Ekklesia/apps/api/src/server.ts

import type { Socket } from "node:net";
import path from "node:path";
import { pathToFileURL } from "node:url";
import { EMPTY_MODULE_SECRET_SOURCE, type ModuleSecretSource } from "@ekklesia/contracts";
import {
  createAuthorizationEvaluator,
  createAuthorizationStore,
} from "@ekklesia/core/authorization";
import { createDashboardStore } from "@ekklesia/core/dashboard";
import { createDeviceStore } from "@ekklesia/core/devices";
import { type HostEnvironment, loadHostConfiguration } from "@ekklesia/core/host";
import {
  createIdentityNotifications,
  createIdentityRuntime,
  type IdentityNotificationReporter,
  type IdentityRateLimitOverride,
} from "@ekklesia/core/identity";
import { createSmtpMailProvider, type MailProvider } from "@ekklesia/core/mail";
import {
  type AnyModuleRuntimeServices,
  createModuleConfigurationStore,
  createModuleLifecycleStore,
  createModuleUiPublicationRegistry,
  type DiscoveredModule,
  type ModuleConfigurationSettings,
  type ModuleEligibilityResolution,
  type ModuleRuntimeServicesLookup,
  resolveModuleEligibility,
  type StartedModuleRuntime,
  startDiscoveredModuleRuntime,
} from "@ekklesia/core/modules";
import {
  createModuleStorageLookup,
  createS3StorageProvider,
  type StorageProvider,
} from "@ekklesia/core/storage";
import { type ServerType, serve } from "@hono/node-server";
import type { Hono } from "hono";
import { createApiApp } from "./app.js";
import {
  type ClientAddressResolver,
  createClientAddressResolver,
  wrapFetchWithClientAddress,
} from "./client-address.js";
import { type ApiRuntimeConfig, apiTrustedOrigins, createApiRuntimeConfig } from "./config.js";
import { loadDashboardHostAssets } from "./dashboard-host-assets.js";
import { type ApiDatabaseHandle, createApiDatabase } from "./database.js";
import { withDevicePrincipals } from "./device-principal.js";
import { loadDisplayHostAssets } from "./display-host-assets.js";
import { READINESS_DATABASE_TIMEOUT_MS } from "./health-routes.js";
import { prepareStartupModuleInventory } from "./platform-permissions.js";

const connectionTimeoutMs = 10_000;
const operationTimeoutMs = 10_000;

export interface NodeServerTimeouts {
  readonly headersTimeoutMs: number;
  readonly requestTimeoutMs: number;
  readonly keepAliveTimeoutMs: number;
  /**
   * How often Node re-scans open connections for a header/request timeout violation
   * (`http.Server`'s `connectionsCheckingInterval`). This is not merely a performance knob: Node
   * enforces `headersTimeout`/`requestTimeout` only when this periodic check runs, and its own
   * default (30s) is verified (`node --version` 24.15.0, direct `http.createServer` reproduction)
   * to let a violating connection sit for up to a full additional interval past its actual
   * timeout before Node ever closes it — so leaving this at Node's default would silently widen
   * both timeouts above by up to 30s in the worst case. Kept small relative to `headersTimeoutMs`
   * so a violation is detected promptly after it actually elapses.
   */
  readonly connectionsCheckingIntervalMs: number;
}

/**
 * Finite Node HTTP server timeouts for this JSON-only API. Node's own defaults
 * (`headersTimeout`: 60s, `requestTimeout`: 300s) are generous enough to let a slow-header or
 * slow-body connection hold a socket for minutes with no reverse proxy in front; these values bound
 * that exposure while remaining comfortable for the small JSON payloads this API ever accepts (see
 * `./body-limits.js`). `headersTimeoutMs` must not exceed `requestTimeoutMs` — Node validates this
 * relationship itself at server construction — so headers alone may take up to 20s, and the
 * complete request (headers plus body) up to 30s, before Node itself responds `408` and closes the
 * connection without ever invoking a handler (subject to `connectionsCheckingIntervalMs`'s own
 * detection latency, above). `keepAliveTimeoutMs` matches Node's own default (5s) made explicit and
 * typed rather than left implicit; it bounds only the idle window between requests on a reused
 * connection, never an in-flight one.
 */
const PRODUCTION_NODE_SERVER_TIMEOUTS: NodeServerTimeouts = Object.freeze({
  connectionsCheckingIntervalMs: 5_000,
  headersTimeoutMs: 20_000,
  keepAliveTimeoutMs: 5_000,
  requestTimeoutMs: 30_000,
});

