Files
PaperClipAI/server/src/app.ts
T
Devin FoleyandPaperclip 78f8c6c3d4 Recover managed bundled plugin workers on demand (#10429)
## Thinking Path

> - Paperclip is the open source app people use to manage AI agents for
work
> - Managed deployments can auto-provision bundled sandbox-provider
plugins so cloud or remote execution environments appear in the board UI
> - In a multi-service deployment, several server processes can share
one database and boot concurrently
> - A sibling process can create a bundled plugin row while the web
process sees it before it reaches `ready`
> - The web process correctly avoids clobbering the existing row, but
its startup `loadAll()` can miss the plugin and never start that worker
locally
> - The environments capabilities route then filters out the sandbox
provider because the plugin is ready in the database but not running in
the web process
> - This pull request adds a narrow managed-bundle recovery path that
lazily starts the missing worker when the capabilities route sees a
ready managed bundled plugin
> - The benefit is that the sandbox provider becomes visible after the
install finishes, without requiring a web-process restart

## Linked Issues or Issue Description

- No public GitHub issue found for this exact deployment race.
- Related broad plugin runtime context: Refs #432.

Bug description:

- What happened: in a managed multi-service deployment with shared
database state and bundled plugin auto-install enabled, the API-serving
process can skip a plugin row while it is still `installed`, run startup
plugin loading before that row becomes `ready`, and then permanently
omit the sandbox provider from environment capabilities.
- Expected behavior: once the managed bundled plugin row reaches
`ready`, the API-serving process should be able to start the plugin
worker and include its sandbox provider without a restart.
- Steps to reproduce: boot a web process and a sibling worker process
concurrently; have the sibling create the bundled plugin row and
transition it to `ready` after the web process has already skipped
auto-install and run `loadAll()`.
- Deployment mode: managed multi-service deployment with shared database
state and `plugins.autoInstall` configured.

## What Changed

- Added a managed bundled plugin worker recovery helper that
single-flights lazy `loadSingle()` starts and only allows configured
managed bundled plugin keys.
- Passed the managed recovery hook into the environments capabilities
route.
- Updated `listReadyPluginEnvironmentDrivers()` to attempt bounded
recovery for ready managed bundled plugins whose worker is missing in
the current process, and only for plugins that actually declare a
`sandbox_provider` environment driver.
- Made request-time recovery use `loadSingle(id, { markErrorOnFailure:
false })` so a local activation failure in one process never transitions
the shared plugin row to `error` (a sibling process may be running the
plugin successfully).
- When error writes are suppressed and activation fails after the worker
was spawned, the loader now tears down the partially-registered local
runtime (scheduler registration, event subscriptions, agent tools,
worker process) instead of leaving a half-activated worker lingering;
the teardown steps are factored out of `unloadSingle()` into a shared
helper.
- A failed recovery attempt now discards the crashed/stopped handle it
left registered in the worker manager (a worker that dies during
initialize is killed without a scheduled restart), so later capability
requests can retry recovery instead of being blocked by the
handle-presence gate until a process restart. Handles in
starting/running/backoff states are left to the worker manager's own
lifecycle; recovery only ever starts when no handle existed, so no
pre-existing worker can be affected.
- Added a regression test suite covering the installed-to-ready race,
allowlist behavior, the driver-kind gate, existing worker handles,
concurrent single-flight recovery, bounded slow recovery attempts,
suppressed shared error-state writes, partial-runtime teardown on late
activation failure, and retry after a dead handle is discarded.

## Verification

- `pnpm vitest run
src/__tests__/plugin-environment-driver-ready-recovery.test.ts` (in
`server/`) passed: 10 tests.
- `pnpm --filter @paperclipai/server typecheck` passed.

## Risks

- Low risk for self-hosted single-process deployments because lazy
recovery is only wired when managed plugin auto-install config is
present; with no managed config the capabilities route takes the exact
pre-change code path.
- The capabilities route can wait briefly while attempting recovery; the
attempt is bounded and defaults to 2 seconds.
- Failed recovery keeps the prior behavior of omitting the provider
until a later successful worker start, and now also cleans up any
partially-started local worker so retries begin from a clean slate.

## Model Used

- Initial implementation: OpenAI GPT-5 via Codex local coding agent,
with repository tool use and command execution.
- Review-feedback follow-ups (driver-kind gate, partial-runtime
teardown, expanded regression tests): Claude Fable 5 (claude-fable-5)
via Claude Code, with repository tool use and command execution.

