基础设施注册

Registrars for hooks, Gateway HTTP routes and RPC methods, CLI entries, and background services, plus the SDK helpers those surfaces depend on. Part of the Plugin SDK overview.

Infrastructure¶

Gateway methods default to profileAccess: "required", so authenticated-profile verification fails closed before plugin dispatch. Set profileAccess: "independent" only for an audited method that neither reads nor mutates durable user or session state. Operator scope remains a separate authorization requirement.

File-watch capacity errors¶

getFileWatchCapacityCode(error) from openclaw/plugin-sdk/file-access-runtime returns EMFILE, ENFILE, or ENOSPC for a native watch failure, or undefined for other errors. It requires syscall: "watch" because watcher libraries can forward directory-scan errors through the same error event. Use the result in the watcher lifecycle owner to stop native retries and select an existing refresh path.

Filesystem observation and worker notifications¶

resolveFsObservationMode(env?) and resolveFsObservationIntervalMs(env?) from openclaw/plugin-sdk/file-access-runtime share the host's preserved CHOKIDAR_* environment contract. Use admitObservationRoot, watch, and their types from the same SDK entrypoint, including ObservationRoot, WatchOptions, and WatchSubscription. These operations share the host's fs-safe instance; a plugin's separate dependency copy cannot observe those Roots. Pass the resolved mode and pollIntervalMs to watch so automatic fallback preserves the polling interval. Keep parsing, settling, retries, and indexing in the consumer. With fs-safe, classify native watch capacity through health.failure.operation === "watch" and health.failure.code === "watch-limit"; getFileWatchCapacityCode retains its existing Node watch-error contract.

For same-version observation workers, createFileWatchNotifier(output, onFailure) from the same SDK entrypoint sends JSON lines through a borrowed writable stream. Call send("change" | "unavailable" | "available") for invalidation and availability updates. It coalesces pending notifications, keeps one write in flight, and calls onFailure when output fails or closes unexpectedly. Await close() to stop accepting notifications and join accepted writes before retiring the worker; the stream remains caller-owned. This carries current observation state, not a complete history of filesystem events.

Streaming file verification¶

sha256File(pathOrHandle, { maxBytes, signal }) from openclaw/plugin-sdk/file-access-runtime returns { bytes, digest } without loading the whole file into memory. It reads through EOF and rejects files that grow beyond the byte limit. A borrowed handle stays open at its original offset; the caller owns admission and close. Path inputs reject final symlinks and close their owned handle. Cancellation settles pending work before rejecting. The optional native helper hashes off the JavaScript event loop; the fallback uses bounded buffers. Neither route provides a snapshot of concurrent writes.

Browser lifecycle cleanup¶

closeTrackedBrowserTabsForSessions from openclaw/plugin-sdk/browser-maintenance accepts an optional prepareCurrent(): Promise<boolean> check after plugin activation and before each new cleanup claim. Returning false skips new claims; the existing isCurrent() callback remains a synchronous owner check after awaited preparation. A host-supplied sessionEntryCurrent check restricts native claim and pre-claim state writes using current session facts; it does not grant store access. Supplying sessionEntryCurrent also requires prepareCurrent, which checks process-local tabs before they acquire a cleanup reservation. Unpaired checks are refused with a warning before tab cleanup begins. Official plugins share the SessionEntryCurrentPreparation and SessionEntryCurrentCheck types through openclaw/plugin-sdk/plugin-state-runtime. Once a tab is claimed, closing and retiring that tab finish under its captured Browser authority even if the cleanup caller subsequently changes. Artifacts advertise this contract with supportsSessionEntryCurrent: true. Guarded cleanup against an older artifact leaves tabs untouched and reports an update warning; callers using only the existing synchronous guard remain supported.

SQLite write admission¶

runSqliteImmediateTransaction(db, prepare, options?) from openclaw/plugin-sdk/sqlite-runtime waits for write admission without blocking the event loop. Its asynchronous prepare function may run more than once when another writer holds the database. Keep preparation repeatable: read and plan there, then return a synchronous transaction callback. Revalidate current owner and row predicates inside that callback before writing.

Returning undefined from preparation skips the write and resolves the helper to undefined, even while another writer remains active. Otherwise, the helper resolves to the transaction callback's result. It rejects an already active transaction or preparation that leaves a transaction open. Once admitted, the callback runs once; callback and commit failures are never replayed.