/**
 * The two fixed operational lines a transport-level client failure may produce. Neither ever
 * carries the remote address, request path, query, headers, cookies, body, forwarding chain, the
 * caught error's own message, or the malformed bytes that caused it — an attacker controls all of
 * those, and a parse failure's message in particular is derived directly from the rejected input.
 */
const NODE_SERVER_TIMEOUT_LOG_MESSAGE =
  "[server] A connection was closed for exceeding a configured header/request timeout.";
const NODE_SERVER_CLIENT_ERROR_LOG_MESSAGE =
  "[server] A connection was closed after a malformed or oversized client request.";

/**
 * Node's own documented default `clientError` responses, reproduced byte-for-byte (verified
 * directly against the pinned Node 24 runtime by driving each case through a real socket with no
 * listener attached). Registering *any* `clientError` listener suppresses that default entirely —
 * `_http_server.js` only falls back when `server.emit("clientError", ...)` returns `false` — so a
 * listener that merely observes would silently turn every 408/400/431 into a bare connection reset.
 * Reproducing the exact bytes is what lets the platform log the event without changing what a
 * caller actually receives.
 */
const NODE_CLIENT_ERROR_RESPONSES: Readonly<Record<string, string>> = Object.freeze({
  ERR_HTTP_REQUEST_TIMEOUT: "HTTP/1.1 408 Request Timeout\r\nConnection: close\r\n\r\n",
  HPE_HEADER_OVERFLOW: "HTTP/1.1 431 Request Header Fields Too Large\r\nConnection: close\r\n\r\n",
});
const NODE_CLIENT_ERROR_DEFAULT_RESPONSE = "HTTP/1.1 400 Bad Request\r\nConnection: close\r\n\r\n";

/**
 * Mirrors Node's own `socketOnError` fallback exactly: write the code-appropriate response only
 * when the socket is still writable *and* no response has already begun for the in-flight request,
 * then destroy the socket. The `_httpMessage` guard is what prevents a second, corrupting status
 * line when a handler already responded and only the (unread) request body subsequently timed out.
 */
function writeNodeClientErrorResponse(error: NodeJS.ErrnoException, socket: Socket): void {
  const inFlightResponse = (socket as Socket & { _httpMessage?: { headersSent?: boolean } })
    ._httpMessage;
  const alreadyResponding = inFlightResponse?.headersSent === true;
  if (socket.writable && !alreadyResponding) {
    const code = typeof error.code === "string" ? error.code : "";
    socket.write(NODE_CLIENT_ERROR_RESPONSES[code] ?? NODE_CLIENT_ERROR_DEFAULT_RESPONSE);
  }
  socket.destroy();
}

export interface RunningApi {
  readonly app: Hono;
  readonly server: ServerType;
  shutdown(): Promise<void>;
}

export interface StartApiOptions {
  /** Test-only escape hatch: never set by the executable entry point. See
   * `IdentityRateLimitOverride` for why production must leave this unset. */
  readonly rateLimitOverride?: IdentityRateLimitOverride;
  /**
   * Where a verification or reset message that could not be delivered is reported.
   *
   * Production composition leaves this unset and one fixed sanitized line is written. It exists as a
   * seam because the *only* externally observable difference between "the relay is down" and "there
   * is no such account" must be this report — every HTTP response is deliberately identical — so a
   * test needs somewhere to observe the operational failure the user is never shown.
   */
  readonly reportNotificationFailure?: IdentityNotificationReporter;
  /** The deployment's raw `EKKLESIA_MODULES_DIR` (ADR 0048), a production input the executable
   * entry point always passes from host configuration. See `prepareStartupModuleInventory`. */
  readonly configuredModulesDirectory?: string | undefined;
  /** Test-only escape hatch: overrides the directory module discovery scans, taking precedence
   * over `configuredModulesDirectory`. Never set by the executable entry point. See
   * `prepareStartupModuleInventory`. */
  readonly modulesDirectoryOverride?: string;
  /** Test-only escape hatch: supplies an already-discovered module list directly, skipping
   * filesystem discovery entirely — this exact list drives both permission composition and module
   * runtime startup. Never set by the executable entry point. See
   * `prepareStartupModuleInventory`. */
  readonly discoveredModulesOverride?: readonly DiscoveredModule[];
  /** Test-only escape hatch: narrows the Node header/request/keep-alive timeouts so a transport
   * test can prove real timeout behavior deterministically without waiting out the production
   * values. Never set by the executable entry point. */
  readonly nodeServerTimeoutOverride?: Partial<NodeServerTimeouts>;
  /** Test-only escape hatch: narrows the graceful HTTP drain deadline
   * (`HTTP_DRAIN_TIMEOUT_MS`) so a forced close can be proven in milliseconds. Never set by the
   * executable entry point. */
  readonly drainTimeoutMsOverride?: number;
  /** Test-only escape hatch: narrows the per-module stop/rollback hook deadline
   * (`MODULE_STOP_HOOK_TIMEOUT_MS`). Never set by the executable entry point. */
  readonly moduleHookTimeoutMsOverride?: number;
  /**
   * Where deployment-supplied module secrets come from (ADR 0029) — the one *production* dependency
   * in this options object rather than a test-only escape hatch.
   *
   * It is a dependency rather than a field on `ApiRuntimeConfig` because it is a lookup, not a
   * validated value, and because the set of secrets depends on which modules are discovered — which
   * is not known when configuration is validated. The executable entry point supplies an
   * environment-backed source; omitting it supplies nothing, so a module declaring a required secret
   * is reported `secret_missing` rather than silently starting without it.
   */
  readonly moduleSecrets?: ModuleSecretSource;
}