## Checklist

- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have searched GitHub for duplicate or related PRs and linked
them above
- [x] I have either (a) linked existing issues with `Fixes: #` / `Closes
#` / `Refs #` OR (b) described the issue in-PR following the relevant
issue template
- [x] I have not referenced internal/instance-local Paperclip issues or
links (only public GitHub `#NNN` / `github.com/paperclipai/paperclip`
URLs)
- [x] My branch name describes the change (e.g. `docs/...`, `fix/...`)
and contains no internal Paperclip ticket id or instance-derived details
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [x] I have updated relevant documentation to reflect my changes
- [x] I have considered and documented any risks above
- [x] All Paperclip CI gates are green
- [x] Greptile is 5/5 with no open P2s, recommendations, or follow-ups
- [x] I will address all Greptile and reviewer comments before
requesting merge

---------

Co-authored-by: Paperclip <noreply@paperclip.ing>
2026-07-29 12:50:16 -07:00

745 lines
29 KiB
TypeScript

import express, { Router, type Request as ExpressRequest } from "express";
import path from "node:path";
import fs from "node:fs";
import { fileURLToPath } from "node:url";
import type { Db } from "@paperclipai/db";
import type { DeploymentExposure, DeploymentMode } from "@paperclipai/shared";
import type { InspectDatabaseBackupHealthOptions } from "./services/database-backup-health.js";
import type { StorageService } from "./storage/types.js";
import { httpLogger, errorHandler } from "./middleware/index.js";
import { actorMiddleware } from "./middleware/auth.js";
import { boardMutationGuard } from "./middleware/board-mutation-guard.js";
import { privateHostnameGuard, resolvePrivateHostnameAllowSet } from "./middleware/private-hostname-guard.js";
import { applyTrustProxy, parseTrustProxyEnv } from "./middleware/trust-proxy.js";
import { healthRoutes } from "./routes/health.js";
import { companyRoutes } from "./routes/companies.js";
import { companySkillRoutes } from "./routes/company-skills.js";
import { companySkillPolicyRoutes } from "./routes/company-skill-policy.js";
import { inboxAgentPolicyRoutes } from "./routes/inbox-agent-policy.js";
import { builtInAgentRoutes } from "./routes/built-in-agents.js";
import { folderRoutes } from "./routes/folders.js";
import { summarySlotRoutes } from "./routes/summary-slots.js";
import { statusCardRoutes } from "./routes/status-cards.js";
import { teamsCatalogRoutes } from "./routes/teams-catalog.js";
import { agentRoutes } from "./routes/agents.js";
import { projectRoutes } from "./routes/projects.js";
import { issueRoutes } from "./routes/issues.js";
import { issueTreeControlRoutes } from "./routes/issue-tree-control.js";
import { caseRoutes } from "./routes/cases.js";
import { fileResourceRoutes } from "./routes/file-resources.js";
import { routineRoutes } from "./routes/routines.js";
import { pipelineRoutes } from "./routes/pipelines.js";
import { environmentRoutes } from "./routes/environments.js";
import { executionWorkspaceRoutes } from "./routes/execution-workspaces.js";
import { goalRoutes } from "./routes/goals.js";
import { boardChatRoutes } from "./routes/board-chat.js";
import { approvalRoutes } from "./routes/approvals.js";
import { secretRoutes } from "./routes/secrets.js";
import { toolAccessRoutes } from "./routes/tool-access.js";
import { smokeLabRoutes } from "./routes/smoke-lab.js";
import { costRoutes } from "./routes/costs.js";
import { activityRoutes } from "./routes/activity.js";
import { dashboardRoutes } from "./routes/dashboard.js";
import { attentionRoutes } from "./routes/attention.js";
import { decisionTrainingRoutes } from "./routes/decision-training.js";
import { userProfileRoutes } from "./routes/user-profiles.js";
import { sidebarBadgeRoutes } from "./routes/sidebar-badges.js";
import { sidebarPreferenceRoutes } from "./routes/sidebar-preferences.js";
