARON_EMERGENCY_SERVER_ENABLED=true; expose tier-2 port 2020; confirm emergency token variable passed.
tests/core/*: adjust locators and toast assertions per shard notes.
tests/utils/TestDataManager.ts: add ACL-disabled check before ACL creation.
global-setup.ts (if needed): add emergency server health probe and state logging.

✅ Completion checklist

CI env starts emergency server (port 2020) and admin API (2019); health probes added.
Security disable verified before data setup; ACL create no longer blocked.
Shard 1 toast flake mitigated with resilient locator/wait.
Shard 2 emergency/tier2 tests pass in CI.
Shard 3 account-settings validation assertion updated and passing.
Shard 4 system-settings toast and user-management locators stabilized.
Vite/coverage remain off during fixes; add a guard/checklist item in workflow to ensure coverage flags stay disabled during triage; plan coverage follow-up separately.

📎 Artifacts reviewed

GH Actions log: .agent_work/run-21377510901.log
Docker logs (shard 2): .agent_work/run-21377510901-artifacts/docker-logs-shard-2.txt (shows emergency server disabled, ACL reset attempts)

6.1 KiB Raw Blame History Unescape Escape