async function safeCloseDatabase(handle: ApiDatabaseHandle): Promise<void> {
  try {
    await handle.close();
  } catch {
    // Cleanup failures are swallowed here deliberately: they must never mask a startup failure
    // that is already propagating, and there is no raw driver detail safe to surface.
  }
}

/** Same swallow-on-cleanup-path rule as `safeCloseDatabase`, for the mail provider's transport. */
async function safeCloseMail(provider: MailProvider | undefined): Promise<void> {
  if (!provider) {
    return;
  }
  try {
    await provider.close();
  } catch {
    // See safeCloseDatabase: never mask a failure that is already propagating.
  }
}

/** Same swallow-on-cleanup-path rule as `safeCloseDatabase`, for the storage provider's own
 * connection pool. */
async function safeCloseStorage(provider: StorageProvider | undefined): Promise<void> {
  if (!provider) {
    return;
  }
  try {
    await provider.close();
  } catch {
    // See safeCloseDatabase: never mask a failure that is already propagating.
  }
}

/** Same swallow-on-cleanup-path rule as `safeCloseDatabase`, used only while unwinding a startup
 * failure that already occurred after the module runtime started — never during normal shutdown,
 * where a module stop failure remains an observable result instead (see `startApi`'s `shutdown`). */
async function safeStopModuleRuntime(moduleRuntime: StartedModuleRuntime): Promise<void> {
  try {
    await moduleRuntime.stop();
  } catch {
    // See safeCloseDatabase: never mask the startup failure already propagating.
  }
}

/**
 * The platform's fixed graceful-drain deadline. In-flight HTTP work has this long to finish after
 * the listener stops accepting new connections; after it, connections are closed so process
 * shutdown stays bounded. Not configurable in this slice.
 */
export const HTTP_DRAIN_TIMEOUT_MS = 30_000;

/**
 * A shutdown that did not complete cleanly. Carries only fixed booleans and the already-sanitized
 * module-runtime error — never a request, socket, URL, credential, SQL, or stack detail — so an
 * operator can see *which* stages failed without any of it becoming a disclosure channel.
 */
export class ApiShutdownError extends Error {
  readonly drainTimedOut: boolean;
  readonly moduleStopFailed: boolean;
  readonly moduleStopError: unknown;

  constructor(drainTimedOut: boolean, moduleStopError: unknown) {
    super("The API did not shut down cleanly.");
    this.name = "ApiShutdownError";
    this.drainTimedOut = drainTimedOut;
    this.moduleStopFailed = moduleStopError !== undefined;
    this.moduleStopError = moduleStopError;
    this.stack = `${this.name}: ${this.message}`;
    Object.freeze(this);
  }
}

/**
 * Stops accepting new connections, then waits up to `timeoutMs` for existing HTTP work to finish.
 *
 * At the deadline it closes idle and then still-active connections using Node's own supported
 * public `Server` APIs (`closeIdleConnections`/`closeAllConnections`, both called only after
 * `server.close()`), and stops waiting for the close callback. No private socket or server field is
 * touched.
 *
 * Force-closing a connection ends the transport. It does **not** prove that arbitrary in-process
 * handler JavaScript was canceled — a handler promise may still be running and may still settle.
 * What is guaranteed is that shutdown no longer waits for it, and that the pool closes afterwards,
 * so its late database work fails.
 */
/** `ServerType` is a union that also covers HTTP/2 servers, which do not expose these two methods.
 * `serve()` here always creates a plain `http.Server`, which does; narrowing with a real capability
 * check keeps that honest instead of asserting a cast. */
interface ConnectionClosingServer {
  closeAllConnections(): void;
  closeIdleConnections(): void;
}

function supportsConnectionClosing(
  server: ServerType,
): server is ServerType & ConnectionClosingServer {
  const candidate = server as Partial<ConnectionClosingServer>;
  return (
    typeof candidate.closeIdleConnections === "function" &&
    typeof candidate.closeAllConnections === "function"
  );
}