import { resourceMembershipRoutes } from "./routes/resource-memberships.js";
import { inboxDismissalRoutes } from "./routes/inbox-dismissals.js";
import { instanceSettingsRoutes } from "./routes/instance-settings.js";
import { openApiRoutes } from "./routes/openapi.js";
import {
instanceDatabaseBackupRoutes,
type InstanceDatabaseBackupService,
} from "./routes/instance-database-backups.js";
import { llmRoutes } from "./routes/llms.js";
import { authRoutes } from "./routes/auth.js";
import { assetRoutes } from "./routes/assets.js";
import { accessRoutes } from "./routes/access.js";
import { pluginRoutes } from "./routes/plugins.js";
import { mcpGatewayProtocolRoutes, toolGatewayRoutes } from "./routes/tool-gateway.js";
import { adapterRoutes } from "./routes/adapters.js";
import { pluginUiStaticRoutes } from "./routes/plugin-ui-static.js";
import { readBrandedStaticIndexHtml } from "./static-index-html.js";
import { applyUiBranding } from "./ui-branding.js";
import { logger } from "./middleware/logger.js";
import { DEFAULT_LOCAL_PLUGIN_DIR, pluginLoader, type PluginLoader } from "./services/plugin-loader.js";
import {
SELF_HOSTED_AUTO_INSTALL_KEYS,
ensureBundledPlugins,
resolveBundledCatalogRoot,
resolveBundledPluginInstalls,
} from "./services/bundled-plugins.js";
import { createPluginWorkerManager, type PluginWorkerManager } from "./services/plugin-worker-manager.js";
import { createPluginJobScheduler } from "./services/plugin-job-scheduler.js";
import { pluginJobStore } from "./services/plugin-job-store.js";
import { createPluginToolDispatcher } from "./services/plugin-tool-dispatcher.js";
import { createToolGatewayService } from "./services/tool-gateway.js";
import { pluginLifecycleManager } from "./services/plugin-lifecycle.js";
import { createPluginJobCoordinator } from "./services/plugin-job-coordinator.js";
import { buildHostServices, flushPluginLogBuffer } from "./services/plugin-host-services.js";
import { createPluginEventBus } from "./services/plugin-event-bus.js";
import { setPluginEventBus } from "./services/activity-log.js";
import { createPluginDevWatcher } from "./services/plugin-dev-watcher.js";
import { createPluginHostServiceCleanup } from "./services/plugin-host-service-cleanup.js";
import { pluginRegistryService } from "./services/plugin-registry.js";
import { createHostClientHandlers } from "@paperclipai/plugin-sdk";
import type { BetterAuthSessionResult } from "./auth/better-auth.js";
import { createCachedViteHtmlRenderer } from "./vite-html-renderer.js";
import { DEFAULT_JSON_BODY_LIMIT, PORTABLE_JSON_BODY_LIMIT } from "./http/body-limits.js";
import { COMPANY_IMPORT_API_PATH } from "./routes/company-import-paths.js";
import { apiCompression } from "./middleware/api-compression.js";
type UiMode = "none" | "static" | "vite-dev";
const FEEDBACK_EXPORT_FLUSH_INTERVAL_MS = 5_000;
const VITE_DEV_ASSET_PREFIXES = [
"/@fs/",
"/@id/",
"/@react-refresh",
"/@vite/",
"/assets/",
"/node_modules/",
"/src/",
];
const VITE_DEV_STATIC_PATHS = new Set([
"/apple-touch-icon.png",
"/favicon-16x16.png",
"/favicon-32x32.png",
"/favicon.ico",
"/favicon.svg",
"/site.webmanifest",
"/sw.js",
]);
export function isDatabaseConnectionUnavailableError(err: unknown): boolean {
const error = err as { code?: unknown; message?: unknown; cause?: unknown };
if (error?.code === "ECONNREFUSED") return true;
return Boolean(error?.cause && isDatabaseConnectionUnavailableError(error.cause));
}
export function resolveViteHmrPort(serverPort: number): number {
if (serverPort <= 55_535) {
return serverPort + 10_000;
}
return Math.max(1_024, serverPort - 10_000);
}
export function resolveViteHmrHost(bindHost: string): string | undefined {
const normalized = bindHost.trim().toLowerCase();
if (normalized === "0.0.0.0" || normalized === "::") return undefined;
return bindHost;
}
export function shouldServeViteDevHtml(req: ExpressRequest): boolean {
const pathname = req.path;
if (VITE_DEV_STATIC_PATHS.has(pathname)) return false;
