Skip to main content

Current operating decision

The supported application path is apps/web on Next.js (scheduled for native TanStack Start + Void + Vite+ on Cloudflare Workers), the satellite apps (Next.js today, moving per app to native TanStack Start + Void + Vite+ on Cloudflare Workers), shared TypeScript packages, and their owning hosted services. The Docker cron runner and apps/tanstack-web are retired; do not revive them. Docker setup and apps/backend (Rust) are not in use until explicitly resumed. They are not prerequisites for documentation, routine development, focused validation, or release delivery. This applies to the live app surfaces, developer onboarding, and incident runbooks. Retained files, Compose services, scripts, dashboards, route manifests, and future-cutover procedures are historical or resumption references. Their presence does not establish an active service. An explicitly isolated experiment or separately approved resumption can use those artifacts; it must not silently become a prerequisite for the active path. TanStack Query remains in use for client fetching/caching. Per-app TanStack Start on Cloudflare Workers is the active target for every web app, including apps/web (scheduled); it is unrelated to the retired apps/tanstack-web tree, whose source refreshes, tests, dependency updates and CI stay disabled. Shared contracts still belong in reusable packages; do not require updates to the retired trees for a live change. Each app migrates under the native Start migration runbook: one integration writer per app, next/* only through @tuturuuu/start-bridge, basePath readiness for both its subdomain and tuturuuu.com/<app>, and exact-commit CI, sealed emitted-Worker acceptance and the app’s reviewed hosted cutover acceptance. Wave-1 uses a private production smoke before routing, then immediate signed-in production acceptance and a 30–60 minute watch; hosted staging is not a flip gate. Current hosting (Vercel or Next.js/OpenNext Workers) stays canonical until that app’s explicit cutover PR. See Oxc and Cloudflare adoption.

Scope and readiness

  • Browser apps take apex mounts; services keep their own ingress and are not browser mounts.
  • Qualification, upload admission, actual upload, subdomain cutover and apex cutover are separate tracked facts per app.
  • Apps drain through a continuously evaluated per-app readiness queue; there is no global lowest-risk-first order or fixed wave labels.
  • Post-deploy watch is 30–60 minutes with immediate rollback, not 5 minutes.

Wave-1 production publication hold

Calendar frontend is admitted for route-free production preparation in productionPublicationPolicy in tuturuuu.ci.ts; Meet, Mail and Apps remain held for wave-1 cutover. The production planner reports Held with that reason and selects neither their native nor legacy publishers, including manual and package recovery selection. Native CI validation stays enabled. Meet’s direct production job also checks the same policy before publication. A production sync therefore retains the currently deployed Meet frontend and realtime Worker versions and the existing Calendar/Mail/Apps Vercel deployments. Other apps and central API suppliers retain their normal production behavior; the hold is not an assertion of compatibility with newly published suppliers. Vercel Git deployment stays disabled; do not change framework, domains, bindings or scheduler ownership. See the wave-1 cutover runbook.

Runtime map

Wave-1 native staging publishers

The Meet, Calendar, Mail and Apps Start/Void source and scoped native CI are active. This does not revive apps/tanstack-web, resume the Rust/Docker stack, or authorize canonical cutover. Calendar, Mail and Apps have separately sealed native staging/production profiles. Calendar has a separately gated private production publisher, which never attaches routes. Other wave-1 apps remain held. Only owner-provisioned, explicitly enabled staging can publish from trusted main runs; no workflow creates provider resources or changes DNS as setup. The native publisher runbook records exact identities, secret names, resource continuity, archive/runtime gates, stop control, rollback receipts and remaining hosted acceptance. Meet’s off-canonical profile is defined separately and has no automated staging publisher here.

Drive staging candidate

apps/drive has an additive Start/RSC bridge candidate. Its current Next app continues to own production. void.config.ts defines the route-free staging profile with workers_dev: false; only the Drive staging host is admitted at the Worker root. No DNS, custom domain, production route, publisher or infrastructure provisioning is part of this candidate. The candidate uses the staging Supabase project pinned in supabase-staging.yaml, forwards storage to the central staging API, and refuses missing or mismatched runtime origins before loading routes. The Vite config also rejects non-staging build origins. Supply only the staging publishable key and staging server secrets; never reuse production credentials. Candidate response cookies stay host-only, with Secure and HSTS. Outgoing Domain cookies are rejected, including prefixed cookies. Global Supabase cookie configuration remains deferred to the issuer migration. Staging responses are private and unindexable. English and Vietnamese routes, Drive session audience and central API ownership remain unchanged. Storage still uses the existing central APIs, path guards and storage RLS from #6879. No candidate-specific storage grants, bucket changes, service-role bypasses or migrations are introduced. Focused regression tests live in apps/drive/src/start-candidate and the Supabase cookie adapter tests. Before cutover, require exact-commit CI and native build acceptance, independent security review, emitted Worker and browser acceptance, central staging auth/API availability, and an explicit promotion/rollback review. A separately authorized staging deploy must verify cross-app login, refresh/logout and cookie deletion, wrong-app/expired sessions, bilingual navigation, upload/list/download/share, consent and private/cross-workspace storage denials against the staging database. Unit fixtures do not establish hosted session or storage RLS acceptance. Keep Wrangler for Worker runtime checks until cf CLI operation parity is verified.