async function drainHttpServer(
  server: ServerType,
  timeoutMs: number,
): Promise<"drained" | "forced"> {
  let timer: ReturnType<typeof setTimeout> | undefined;
  const closed = new Promise<void>((resolve) => {
    server.close(() => resolve());
  });
  try {
    const outcome = await Promise.race([
      closed.then(() => "drained" as const),
      new Promise<"forced">((resolve) => {
        timer = setTimeout(() => resolve("forced"), timeoutMs);
      }),
    ]);
    if (outcome === "drained") {
      return "drained";
    }
    if (supportsConnectionClosing(server)) {
      server.closeIdleConnections();
      server.closeAllConnections();
    }
    return "forced";
  } finally {
    clearTimeout(timer);
    // The close callback may still fire later; nothing awaits it, and it carries no rejection.
  }
}

/**
 * Binds the Node HTTP server. Two platform-owned transport concerns are applied at this exact
 * boundary, below Hono entirely: `wrapFetchWithClientAddress` resolves and sanitizes the caller's
 * address from the real socket peer before any route (including Better Auth's own rate limiter)
 * ever sees the request (see `./client-address.js`), and `serverOptions` gives Node finite
 * header/request/keep-alive timeouts so a slow-header or slow-body connection cannot hold a socket
 * indefinitely. A connection Node itself closes for exceeding one of these timeouts never reaches
 * `app.fetch` and never receives a second response — both are Node's own documented behavior for
 * `headersTimeout`/`requestTimeout`, not something this function re-implements.
 */
function bindServer(
  app: Hono,
  host: string,
  port: number,
  clientAddressResolver: ClientAddressResolver,
  timeouts: NodeServerTimeouts,
): Promise<ServerType> {
  return new Promise((resolve, reject) => {
    let settled = false;
    const server = serve(
      {
        fetch: wrapFetchWithClientAddress(app.fetch, clientAddressResolver),
        hostname: host,
        port,
        serverOptions: {
          connectionsCheckingInterval: timeouts.connectionsCheckingIntervalMs,
          headersTimeout: timeouts.headersTimeoutMs,
          keepAliveTimeout: timeouts.keepAliveTimeoutMs,
          requestTimeout: timeouts.requestTimeoutMs,
        },
      },
      () => {
        if (!settled) {
          settled = true;
          resolve(server);
        }
      },
    );
    // `headersTimeout`/`requestTimeout` do NOT surface as the server's general `'timeout'` event —
    // that event only fires for `server.setTimeout()`'s idle-socket timer, which this server never
    // arms. Verified directly against the pinned Node 24 runtime by driving a real slow-header and
    // slow-body connection through a real socket: the only server-level event either produces is
    // `'clientError'`, carrying `err.code === "ERR_HTTP_REQUEST_TIMEOUT"`. Observing the failure
    // therefore has to happen here, and — because registering a listener suppresses Node's own
    // default response — this handler must reproduce that default itself.
    server.on("clientError", (error: NodeJS.ErrnoException, socket: Socket) => {
      console.error(
        error.code === "ERR_HTTP_REQUEST_TIMEOUT"
          ? NODE_SERVER_TIMEOUT_LOG_MESSAGE
          : NODE_SERVER_CLIENT_ERROR_LOG_MESSAGE,
      );
      writeNodeClientErrorResponse(error, socket);
    });
    server.once("error", () => {
      if (!settled) {
        settled = true;
        reject(new Error("The API server failed to start listening."));
      }
    });
  });
}