if (VITE_DEV_ASSET_PREFIXES.some((prefix) => pathname.startsWith(prefix))) return false;
return req.accepts(["html"]) === "html";
}
export function shouldEnablePrivateHostnameGuard(opts: {
deploymentMode: DeploymentMode;
deploymentExposure: DeploymentExposure;
}): boolean {
return (
opts.deploymentExposure === "private" &&
(opts.deploymentMode === "local_trusted" || opts.deploymentMode === "authenticated")
);
}
export function createManagedBundledPluginWorkerRecovery(input: {
managedBundledPluginKeys: readonly string[];
workerManager: Pick<PluginWorkerManager, "getWorker" | "isRunning" | "stopWorker">;
getLoader: () => Pick<PluginLoader, "loadSingle"> | null;
}): (plugin: { id: string; pluginKey: string }) => Promise<boolean> {
const recoverablePluginKeys = new Set(input.managedBundledPluginKeys);
const inFlightStarts = new Map<string, Promise<boolean>>();
// A failed attempt can leave behind the dead handle it registered (e.g. the
// worker process died during initialize, which kills the process without
// scheduling a restart). No pre-existing handle survives to a recovery
// attempt — recovery only starts when getWorker() was empty — so discarding
// the dead handle lets a later capability request retry instead of being
// blocked by the handle-presence gate until the process restarts. Handles
// in starting/running/backoff states belong to the worker manager's own
// lifecycle and are left alone.
const discardDeadRecoveryHandle = async (plugin: { id: string; pluginKey: string }) => {
const handle = input.workerManager.getWorker(plugin.id);
if (!handle || (handle.status !== "crashed" && handle.status !== "stopped")) return;
try {
await input.workerManager.stopWorker(plugin.id);
} catch (err) {
logger.warn(
{
pluginId: plugin.id,
pluginKey: plugin.pluginKey,
err: err instanceof Error ? err.message : String(err),
},
"failed to discard dead plugin worker handle after recovery failure",
);
}
};
return async (plugin) => {
if (!recoverablePluginKeys.has(plugin.pluginKey)) return false;
const inFlight = inFlightStarts.get(plugin.id);
if (inFlight) return inFlight;
const startPromise = (async () => {
if (input.workerManager.getWorker(plugin.id)) {
return input.workerManager.isRunning(plugin.id);
}
const loader = input.getLoader();
if (!loader) return false;
try {
const result = await loader.loadSingle(plugin.id, {
markErrorOnFailure: false,
});
if (result.success === true || input.workerManager.isRunning(plugin.id)) {
return true;
}
await discardDeadRecoveryHandle(plugin);
return false;
} catch (err) {
logger.warn(
{
pluginId: plugin.id,
pluginKey: plugin.pluginKey,
err: err instanceof Error ? err.message : String(err),
},
"managed bundled plugin lazy worker recovery failed",
);
await discardDeadRecoveryHandle(plugin);
throw err;
}
})();
inFlightStarts.set(plugin.id, startPromise);
try {
return await startPromise;
} finally {
if (inFlightStarts.get(plugin.id) === startPromise) {
inFlightStarts.delete(plugin.id);
}
}
};
}
export async function createApp(
db: Db,
opts: {
uiMode: UiMode;
serverPort: number;
storageService: StorageService;
feedbackExportService?: {
flushPendingFeedbackTraces(input?: {
companyId?: string;
traceId?: string;
limit?: number;
now?: Date;
}): Promise<unknown>;
};
databaseBackupService?: InstanceDatabaseBackupService;
databaseBackupHealth?: InspectDatabaseBackupHealthOptions;
deploymentMode: DeploymentMode;
deploymentExposure: DeploymentExposure;
allowedHostnames: string[];
bindHost: string;
authReady: boolean;
companyDeletionEnabled: boolean;
instanceId?: string;
hostVersion?: string;
localPluginDir?: string;
pluginMigrationDb?: Db;
pluginWorkerManager?: PluginWorkerManager;
betterAuthHandler?: express.RequestHandler;
resolveSession?: (req: ExpressRequest) => Promise<BetterAuthSessionResult | null>;
/**
* `plugins.autoInstall` from the managed config (PAPERCLIP_MANAGED_CONFIG).
* `null`/absent ⇒ self-hosted: only the built-in kubernetes bundle is
