Skip to main content
Finance, Inventory, Contacts, CMS, and Tasks depend on the live Web API. A successful satellite build does not prove that this shared dependency is usable. Production releases must preserve daily enterprise operations during optional download-provider failures.

Ordinary operations and bulk downloads

Ordinary GET and HEAD requests never enter the optional offline-download guard, even when OFFLINE_DOWNLOAD_PROTECTION_ENABLED is enabled. Explicit bulk requests marked with the offline-download header retain their protection and return a retryable failure when its provider is unavailable. Required MFA, verified IP blocks, and baseline rate limits remain effective. This rule applies to Web and the registered satellite API proxies. It prevents optional Redis, quota, or Turnstile failures from blocking normal pages and routine reads. It does not change permission checks or authorize bulk exports. The real proxy regression suite is packages/utils/src/__tests__/daily-operation-download-isolation.test.ts.

Production domain assignment

The Web, Finance, Inventory, Contacts, CMS, and Tasks production workflows stage production artifacts with vercel deploy --prod --skip-domain. They then run scripts/ci/verify-staged-critical-app.js against the immutable deployment URL. The gate requires production build metadata for the exact workflow SHA and app. GET and HEAD probes with a deliberately invalid machine-key token must reach the normal JSON API authentication boundary and return 401. Unexpected statuses, redirects, HTML, missing metadata, and mismatched source identity stop the workflow before domain assignment. Only a passing staged deployment is promoted, then recorded as successfully deployed. Vercel staging and promotion retain the same built artifact. The CLI probe uses deployment-protection bypass through vercel curl; it supplies no application valid session and reads no customer records. Public GET/HEAD build metadata reads do not require satellite cookies. CMS machine-key requests remove ambient cookies from the guard and forwarded request, so invalid keys cannot borrow a browser actor and expired cookies do not interfere with independent API-key verification. JWT-shaped headers retain browser session refresh and assurance checks; their shape is not proof of authentication. Public metadata reads discard incidental session credentials. CMS middleware regressions cover metadata access, invalid machine-key forwarding, guard denial, and browser assurance failures in apps/cms/src/proxy.test.ts. When the production planner selects Web, every selected Vercel satellite waits for Web’s explicit promoted output. A failed, cancelled, skipped, or package-release-blocked Web deployment prevents satellite rollout. When Web is not selected, satellites can deploy independently. Cloudflare jobs keep their own dependency rules. These probes establish source identity and API ingress health. They do not prove signed-in permissions, business CRUD, mobile behavior, or provider integrations. The Web probe includes Learn’s Programming catalog API: its sidebar page must not advance against a platform artifact that lacks the route. An out-of-band Instant Rollback can still create mixed versions, so recovery must restore a compatible platform artifact and verify the signed-in Programming sidebar flow. An HTML API 404 is not evidence that the Learn page or workspace is missing. After promotion, verify the requested signed-in workflows separately and record any missing evidence. Production database migrations still require their gated exact-SHA workflow; never push them manually from an agent checkout. Regression coverage lives in scripts/ci/critical-rollout-workflow.test.js, scripts/ci/verify-staged-critical-app.test.js, and the existing release workflow suite. The dependency test exercises selected and unselected Web, failed and cancelled runs, missing promotion output, failed planning, and unselected apps.

Lettin Worker health and cache initialization

Lettin uses the existing URL-pinned OpenNext adapter shared with Parley. Its cache registration patch installs the composable handlers and returns before the original Next loadCustomCacheHandlers loader. Replacing only minified declarations leaves references to removed aliases; after bundling those names can capture unrelated module locals and attempt to import file:/.next/None. This failure occurs before /api/v1/lettin/health reaches D1 or R2 and can return an empty 500. Run node --test apps/lettin/scripts/opennext-health.node.mjs after adapter or Next updates. The regression patches and executes the installed Next production route and page cache methods, registers the composable handlers, and rejects fallthrough to the original loader. Lettin CI runs it before the real Worker build. Keep the shared adapter patch and its exact URL dependency selected by the workflow matcher. To distinguish the boundary from a storage failure, replay a retained artifact privately with no outbound requests, local disposable D1 and R2, and a synthetic worlds(id) table. Preserve the original and record runtime/version, normalization and fixture changes. In the retained-artifact control, replacing only cache registration changes health from empty 500 to 200; removing worlds then reaches the health handler and returns its JSON Request failed 500. This proves the cache boundary defect without establishing production storage health. Missing retained WASM must be supplied from the matching installed Next package and recorded as a harness dependency, not represented as original artifact bytes. Require exact-commit CI, the emitted candidate Worker and hosted health verification with deployment identity before claiming recovery. Keep production D1 migration and authenticated workflow acceptance separate; an empty 500 alone does not authorize database repairs.