/**
 * Explicit runtime composition, in a fixed order:
 *
 * 1. Discover every installed module's manifest exactly once and compose the one immutable
 *    platform permission registry from Core's own declarations plus every module's
 *    (`prepareStartupModuleInventory`, `./platform-permissions.js`).
 * 2. Construct the one shared PostgreSQL pool (identity, authorization, and the module database
 *    facade each get a typed Drizzle view over it — see `./database.js`).
 * 3. Construct the Core identity runtime, authorization store, and evaluator from that pool.
 * 4. Read the stored module enablement configuration and resolve, from that exact same discovered
 *    inventory, which modules are eligible to run (`resolveModuleEligibility`).
 * 5. Start the module runtime from the eligible subset (`startDiscoveredModuleRuntime`) — never
 *    rediscovering modules a second time — passing the manifest-selected service map.
 * 6. Create the Hono application, mounting the started module runtime's published routes, publishing
 *    UI for that same eligible set, and passing the same per-module service map to HTTP handlers.
 * 7. Bind the Node HTTP server.
 *
 * Discovery and permission composition happen before any pool, Better Auth instance, module
 * entrypoint import, or HTTP server exists, so a discovery or composition failure — an invalid
 * manifest, a dependency-graph problem, a module claiming the reserved `core.*` permission
 * namespace, or a duplicate permission declaration — leaves nothing to close: it propagates
 * directly, and startup never falls back to a Core-only registry. Core infrastructure (the pool,
 * module database facade, identity, authorization) exists before any executable module lifecycle
 * code runs, and every module route is published only after the entire module runtime has
 * registered and started successfully, so no request can ever reach a module before every module
 * has started. A module that is disabled, or blocked by a disabled required dependency, is never
 * imported at all: no entrypoint evaluation, no lifecycle hook, no route, no capability, and no
 * published UI artifact — there is no per-request "is this module enabled?" check anywhere, because
 * there is nothing mounted to check.
 *
 * Step 4 makes startup depend on a reachable, migrated database. If the enablement configuration
 * cannot be read, startup fails closed with Core's sanitized `MODULE_LIFECYCLE_READ_FAILED` rather
 * than starting every discovered module: the platform cannot know what an administrator decided, and
 * starting a module they disabled is worse than not starting at all.
 *
 * Returns a handle with an idempotent `shutdown()` so both the executable entry point and tests can
 * stop the server, the module runtime, and the shared pool deterministically. Never applies
 * migrations, generates schema, or issues a module database query merely because modules were
 * discovered — the database must already be migrated through `pnpm migrations:apply`.
 */
export async function startApi(
  config: ApiRuntimeConfig,
  options: StartApiOptions = {},
): Promise<RunningApi> {
  const inventory = await prepareStartupModuleInventory({
    configuredModulesDirectory: options.configuredModulesDirectory,
    ...(options.modulesDirectoryOverride === undefined
      ? {}
      : { modulesDirectoryOverride: options.modulesDirectoryOverride }),
    ...(options.discoveredModulesOverride === undefined
      ? {}
      : { discoveredModulesOverride: options.discoveredModulesOverride }),
  });

  const databaseHandle = createApiDatabase(config.databaseUrl, {
    connectionTimeoutMs,
    operationTimeoutMs,
    // The one validated transport policy reaches this runtime's single physical pool here; the
    // identity, authorization, and module Drizzle views are all typed views over that same pool,
    // so TLS never introduces a second connection.
    transport: config.databaseTransport,
  });

  // Both host providers are constructed from validated configuration **before** anything that uses
  // them, and neither touches the network while being built: an S3 client opens no connection and a
  // Nodemailer transport opens no socket, so an unreachable relay or storage service can never stop
  // the platform from booting (ADR 0030, ADR 0037). An installation with neither configured gets
  // `undefined` for both, which is a complete, valid deployment.
  const storageProvider = config.storage ? createS3StorageProvider(config.storage) : undefined;
  const mailProvider = config.mail ? createSmtpMailProvider(config.mail) : undefined;

  let runtime: ReturnType<typeof createIdentityRuntime>;
  let authorizationStore: ReturnType<typeof createAuthorizationStore>;
  let deviceStore: ReturnType<typeof createDeviceStore>;
  try {
    runtime = createIdentityRuntime({
      config: config.identity,
      db: databaseHandle.identity,
      // Identity's two outbound messages, composed here — the one place that knows both the
      // validated identity configuration and the host's mail provider. Core's identity code learns
      // a recipient, a subject, and two bodies and nothing whatsoever about SMTP (ADR 0038).
      notifications: createIdentityNotifications({
        authBaseUrl: config.identity.baseUrl,
        dashboardOrigin: config.identity.dashboardOrigin,
        mail: mailProvider,
        ...(options.reportNotificationFailure ? { report: options.reportNotificationFailure } : {}),
      }),
      ...(options.rateLimitOverride ? { rateLimitOverride: options.rateLimitOverride } : {}),
    });
    authorizationStore = createAuthorizationStore(databaseHandle.authorization);
    deviceStore = createDeviceStore(databaseHandle.devices);
  } catch (error) {
    await safeCloseStorage(storageProvider);
    await safeCloseMail(mailProvider);
    await safeCloseDatabase(databaseHandle);
    throw error;
  }

  const evaluator = createAuthorizationEvaluator({
    // Device authority reaches the one evaluator here. There is no second decision point: a device
    // is allowed or denied by the same function, with the same deny-by-default, as a person.
    devices: deviceStore,
    grants: authorizationStore,
    platformOwners: authorizationStore,
  });