* ensured, exactly as before. A managed list is resolved against the
* bundled catalog fail-to-start (see services/bundled-plugins.ts).
*/
managedPluginAutoInstall?: readonly string[] | null;
/** Test override for the bundled plugin catalog root. */
bundledPluginCatalogRoot?: string;
},
) {
const app = express();
app.locals.paperclipDb = db;
const captureRawBody = (req: express.Request, _res: express.Response, buf: Buffer) => {
(req as unknown as { rawBody: Buffer }).rawBody = buf;
};
// Respect the operator's `TRUST_PROXY` env var (see middleware/trust-proxy.ts).
// Default is unset → Express trusts nothing, which is the only safe choice
// when the server may be reachable without a known reverse proxy in front.
applyTrustProxy(app, parseTrustProxyEnv(process.env.TRUST_PROXY));
app.use(COMPANY_IMPORT_API_PATH, express.json({
limit: PORTABLE_JSON_BODY_LIMIT,
verify: captureRawBody,
}));
app.use(express.json({
limit: DEFAULT_JSON_BODY_LIMIT,
verify: captureRawBody,
}));
app.use("/api", apiCompression());
app.use(httpLogger);
const privateHostnameGateEnabled = shouldEnablePrivateHostnameGuard({
deploymentMode: opts.deploymentMode,
deploymentExposure: opts.deploymentExposure,
});
const privateHostnameAllowSet = resolvePrivateHostnameAllowSet({
allowedHostnames: opts.allowedHostnames,
bindHost: opts.bindHost,
});
app.use(
privateHostnameGuard({
enabled: privateHostnameGateEnabled,
allowedHostnames: opts.allowedHostnames,
bindHost: opts.bindHost,
}),
);
app.use(
actorMiddleware(db, {
deploymentMode: opts.deploymentMode,
resolveSession: opts.resolveSession,
}),
);
app.use("/api/auth", authRoutes(db));
if (opts.betterAuthHandler) {
app.all("/api/auth/{*authPath}", opts.betterAuthHandler);
}
app.use(llmRoutes(db));
const hostServicesDisposers = new Map<string, () => void>();
const workerManager = opts.pluginWorkerManager ?? createPluginWorkerManager();
const managedAutoInstallKeys = opts.managedPluginAutoInstall ?? null;
const bundledCatalogRoot =
opts.bundledPluginCatalogRoot ?? resolveBundledCatalogRoot(process.env);
const bundledPluginInstalls = resolveBundledPluginInstalls(
managedAutoInstallKeys ?? SELF_HOSTED_AUTO_INSTALL_KEYS,
{
catalogRoot: bundledCatalogRoot,
env: process.env,
enforceCatalogRoot: managedAutoInstallKeys !== null,
},
);
const managedBundledPluginKeys =
managedAutoInstallKeys !== null
? bundledPluginInstalls.map((install) => install.pluginKey)
: [];
let runtimePluginLoader: Pick<PluginLoader, "loadSingle"> | null = null;
// A sibling process can install a managed bundled plugin while this process
// skips the mid-install row, then finish the row after this process's
// loadAll() pass. The capabilities route may recover only those managed
// bundles by starting their ready-but-unstarted worker lazily.
const recoverManagedBundledPluginWorker =
managedAutoInstallKeys !== null
? createManagedBundledPluginWorkerRecovery({
managedBundledPluginKeys,
workerManager,
getLoader: () => runtimePluginLoader,
})
: undefined;
// Mount API routes
const api = Router();
api.use(boardMutationGuard());
api.use(
"/health",
healthRoutes(db, {
deploymentMode: opts.deploymentMode,
deploymentExposure: opts.deploymentExposure,
authReady: opts.authReady,
companyDeletionEnabled: opts.companyDeletionEnabled,
databaseBackupHealth: opts.databaseBackupHealth,
}),
);
api.use(openApiRoutes());
api.use("/companies", companyRoutes(db, opts.storageService));
api.use(llmRoutes(db));
api.use(folderRoutes(db));
api.use(companySkillRoutes(db));
api.use(companySkillPolicyRoutes(db));
api.use(inboxAgentPolicyRoutes(db));
api.use(builtInAgentRoutes(db));
api.use(summarySlotRoutes(db));
api.use(statusCardRoutes(db));
api.use(teamsCatalogRoutes(db));
api.use(agentRoutes(db, { pluginWorkerManager: workerManager }));
api.use(assetRoutes(db, opts.storageService));
api.use(projectRoutes(db));
api.use(caseRoutes(db, opts.storageService));
api.use(issueTreeControlRoutes(db));
api.use(fileResourceRoutes(db));