Admission retries use the connection's existing busy_timeout; this is not a total deadline for preparation or transaction execution. options supplies the same transaction diagnostics as runSqliteImmediateTransactionSync. Keep the database handle and its owning operation alive until the returned promise settles. The callback and SQLite calls still run synchronously on the caller's thread.

For a repeated fixed query, prepareSqliteQuerySync(db, build) compiles its Kysely shape once and binds fresh parameters on each call. It uses the normal synchronous executor and the connection's bounded statement cache when enabled. Keep the prepared function with its database owner and discard it when closing the connection; transaction callbacks must remain synchronous.

Worker task admission¶

WorkerTaskPool from openclaw/plugin-sdk/process-runtime supports reusable computation workers for bundled and separately published official plugins. Inside those workers, import serveWorkerTasks and the WorkerTaskControl type from openclaw/plugin-sdk/worker-task-server to avoid loading the host process and pool runtime. Both paths use the same task protocol.

The older serving exports in process-runtime remain for released official plugins. Bundled workers use worker-task-server; remove the older exports only after supported official plugin versions have migrated to hosts with this subpath.

Each pool defaults to 128 outstanding tasks and 256 MiB of reported input bytes, including queued, preparing, and running tasks. Set maxPendingTasks and maxPendingBytes when constructing a pool to choose different positive limits. Report known retained input with run(input, { inputBytes }), including buffers captured by an input factory. Omitted inputBytes counts as zero; this accounting does not measure serialized payload size, decoded data, results, or worker heaps.

Capacity exhaustion rejects run() with WorkerTaskError.code = "overloaded" before preparing or executing that input. Accepted work remains ordered within a single-worker pool. Report the rejected operation as unsuccessful; do not substitute an empty result or bypass the limit with synchronous execution. After accepted work settles, the same pool accepts new work again. A caller may retry a rejected operation after pressure drains and its original authority and deadline are revalidated; the pool does not retry it automatically.

For stateless computation, sharedCompute: true also shares an aggregate 128-task/256-MiB admission budget and CPU execution capacity with participating pools in the same isolate. Dedicated ordered pools retain their own execution capacity and still enforce their individual admission limits.

Pass static Node.js Worker settings in workerOptions. For per-worker settings, prepareWorker() runs once per Worker creation attempt and returns { options, temporaryDirectory? }. Its options shallowly override workerOptions: properties such as env, workerData, and resourceLimits replace the whole static property rather than merging nested values.

A returned temporaryDirectory transfers a newly allocated disposable directory to the pool. Preparation owns cleanup if it fails before returning. The pool removes the directory only after that Worker exits, including startup failure or cancellation, and reports deletion failures without replacing the task outcome. Worker exit releases execution capacity; close() also waits for pending file cleanup. Keep persistent data and files borrowed outside the Worker out of this directory.

When native termination fails, the pool retains that worker's input custody and capacity. retryFailedRetirements() retries only those failed retirements and joins native exit and pending file cleanup without interrupting healthy tasks or waiting for them to finish. It does not replay failed work or close the pool. An owner that is shutting down must stop new admissions, drain healthy tasks, and finish with close().

serveWorkerTasks supplies a third handler argument, WorkerTaskControl. Await control.runNativeSection(() => nativeOperation()) around each bounded native operation that must finish before its worker can be terminated. The fence also awaits a returned promise, for native libraries with asynchronous entrypoints. Keep unrelated work and rendering outside the fence; do not fence an entire document or a host request. Call control.throwIfCancelled() between pages or other units of work so cancellation cannot start another native operation.

Cancellation, deadlines, pool closure, and worker retirement close native-section admission atomically. An unfenced worker is terminated immediately; a fenced worker remains charged against admission and execution capacity until its current native operation finishes and the worker exits. A deadline requests cancellation; it cannot safely interrupt a stuck native call. Native sections must therefore have bounded inputs and must not wait for network, user input, or unbounded work.

SQLite worker stores¶

Use openSqliteWorkerStore<Operations> from openclaw/plugin-sdk/sqlite-runtime to move a feature's SQLite lifecycle off the application event loop. Call it from the main application thread with { moduleUrl, databasePath, input }. Operations maps each domain operation to its { input, output } types; store.execute({ type, input }, { signal }?) returns the corresponding output promise. The host currently rejects calls from other application workers, which need a shared host-broker connection.

This first host supports filesystem-backed databases only. Empty paths, SQLite URIs, :memory:, and OpenClaw's reserved incognito database basename are refused before normalization or worker admission. In-memory and incognito ownership remain pending; these locators must never become disk filenames.