  // Read the administrator's enablement decisions and resolve which discovered modules may actually
  // start, before a single module entrypoint is imported. A failure here fails startup rather than
  // falling back to "start everything": the platform cannot know what it was told, and starting a
  // module an administrator disabled is worse than not starting at all.
  const moduleLifecycleStore = createModuleLifecycleStore(databaseHandle.moduleLifecycle);
  const moduleConfigurationStore = createModuleConfigurationStore(
    databaseHandle.moduleConfiguration,
  );
  const moduleSecrets = options.moduleSecrets ?? EMPTY_MODULE_SECRET_SOURCE;
  let eligibility: ModuleEligibilityResolution;
  let moduleConfigurationSettings: ModuleConfigurationSettings;
  try {
    // Both axes are read before a single module entrypoint is imported, and resolved together in
    // one pass (ADR 0029): a module is eligible only if an administrator enabled it, its own
    // required configuration is satisfiable, and every module it requires is eligible too. The pass
    // also produces each eligible module's frozen configuration snapshot, so the configuration that
    // made a module eligible is the identical object it later receives.
    // Read sequentially rather than concurrently, deliberately: these are two tiny queries, and
    // racing them would make the reported failure depend on which rejected first. An operator
    // debugging a broken deployment must get the same error code every time.
    const lifecycleConfiguration = await moduleLifecycleStore.readConfiguration();
    const settings = await moduleConfigurationStore.readSettings();
    moduleConfigurationSettings = settings;
    eligibility = resolveModuleEligibility(
      inventory.modules,
      lifecycleConfiguration,
      { secrets: moduleSecrets, settings },
      { providerConfigured: storageProvider !== undefined },
    );
  } catch (error) {
    await safeCloseStorage(storageProvider);
    await safeCloseMail(mailProvider);
    await safeCloseDatabase(databaseHandle);
    throw error;
  }

  const services: ModuleRuntimeServicesLookup = new Map<string, AnyModuleRuntimeServices>(
    eligibility.eligibleModules.map((module) => [
      module.manifest.id,
      module.manifest.database?.abi === 1
        ? Object.freeze({ database: databaseHandle.moduleExecutor(module.manifest.id) })
        : Object.freeze({ database: databaseHandle.module }),
    ]),
  );

  // One scoped facade per eligible module, over the exact buckets that module's own manifest
  // declares. Built from the eligible set so a blocked module never receives storage at all.
  const moduleStorages = createModuleStorageLookup(
    storageProvider,
    config.storage?.bucketPrefix ?? "",
    eligibility.eligibleModules.map((module) => ({
      logicalBucketIds: module.manifest.storage.buckets.map((bucket) => bucket.id),
      moduleId: module.manifest.id,
    })),
  );

  // This snapshot is inert metadata only: construction reads no artifact and executes no module.
  // It lets the API prove that a display host build is available *before* importing any eligible
  // module code. The snapshot is not exposed anywhere unless the complete runtime subsequently
  // starts and application composition succeeds.
  const moduleUi = createModuleUiPublicationRegistry(eligibility.eligibleModules);
  let displayHostAssets: Awaited<ReturnType<typeof loadDisplayHostAssets>> | undefined;
  let dashboardHostAssets: Awaited<ReturnType<typeof loadDashboardHostAssets>>;
  try {
    displayHostAssets =
      moduleUi.listDisplays().length === 0 ? undefined : await loadDisplayHostAssets();
    // Unconditional, unlike the display host: every installation serves its Dashboard (ADR 0044),
    // so a missing shell build fails here, before any module code is imported.
    // The page learns its API origin from the API that serves it, so one build serves either
    // topology (ADR 0045).
    dashboardHostAssets = await loadDashboardHostAssets(config.identity.baseUrl);
  } catch (error) {
    await safeCloseStorage(storageProvider);
    await safeCloseMail(mailProvider);
    await safeCloseDatabase(databaseHandle);
    throw error;
  }

  // Module load/register/start failures use Core's own existing rollback (already-started
  // modules are stopped, in reverse order, before this rejects) — there is no separate runtime
  // handle to stop here, only the pool this failure leaves open.
  //
  // Only eligible modules are started, and startup remains all-or-nothing across *that* set: an
  // ineligible module is never imported, so its `register`/`start` hooks never run, it publishes no
  // routes and no capabilities, and it cannot take startup down. Eligibility filtering preserves
  // discovery's topological order, so a started module's required dependencies still precede it.
  let moduleRuntime: StartedModuleRuntime;
  try {
    moduleRuntime = await startDiscoveredModuleRuntime(eligibility.eligibleModules, services, {
      configurations: eligibility.configurations,
      storages: moduleStorages,
      ...(options.moduleHookTimeoutMsOverride === undefined
        ? {}
        : { hookTimeoutMs: options.moduleHookTimeoutMsOverride }),
    });
  } catch (error) {
    await safeCloseStorage(storageProvider);
    await safeCloseMail(mailProvider);
    await safeCloseDatabase(databaseHandle);
    throw error;
  }