api.use(routineRoutes(db, { pluginWorkerManager: workerManager }));
api.use(pipelineRoutes(db));
api.use(environmentRoutes(db, {
pluginWorkerManager: workerManager,
recoverMissingPluginWorker: recoverManagedBundledPluginWorker
? {
pluginKeys: managedBundledPluginKeys,
startWorker: recoverManagedBundledPluginWorker,
}
: undefined,
}));
api.use(executionWorkspaceRoutes(db, { pluginWorkerManager: workerManager }));
api.use(goalRoutes(db));
api.use(boardChatRoutes(db, { deploymentMode: opts.deploymentMode }));
api.use(approvalRoutes(db, { pluginWorkerManager: workerManager }));
api.use(secretRoutes(db));
const trustedLocalStdioRuntimeHost =
process.env.PAPERCLIP_TRUSTED_MCP_RUNTIME_HOST
?? process.env.PAPERCLIP_TOOL_RUNTIME_TRUSTED_HOST
?? null;
api.use(costRoutes(db, { pluginWorkerManager: workerManager }));
api.use(activityRoutes(db));
api.use(dashboardRoutes(db));
api.use(attentionRoutes(db));
api.use(decisionTrainingRoutes(db));
api.use(userProfileRoutes(db));
api.use(sidebarBadgeRoutes(db));
api.use(sidebarPreferenceRoutes(db));
api.use(resourceMembershipRoutes(db));
api.use(inboxDismissalRoutes(db));
api.use(instanceSettingsRoutes(db));
if (opts.databaseBackupService) {
api.use(instanceDatabaseBackupRoutes(opts.databaseBackupService));
}
const pluginRegistry = pluginRegistryService(db);
const eventBus = createPluginEventBus();
setPluginEventBus(eventBus);
const jobStore = pluginJobStore(db);
const lifecycle = pluginLifecycleManager(db, { workerManager });
const scheduler = createPluginJobScheduler({
db,
jobStore,
workerManager,
});
const toolDispatcher = createPluginToolDispatcher({
workerManager,
lifecycleManager: lifecycle,
db,
});
const toolGateway = createToolGatewayService(db, {
pluginToolDispatcher: toolDispatcher,
deploymentMode: opts.deploymentMode,
deploymentExposure: opts.deploymentExposure,
trustedLocalStdioRuntimeHost,
});
// Issue routes are intentionally mounted after the gateway is constructed because
// issue approval endpoints delegate to it. The intervening routers use distinct
// route prefixes, so this dependency does not change issue-route precedence.
api.use(issueRoutes(db, opts.storageService, {
feedbackExportService: opts.feedbackExportService,
pluginWorkerManager: workerManager,
approveToolActionRequest: (input) => toolGateway.approveActionRequest(input),
}));
app.use(mcpGatewayProtocolRoutes(toolGateway));
api.use(toolAccessRoutes(db, {
deploymentMode: opts.deploymentMode,
deploymentExposure: opts.deploymentExposure,
trustedLocalStdioRuntimeHost,
toolGateway,
}));
api.use(smokeLabRoutes(db, {
deploymentMode: opts.deploymentMode,
deploymentExposure: opts.deploymentExposure,
}));
const jobCoordinator = createPluginJobCoordinator({
db,
lifecycle,
scheduler,
jobStore,
});
const hostServiceCleanup = createPluginHostServiceCleanup(lifecycle, hostServicesDisposers);
let viteHtmlRenderer: ReturnType<typeof createCachedViteHtmlRenderer> | null = null;
const loader = pluginLoader(
db,
{
localPluginDir: opts.localPluginDir ?? DEFAULT_LOCAL_PLUGIN_DIR,
migrationDb: opts.pluginMigrationDb,
},
{
workerManager,
eventBus,
jobScheduler: scheduler,
jobStore,
toolDispatcher,
lifecycleManager: lifecycle,
instanceInfo: {
instanceId: opts.instanceId ?? "default",
hostVersion: opts.hostVersion ?? "0.0.0",
deploymentMode: opts.deploymentMode,
deploymentExposure: opts.deploymentExposure,
},
buildHostHandlers: (pluginId, manifest) => {
const notifyWorker = (method: string, params: unknown) => {
const handle = workerManager.getWorker(pluginId);
if (handle) handle.notify(method, params);
};
const services = buildHostServices(db, pluginId, manifest.id, eventBus, notifyWorker, {
pluginWorkerManager: workerManager,
manifest,
});
hostServicesDisposers.set(pluginId, () => services.dispose());
return createHostClientHandlers({
pluginId,
capabilities: manifest.capabilities,
services,
});
},
},
);
runtimePluginLoader = loader;
api.use(
toolGatewayRoutes(db, toolGateway),
);
api.use(
pluginRoutes(
db,
loader,