Hosted scheduler map

New schedules go to Cloudflare; the background work inventory lists every remaining legacy owner and its migration target. The Cloudflare Cron Control guide explains the idle skip, requeue RPC, pending-batch check, and delivery token. Calendar’s Infrastructure gateway authenticates the scheduler and uses its configured credential to reach Calendar. It must not be bypassed by adding a second direct Calendar schedule. vercel: false in the old Web cron catalog excludes that entry from Web’s Vercel scheduler. It does not mean the retired Docker runner will execute it. A route or enabled source configuration is not proof of an active schedule. If a job appears only in retained Docker configuration, identify its hosted owner explicitly before calling it active. Do not activate a replacement or change production schedules merely to resolve documentation ambiguity.

Diagnose missed work without Docker

  1. Identify the owning scheduler declaration and deployed revision.
  2. Inspect the hosted invocation result, authentication failure, timeout, and retry behavior for the relevant bounded time window.
  3. Trace the processor/gateway response and terminal domain state. For Calendar, confirm a fresh provider-sync timestamp with all clients closed. For notifications, confirm claimed/terminal delivery state rather than merely a scheduled invocation.
  4. Use an authorized, scoped manual processor recovery only when necessary; retain deduplication and recipient/workspace rules.
  5. Report missing deployment or scheduler evidence separately from source correctness.
Stale tmp/docker-web/cron/status.json, old runner heartbeats, and queued local Docker recovery requests are not health signals for these hosted schedulers. Retained tooling may still contain runner/watchdog code. The standard Docker commands, Docker setup CI, Docker-backed E2E jobs, inactive implementation test suites, and migration refresh commands are disabled. This does not delete the retained files or cancel already-running external jobs.

Regression and maintenance evidence

node --test scripts/ci/paused-migration-workflows.test.js guards the Docker setup and migration pause, including manual workflow jobs. scripts/docs-audit.test.js guards the active-runtime decision, links, navigation, and reproducible documentation/SEO inventories. Hosted job configuration and actual execution evidence remain necessary for operational claims. Without Docker gives the minimal development path.

Update and test pause until further notice

Default Node script-test discovery excludes inactive Docker/Rust/TanStack implementation suites while retaining tests that verify the pause. Shared Next.js unit tests, type checks, lint, and app builds stay active. The old Docker-backed E2E workflow is paused, so it no longer establishes browser integration proof. Direct Playwright commands can target an already prepared, safe local Next.js environment without starting Docker; their seeded database requirements still apply. A hosted build is not a replacement for that missing integration evidence. bun update-all targets maintained workspace names explicitly. Legacy shell and batch entrypoints delegate to the same updater. Renovate ignores inactive app manifests and Docker files; broad lockfile maintenance is disabled to avoid refreshing retained dependencies. Shared lockfile resolutions can still reflect active package changes and must be reviewed separately from paused manifests. Generated documentation inventories do not refresh Rust/TanStack source or versions. Existing migration runbooks remain retained reference material. Resumption requires an explicit instruction and a coordinated change to the command guards, test discovery, CI job/switchboard gates, dependency policy, docs inventory, and operating docs. Do not selectively re-enable one gate and assume the rest of the runtime is maintained. Release Please excludes the inactive TanStack Start package from release generation; its historical manifest version remains frozen. Renovate also ignores the inactive runtime and Docker workflow files, in addition to the paused source/Dockerfiles. Policy tests guard these exclusions. Do not reactivate them indirectly during ordinary active-package release maintenance. The only Docker CI exception is the trusted-main, GitHub-hosted Linux e2e-smoke-report-only.yaml Web browser workflow (authorized 2026-10-28). Its scoped launcher enforces admission, per-run identities and owned cleanup; the general paused commands and existing disabled jobs remain paused, including Rust, TanStack and the cron runner. See the GitHub Actions runbook’s Report-only Web E2E section for runtime bounds, host-lock recovery and report-only acceptance.