  // Set by the first `shutdown()` call, before anything is drained or stopped, so readiness turns
  // `503` while the instance still finishes in-flight requests.
  let shuttingDown = false;

  let app: Hono;
  try {
    app = createApiApp({
      authorization: {
        evaluator,
        registry: inventory.permissionRegistry,
        store: authorizationStore,
      },
      moduleRoutes: moduleRuntime.routes,
      moduleRealtime: moduleRuntime.realtime,
      moduleServices: services,
      // Built from the exact **eligible** set the runtime just started, never the raw discovered
      // inventory, and exposed only after startup succeeded — so the published set is the running
      // set. A disabled or dependency-blocked module therefore publishes no artifact, appears in no
      // browser inventory, and appears in no resolved Dashboard navigation, by construction rather
      // than by a filter somewhere downstream. Modules declaring no UI role are simply absent.
      moduleUi,
      ...(displayHostAssets === undefined ? {} : { displayHostAssets }),
      moduleLifecycle: {
        // The boot snapshot: what this process discovered, what it resolved, and what the runtime
        // reports it actually started — read from the runtime rather than assumed equal to the
        // eligible set.
        snapshot: Object.freeze({
          discoveredModules: inventory.modules,
          eligibility: eligibility.entries,
          startedModuleIds: Object.freeze(moduleRuntime.modules.map((module) => module.id)),
        }),
        store: moduleLifecycleStore,
      },
      moduleConfiguration: {
        // The settings this process actually resolved from, so `restartRequired` is a comparison
        // against what is running rather than a guess.
        bootSettings: moduleConfigurationSettings,
        discoveredModules: inventory.modules,
        secrets: moduleSecrets,
        snapshots: eligibility.configurations,
        store: moduleConfigurationStore,
      },
      moduleStorage: {
        providerConfigured: storageProvider !== undefined,
        snapshots: moduleStorages,
      },
      // A typed view over the same single pool as identity, authorization, and the module facade.
      dashboard: createDashboardStore(databaseHandle.dashboard),
      dashboardHostAssets,
      // The app is only composed after `startDiscoveredModuleRuntime` resolved, so module startup
      // has completed whenever it can answer; it stops counting as complete once shutdown begins.
      health: {
        checkDatabase: () => databaseHandle.ping(READINESS_DATABASE_TIMEOUT_MS),
        isModuleStartupComplete: () => !shuttingDown,
      },
      devices: {
        // `Secure` follows the deployment's own scheme, read from the already-validated identity
        // base URL rather than from any request. Production is https by configuration policy, so a
        // real deployment always gets a Secure device cookie; a plaintext development origin would
        // silently drop one.
        secureCookies: config.identity.baseUrl.startsWith("https://"),
        store: deviceStore,
      },
      // Principal resolution, extended with managed devices. A human session still wins, and this
      // is the only place a device credential ever becomes a principal.
      runtime: withDevicePrincipals({ devices: deviceStore, runtime }),
      // The configured origins plus the API's own (ADR 0045).
      trustedOrigins: apiTrustedOrigins(config.identity),
    });
  } catch (error) {
    await safeStopModuleRuntime(moduleRuntime);
    await safeCloseStorage(storageProvider);
    await safeCloseMail(mailProvider);
    await safeCloseDatabase(databaseHandle);
    throw error;
  }

  const clientAddressResolver = createClientAddressResolver(config.clientAddress);
  const timeouts: NodeServerTimeouts = {
    ...PRODUCTION_NODE_SERVER_TIMEOUTS,
    ...options.nodeServerTimeoutOverride,
  };

  let server: ServerType;
  try {
    server = await bindServer(app, config.host, config.port, clientAddressResolver, timeouts);
  } catch (error) {
    await safeStopModuleRuntime(moduleRuntime);
    await safeCloseStorage(storageProvider);
    await safeCloseMail(mailProvider);
    await safeCloseDatabase(databaseHandle);
    throw error;
  }