{ scheduler, jobStore },
{ workerManager },
{ toolDispatcher },
{ workerManager },
{ toolGateway },
),
);
api.use(adapterRoutes());
api.use(
accessRoutes(db, {
deploymentMode: opts.deploymentMode,
deploymentExposure: opts.deploymentExposure,
bindHost: opts.bindHost,
allowedHostnames: opts.allowedHostnames,
}),
);
app.use("/api", api);
app.use("/api", (_req, res) => {
res.status(404).json({ error: "API route not found" });
});
app.use(pluginUiStaticRoutes(db, {
localPluginDir: opts.localPluginDir ?? DEFAULT_LOCAL_PLUGIN_DIR,
}));
const __dirname = path.dirname(fileURLToPath(import.meta.url));
if (opts.uiMode === "static") {
// Try published location first (server/ui-dist/), then monorepo dev location (../../ui/dist)
const candidates = [
path.resolve(__dirname, "../ui-dist"),
path.resolve(__dirname, "../../ui/dist"),
];
const uiDist = candidates.find((p) => fs.existsSync(path.join(p, "index.html")));
if (uiDist) {
// Hashed asset files (Vite emits them under /assets/<name>.<hash>.<ext>)
// never change once built, so they can be cached aggressively.
app.use(
"/assets",
express.static(path.join(uiDist, "assets"), {
maxAge: "1y",
immutable: true,
}),
);
// Non-hashed static files (favicon.ico, manifest, robots.txt, etc.):
// short cache so operators who swap them out see the new version
// reasonably fast. Override for `index.html` specifically — it is
// served by this middleware for `/` and `/index.html`, and it must
// never outlive the asset hashes it points at.
app.use(
express.static(uiDist, {
maxAge: "1h",
setHeaders(res, filePath) {
if (path.basename(filePath) === "index.html") {
res.set("Cache-Control", "no-cache");
}
},
}),
);
// SPA fallback. Only for non-asset routes — if the browser asks for
// /assets/something.js that doesn't exist, we must NOT serve the HTML
// shell: the browser would try to load it as a JavaScript module, fail
// with a MIME-type error, and cache that broken response. Return 404
// instead. The index.html response itself is no-cache so a subsequent
// deploy's updated asset hashes are picked up on next load.
app.get(/.*/, (req, res) => {
if (req.path.startsWith("/assets/")) {
res.status(404).end();
return;
}
res
.status(200)
.set("Content-Type", "text/html")
.set("Cache-Control", "no-cache")
.end(readBrandedStaticIndexHtml(uiDist));
});
} else {
console.warn("[paperclip] UI dist not found; running in API-only mode");
}
}
if (opts.uiMode === "vite-dev") {
const uiRoot = path.resolve(__dirname, "../../ui");
const publicUiRoot = path.resolve(uiRoot, "public");
const hmrPort = resolveViteHmrPort(opts.serverPort);
const hmrHost = resolveViteHmrHost(opts.bindHost);
const { createServer: createViteServer } = await import("vite");
const vite = await createViteServer({
root: uiRoot,
appType: "custom",
server: {
middlewareMode: true,
hmr: {
...(hmrHost ? { host: hmrHost } : {}),
port: hmrPort,
clientPort: hmrPort,
},
allowedHosts: privateHostnameGateEnabled ? Array.from(privateHostnameAllowSet) : undefined,
},
});
viteHtmlRenderer = createCachedViteHtmlRenderer({
vite,
uiRoot,
brandHtml: applyUiBranding,
});
const renderViteHtml = viteHtmlRenderer;
if (fs.existsSync(publicUiRoot)) {
app.use(express.static(publicUiRoot, { index: false }));
}
app.get(/.*/, async (req, res, next) => {
if (!shouldServeViteDevHtml(req)) {
next();
return;
}
try {
const html = await renderViteHtml.render(req.originalUrl);
res.status(200).set({ "Content-Type": "text/html" }).end(html);
} catch (err) {
next(err);
}
});
app.use(vite.middlewares);
}
app.use(errorHandler);
jobCoordinator.start();
scheduler.start();
let feedbackExportShuttingDown = false;
let feedbackExportTimer: ReturnType<typeof setInterval> | null = null;
const disableFeedbackExportFlushes = () => {
feedbackExportShuttingDown = true;
if (feedbackExportTimer) {
clearInterval(feedbackExportTimer);
feedbackExportTimer = null;
}
};
const flushPendingFeedbackExports = async () => {
if (feedbackExportShuttingDown) return;
