Changelog

Breaking for imports of @template-cloudflare/api. Server code moved from routers/, lib/ and topic folders to packages/api/src/features/<feature>/<feature>.<role>.ts, with shared building blocks in packages/api/src/platform/. Public subpaths now follow ./<feature>/<role>, for example ./todo/contract, ./app-router, ./uploads/service and ./assistant/http. If your fork imports an earlier subpath, use the table in the api package reference. The four subpaths the mobile app imports (./routers/{todo,integrations,billing,notifications}.schema) remain as aliases, so the native app builds unchanged.

• Feature shape. Every feature has the same role files: a leaf contract, procedures, a service, errors, and optional jobs and HTTP edges. There is no barrel and no repository layer; row-level security already isolates Organizations. The @template-cloudflare/api/mcp/* subpaths are server-only in the client boundary. New tests and lint rules hold the shape: leaf modules import only leaves, features import each other through declared edges, there are no import cycles, type imports are explicit, and configuration is read through packages/env. Backend shape, ADR-0011.
• Tagged domain errors and Effect in procedures. In todo and documents, services fail with tagged errors, and one mapper per feature turns them into the same APP_ERRORS codes and messages clients saw before. Each procedure runs its service once through runProcedure. A client abort is honoured only before the program starts; once it runs, a client that goes away no longer rolls back the write, so the write and its Audit Event commit and the response goes unread. Other features keep their services until they are migrated. ADR-0010, error handling.
• Request ID and quieter logs. Every RPC response carries x-request-id, including the 500 when building the request context fails: Cloudflare's cf-ray when it is well formed, otherwise a UUID. The MCP context derives its own request ID for the logs of its document tools. The same ID is context.requestId on the server and is part of the error log, so a user can quote it. It correlates logs only: outside Cloudflare a client can supply the cf-ray, so nothing decides by it. Expected 4xx failures are logged at debug level instead of as errors. Runtime switches (DB_TENANT_PIPELINE, TRUSTED_PROXY_HEADER, TRUSTED_PROXIES, TENANCY_PROBE) are declared in the env schema. Their meaning is unchanged, but an invalid value is now reported instead of silently replaced by a default; check these variables in your Worker and Node deployments.
• Input validation answers HTTP 400. An invalid input now returns HTTP 400 with the same INPUT_VALIDATION code, message and field errors as before; it used to come back as HTTP 500, because oRPC's default status map did not know the code. The RPC and OpenAPI handlers use errorStatusMap from @template-cloudflare/errors/server, and APP_ERRORS.INPUT_VALIDATION.status is 400 instead of 422, which also changes the status in the generated OpenAPI document. The assistant's HTTP route maps statuses with the same shared map, so INPUT_VALIDATION is 400 there as well. Clients that branch on code see no change. Error handling.
• Org roles, ready when needed. orgProcedure reads the Member's Org role with its membership check, and requireOrgRole(...) restricts a procedure to some roles, failing closed. Roles are not hierarchical (list every role that may act), and using it outside orgProcedure is a type error. No procedure applies it yet. Backend conventions.
• Written rules with their checks. Backend conventions state each rule with its reason, example, anti-pattern and enforcement. The enforcement matrix shows which rules a check holds and which rely on review. A review checklist, a pull request template and a root CODING_STANDARDS.md index the same rules.
• Runbooks. Runbooks cover a database restore or host move, the pipelining kill switch, Hyperdrive caching, the RLS proof on a preview, and the emergency shutdown. The emergency-shutdown workflow does not block a custom domain; the runbook gives the extra step.
• Decisions show their status. Every ADR starts with its status and acceptance date, and the decision index lists them. The pagination and testing rules are written down.
• A skill for new features. The domain-feature skill walks from a glossary term to an RLS table, contract, tagged errors, service, procedures, isolation tests and docs.

Measurements (report): the server code grows by 6.9 KiB gzip, within the 10 KiB budget for the Effect pilots, and the client carries no Effect. The happy-path p50 of the pilot procedures is within the noise of an A/A run; the todo.byId error path costs +0.040 ms. Separately, the Worker as a whole grows by 70.2 KiB gzip, of which 63.4 KiB are the documentation pages, ADRs and changelog the site renders on the server: that is content and has nothing to do with Effect, and whether it counts against the Worker budget is still open.

Sources: platform move, feature folders, first of 18 moves, graph tests, direction rules, runProcedure, todo pilot, documents pilot, request ID, error log levels, runtime switches, trusted proxy, requireOrgRole, 400 for input validation, measurement.

Breaking for forks that changed apps/web/src. The folders components/, hooks/, lib/, utils/, functions/, middleware/, og/ and data/ are gone. Code now lives in apps/web/src/features/<feature>/<feature>.<role>.ts(x), with the same feature names as the backend where one exists (todo, documents, billing, …), and in apps/web/src/platform/<area>/ for what has no domain (the oRPC client, session, i18n, errors, SEO, OG). Routes stay where they are and get thinner. No URL, route or translation key changes, and every route answers as before. The only visible copy change is the web code sample on the landing page (and in llms-full.txt), which now names apps/web/src/features/landing/landing.hooks.ts. If your fork edited a moved file, re-apply the change at its new path; git log --follow finds every file.

• Fixed roles. A feature has *.queries.ts (query options and the loader's prefetch), *.hooks.ts (the logic), *.page.tsx, presentational components, pure model files and, for server-only code, *.server.ts(x). Components no longer import TanStack Query, the oRPC client or the auth client directly. Web conventions, ADR-0012.
• Server-only code is named .server. TanStack Start refuses *.server.* modules in the client build, so the name is checked by the framework. Client code reaches server code only through two isomorphic bridges (the session query and the docs page loader), and the web keeps zero createServerFn. Client/server boundary.
• Rules with checks. A boundary test (scripts/web-feature-boundaries.test.mjs) checks layers, cross-feature imports, the server suffix, barrels, .client. names, dynamic imports and stale vi.mock or worker URLs. The same test checks import cycles, and file size is a lint warning. The enforcement matrix lists W1–W13.
• Locales stay central. apps/web/src/locales/{en,de,fr}.ts keep every key; new keys use their feature's prefix.
• A skill section for new screens. The domain-feature skill gains a web section with templates for queries, hooks, the page and a thin route.

Known leftover: a comment in apps/mobile/src/lib/user-agent.ts still names the old web path; it changes with the next mobile change. Unused vendored components moved as they were; removing them is a separate decision.

Measurements (report): the client grows by 425 B gzip in total, and its entry chunk shrinks by 1,128 B because loaders import *.queries instead of the hooks. The server code grows by 850 B gzip, and no Effect module reaches the client. The docs pages added or edited with this change (ADR-0012, the web conventions, this entry and the updated guides) add about 23.6 KiB gzip to the server's docs chunk, so the server grows by about 24.5 KiB gzip in total; the docs do not reach the client.

Sources: platform moves, platform, second part, OG with server suffixes, feature folders, first of 23 moves, last feature move, boundary test, boundary test in the root tests, file size warning, bundle markers for web server modules, thin todos route, files gateway, root document, loaders out of routes, measurement.

The web assistant adds persisted Threads, streamed responses and explicit approval for mutating tools. Tool execution uses one server catalog shared with MCP, preserving current Session and Organization authorization. Provider configuration and the ai.assistant flag jointly control availability.

Agent discovery now exposes public instructions and Markdown pages alongside OAuth connection metadata. The supported transport keeps TanStack AI’s Promise/SSE lifecycle separate from the Effect job boundary.

• AI assistant and approvals
• MCP server and connecting an agent
• Agent-readable pages

Source: assistant integration