Public roadmap activity

The roadmap Worker uses a Cloudflare Cron Trigger at minute 17 each hour. Each event reserves one hour slot in a singleton SQLite-backed Durable Object before querying GitHub. A completed reservation publishes one snapshot atomically. An interrupted reservation still uses the slot. The public EN and VI pages read cached results and fall back to their build snapshot when the scheduled observation is missing or older than the build. The commit count covers commits reachable from the observed main SHA over thirty Vietnam calendar days, including a separate seven-day total. It excludes bot and release noise and groups commits by conventional scope. The existing build-time deployment and PR history retain their respective sources. The scheduled read makes at most 49 GitHub requests and one public Web build-info request per hour, with an eight-second request timeout and no retries. A response beyond the page cap leaves the prior snapshot intact. Each scheduled GitHub call emits one roadmap.github.rate_limit console event for Worker tail capture. The event records the expected and observed resource bucket, HTTP status, numeric limit, remaining count and reset epoch, and the Worker observation time. Missing or invalid fields are null. The event omits request URLs, bodies, credentials and unrelated headers. Prepare the switch-off recovery artifact from the hosted, sealed primary artifact on a qualified Linux host. Run node apps/roadmap/scripts/prepare-recovery-artifact.mjs <primary-dir> <exact-sha> <output-parent>. The script verifies the descriptor seal before copying its exact bytes to a fresh directory. It reseals the emitted files with empty cron triggers and ROADMAP_REFRESH_ENABLED=false; the Worker modules, assets, binding and SQLite migration remain identical. With the production switch set to true, the scheduled handler can reserve refresh work and route loaders may read the durable snapshot. With it set to false, the page serves build fallback data. Use the switch-off artifact after the class migration if the scheduled path needs to be disabled. Qualify the sealed artifact and its emitted configuration on an isolated Worker before production release. The roadmap observation carries a distinct count of commits reachable from main for the current and prior seven Vietnam calendar days. Product counts exclude bot and release noise. Each product row includes a 30-day series (oldest first), both seven-day totals, a thirty-day total, merged PRs in the current seven days, and up to five recent subjects that pass the public-title filter. The Mind product key is mind. Unmapped work rolls up to other. A recent commit is marked live only when the fetched parent links prove ancestry from the verified production SHA. Missing ancestry is null. The release train groups the build artifact’s public app observations and the scheduled Web build observation by SHA. Each production marker uses the earliest observed build time and the distinct app count. Across refreshes it retains the highest observed app count for each SHA. This source currently observes production builds. When the Web endpoint cannot verify the active SHA, the response records a null production SHA and retains earlier markers. To stop the job, remove the cron trigger and set ROADMAP_REFRESH_ENABLED to false in the Worker configuration. This also stops snapshot reads and returns the page to its build snapshot. Allow any already-admitted execution to drain after the change takes effect. After three consecutive failed or interrupted slots, the object stops admitting outbound work until an operator diagnoses the cause and deploys a new singleton identity for both readers and writers. The previous snapshot retains its observation time until the next successful publish or a newer build. Run node --test apps/roadmap/src/activity-fence.test.ts apps/roadmap/src/roadmap-snapshot.test.ts for the admission and snapshot rules. In the Worker acceptance run, exercise duplicate scheduled delivery, restart with an outstanding reservation, three failures, a disabled cron, and a successful observation. Verify the emitted configuration carries the cron, binding and SQLite migration, and both locales render the same cached timestamp. The build artifact workflow produces an artifact. Production deployment remains a separate release operation. The public EN and VI routes lead with the Pulse view. Its hero uses distinct main commits for both seven-day totals and the 30-day chart. The product cards use scoped counts and derive roadmap progress from the curated tracks. Commit subjects remain in English in both locales. The release train hides environments without observations, and narrow screens use a vertical rail and the last 14 days of the heatmap. Until a scheduled snapshot is available, the page labels its build-time activity and leaves the active production SHA unavailable. The detailed roadmap tabs remain available below Pulse. The Pulse layout renders store listings and WebP galleries from apps/mobile/store/metadata and apps/mobile/store/screenshots. Public submission and approval dates come from the listing metadata. Listing and screenshot updates ship with a new sealed artifact. The cron refreshes GitHub activity independently of those assets. The EN and VI Pulse views include a Tech stack inventory of the checked-in apps/ directories. The build classifies the source stack using package manifests, provider configuration, gateway mounts and publication holds. Progress counts web migration entries with both a TanStack Start Worker and a configured production route. Route-free candidates, including admitted uploads, stay outside the numerator. Paused and non-web app rows stay outside the denominator. This source inventory updates when a new Roadmap artifact is built.