try {
await opts.feedbackExportService?.flushPendingFeedbackTraces();
} catch (err) {
if (isDatabaseConnectionUnavailableError(err)) {
disableFeedbackExportFlushes();
logger.warn({ err }, "Disabling pending feedback export flushes because the database is unavailable");
return;
}
logger.error({ err }, "Failed to flush pending feedback exports");
}
};
feedbackExportTimer = opts.feedbackExportService
? setInterval(() => {
void flushPendingFeedbackExports();
}, FEEDBACK_EXPORT_FLUSH_INTERVAL_MS)
: null;
feedbackExportTimer?.unref?.();
if (opts.feedbackExportService) {
void flushPendingFeedbackExports();
}
void toolDispatcher.initialize().catch((err) => {
logger.error({ err }, "Failed to initialize plugin tool dispatcher");
});
const devWatcher = createPluginDevWatcher(
lifecycle,
async (pluginId) => (await pluginRegistry.getById(pluginId))?.packagePath ?? null,
);
// Auto-provision bundled plugins so their providers are registered for
// agent runs. Bundles are excluded from the pnpm
// workspace and built standalone into the image (see Dockerfile), then
// installed here from their local paths. This runs BEFORE loadAll() so
// loadAll() can activate them in the same startup pass.
//
// Workers are started exactly once, by loadAll(): the `lifecycle` manager
// above is constructed without a runtime-capable loader
// (pluginLifecycleManager(db, { workerManager }) — no `loader` option), so
// the lifecycle.load() that ensureBundledPlugins performs per newly
// installed bundle only records the `ready` status and does not spawn a
// worker (see activateReadyPlugin in services/plugin-lifecycle.ts).
//
// Managed instances (`plugins.autoInstall` from PAPERCLIP_MANAGED_CONFIG)
// drive the key list from the control plane; self-hosted instances keep
// the pre-existing behavior of ensuring only the kubernetes bundle.
//
// Resolution is deliberately synchronous and NOT fail-safe: an
// unknown key or a path escaping the bundled catalog root throws out of
// createApp so a managed instance refuses to start (positive allowlist,
// fail closed).
// SAFETY: installation is fully fail-safe. Any failure
// (missing bundle, install error, load error) is caught, logged, and
// swallowed per plugin so the server ALWAYS finishes booting. A degraded
// boot (a provider unavailable, some agents cannot run) is strictly
// preferable to a crash loop.
//
// The chain is not awaited here (createApp stays fast), but the settled
// promise is exposed via `app.locals.bundledPluginsStartup` so boot steps
// that must not outrun plugin availability — managed sandbox environments
// (`applyManagedEnvironments`) run before the heartbeat resumes queued
// runs — can sequence on it. It never rejects.
const bundledPluginsStartup = ensureBundledPlugins(
bundledPluginInstalls,
{ registry: pluginRegistry, loader, lifecycle, logger },
// Managed mode reinstalls soft-uninstalled bundles (the control plane
// owns provisioning); self-hosted leaves an operator's uninstall alone.
// Operator-DISABLED plugins are never touched in either mode.
{ reinstallUninstalled: managedAutoInstallKeys !== null },
)
.then(() => loader.loadAll())
.then((result) => {
if (!result) return;
for (const loaded of result.results) {
if (devWatcher && loaded.success && loaded.plugin.packagePath) {
devWatcher.watch(loaded.plugin.id, loaded.plugin.packagePath);
}
}
}).catch((err) => {
logger.error({ err }, "Failed to load ready plugins on startup");
});
app.locals.bundledPluginsStartup = bundledPluginsStartup;
let appServicesShutdown = false;
const shutdownAppServices = () => {
if (appServicesShutdown) return;
appServicesShutdown = true;
disableFeedbackExportFlushes();
devWatcher?.close();
viteHtmlRenderer?.dispose();
hostServiceCleanup.disposeAll();
hostServiceCleanup.teardown();
};
app.locals.paperclipShutdown = shutdownAppServices;
process.once("exit", shutdownAppServices);
process.once("beforeExit", () => {
void flushPluginLogBuffer();
});
return app;
}