  // Memoizes one shared promise rather than a boolean flag: a boolean only prevents a *second*
  // cleanup from starting, but a caller that arrives while the first cleanup is still in flight
  // would previously get an already-resolved `undefined` immediately, before the listener had
  // actually drained or the pool had actually closed. Storing the promise itself means every
  // caller — concurrent or sequential — awaits the exact same drain-then-close operation and
  // observes it complete (or fail) exactly once. Realtime subscriptions are closed *before* HTTP
  // drain begins: they are intentionally long-lived HTTP responses and would otherwise keep every
  // clean shutdown open until the forced-drain deadline. Order matters, and is now additionally
  // load-bearing for the module database facade: the HTTP listener is drained and closed first so
  // no new or in-flight request can reach a module handler while the module runtime is stopping;
  // then the module runtime stops — its stop hooks may still use `services.database`, which is
  // safe because the pool is not yet closed; only then does the shared pool close, so no module
  // code can run against it afterward. Unlike the startup-cleanup paths above, a module-stop
  // failure here is not swallowed — it is a genuine shutdown result a caller should observe — but
  // database cleanup still always runs first, so a stop failure never leaves the pool open.
  const drainTimeoutMs = options.drainTimeoutMsOverride ?? HTTP_DRAIN_TIMEOUT_MS;
  let shutdownPromise: Promise<void> | undefined;
  const shutdown = (): Promise<void> => {
    if (!shutdownPromise) {
      shuttingDown = true;
      shutdownPromise = (async () => {
        // Close long-lived streams so HTTP can drain, but leave publishers active until the drain
        // completes: an in-flight module command may still commit and publish its resulting state.
        // `moduleRuntime.stop()` deactivates publishers immediately before stop hooks run.
        moduleRuntime.realtime.closeSubscriptions();
        const drainOutcome = await drainHttpServer(server, drainTimeoutMs);
        let moduleStopError: unknown;
        try {
          await moduleRuntime.stop();
        } catch (error) {
          moduleStopError = error;
        }
        // After the module runtime has stopped, so a `stop` hook may still use its storage
        // facade, and before the pool closes, matching the existing ordering discipline exactly.
        await safeCloseStorage(storageProvider);
        await safeCloseMail(mailProvider);
        await safeCloseDatabase(databaseHandle);
        // Every stage above ran regardless of what the earlier ones did, so a drain timeout or a
        // module-stop failure can never leave the pool open. Both failures are reported together,
        // as fixed metadata: a forced drain is a genuine shutdown failure, not a success.
        if (drainOutcome === "forced" || moduleStopError) {
          throw new ApiShutdownError(
            drainOutcome === "forced",
            moduleStopError === undefined ? undefined : moduleStopError,
          );
        }
      })();
    }
    return shutdownPromise;
  };

  return Object.freeze({ app, server, shutdown });
}

/**
 * Wires `SIGINT`/`SIGTERM` to `running.shutdown()` and only then exits — extracted from `main()`
 * so it can be exercised directly in tests with a fake `running`/`exit`, without sending a real
 * process signal or invoking the real `process.exit` (which would terminate the test runner
 * itself). Production composition (`main()`, below) calls this with real `process.exit`.
 */
export function installShutdownSignalHandlers(
  running: Pick<RunningApi, "shutdown">,
  exit: (code: number) => void = (code) => process.exit(code),
): void {
  let signalReceived = false;
  const handleSignal = (): void => {
    if (signalReceived) {
      return;
    }
    signalReceived = true;
    running
      .shutdown()
      .then(() => exit(0))
      .catch(() => exit(1));
  };
  process.on("SIGINT", handleSignal);
  process.on("SIGTERM", handleSignal);
}

/**
 * Turns one deployment environment into this process's startup inputs.
 *
 * Every host infrastructure value comes from `loadHostConfiguration` — the same function
 * `pnpm doctor` calls with the same environment — so the runtime and the diagnostic that reports on
 * it can never disagree about what this deployment is configured to use. Only the two values that
 * describe *this listener* rather than the host's infrastructure are read here.
 */
export function readConfigFromEnvironment(processEnvironment: HostEnvironment): {
  readonly input: Parameters<typeof createApiRuntimeConfig>[0];
  readonly environment: "development" | "production";
  readonly moduleSecrets: ModuleSecretSource;
  readonly modulesDirectory: string | undefined;
} {
  const host = loadHostConfiguration(processEnvironment);
  return {
    environment: host.environment,
    input: {
      clientAddress: host.clientAddress,
      databaseTransport: host.databaseTransport,
      databaseUrl: host.databaseUrl,
      host: processEnvironment.EKKLESIA_API_HOST,
      identity: host.identity,
      mail: host.mail,
      port: processEnvironment.EKKLESIA_API_PORT,
      storage: host.storage,
    },
    moduleSecrets: host.moduleSecrets,
    modulesDirectory: host.modulesDirectory,
  };
}

export async function main(): Promise<void> {
  const { environment, input, moduleSecrets, modulesDirectory } = readConfigFromEnvironment(
    process.env,
  );
  const config = createApiRuntimeConfig(input, { environment });
  const running = await startApi(config, {
    configuredModulesDirectory: modulesDirectory,
    moduleSecrets,
  });

  const address = running.server.address();
  const boundPort = address && typeof address === "object" ? address.port : config.port;
  console.log(`Ekklesia API listening on ${config.host}:${boundPort}`);

  installShutdownSignalHandlers(running);
}

function isDirectExecution(): boolean {
  const entrypoint = process.argv[1];
  return (
    entrypoint !== undefined && import.meta.url === pathToFileURL(path.resolve(entrypoint)).href
  );
}

if (isDirectExecution()) {
  main().catch(() => {
    console.error("Ekklesia API failed to start.");
    process.exitCode = 1;
  });
}