The static local moduleUrl identifies a trusted feature module exporting createSqliteWorkerBackend(input, { databasePath }). Its factory owns database opening and schema setup; its synchronous execute(command) owns queries and transactions. close() may await cleanup outside transactions; the host retains the actor until that cleanup settles. Keep native handles, WAL maintenance, and prepared statements inside that backend. Send serializable domain commands and results across the boundary, never SQL strings, callbacks, or Kysely builders. Declare private build entries with openclaw.build.workerEntries and derive their locations from the loader's api.runtimeSource fact.

Pass existingOnly: true when acquisition must preserve a missing database. The call returns undefined without starting a worker or invoking a factory when the file is absent and no active actor retains that path. A cold existing open requires the module's explicit openExistingSqliteWorkerBackend(input, { databasePath }) export. The host never substitutes the ordinary creation factory. The existing factory must use SQLite's native read-only or existing-file opening mode and validate the current schema without creating or migrating it. A filesystem existence check followed by ordinary create-if-missing opening does not satisfy this contract.

The host checks physical identity before dispatching the existing factory and again before returning the store. Disappearance or replacement after admission rejects acquisition. Existing-only and ordinary clients share the same physical actor when their module and initialization input match; changing open intent does not rerun a factory or create another connection. Domain commands still own write permission and any later schema initialization. Existing-only acquisition provides no read-only capability for subsequent commands.

Abort signals remove operations that are still queued. Once dispatched, an operation retains its result or failure; cancellation does not prove rollback. close() rejects new work and drains that client's accepted operations. The last client also closes its database backend; the last backend releases its worker. Worker loss or failure to serialize a completed operation's result can reject with code: "outcome-unknown". The operation may have committed: inspect authoritative state before deciding what to do next. The host never automatically retries a write.

An asynchronous execute return violates the command contract. The host retires and joins that worker before reporting outcome-unknown; it does the same when a completed reply cannot be decoded. Failed cleanup retains its original error while the worker is drained.

The process-wide host starts lazily and uses two to eight shared Node workers based on available CPUs. Bun uses up to 64 dedicated workers until its native SQLite close fix ships. The host permits 64 opening or live store clients (including clients sharing a database), 128 outstanding operations per worker, and 256 MiB of queued and retained input across all workers. Count-only overflow waits in FIFO order on its worker for up to ten seconds; an independent worker keeps its own request capacity. Byte, message, and store limits refuse immediately with code: "overloaded". Each input message is limited to 32 MiB. Larger execute inputs arrive in 8 MiB chunks; the backend runs once after the complete command is validated. Factory initialization input remains a single bounded message.

Commands retaining at most 64 MiB of serialized input can queue, with their full byte length charged until settlement. A larger command must start immediately on an idle worker with a reserved 32 MiB transport window; otherwise it rejects with overloaded before dispatch. The aggregate budget bounds admitted queue bytes and reserved transport windows, not the complete value held by an active oversized command or result. Once staging starts, the existing post-dispatch cancellation and drainage rules apply.

Results up to 64 MiB use an inline reply; larger results transfer their complete serialized value in bounded 8 MiB chunks. Callers still materialize the complete result in memory, and the original operation remains owned through transfer validation and cleanup. Operations for one database share its connection owner and execute in order. There are no reader replicas or worker-pool configuration options.

File identities are admission facts, not native-handle attestations. Close and drain a database's clients before replacing or relocating its file. The host refuses observed identity changes and collisions with an existing owner; path checks cannot protect against an uncoordinated filesystem replacement. Each client retains its admitted lexical and canonical pathnames through drainage and close. The backend's opening paths remain pinned for its native lifetime; released secondary aliases do not accumulate while other clients live.

Computation worker entrypoints¶

For a plugin-owned worker, pass package: { name, distWorkerPath } to resolveRuntimeWorkerUrl from the same SDK subpath. Use the plugin's package.json name and a worker path relative to its dist directory. The descriptor then supports bundled and standalone installations, including renamed installation directories. Declare the worker's source entry in openclaw.build.workerEntries so package builds emit it.

Webhook body rejection¶

Use readWebhookBodyOrReject or readJsonWebhookBodyOrReject from openclaw/plugin-sdk/webhook-request-guards for bounded body reads. Return when the result is { ok: false }; the helper owns the error response and connection cleanup. Body byte limits and read timeouts remain separate from transport cleanup.

For a custom error representation after a response-first body read, await sendHttpRequestRejection(req, res, statusCode, body, contentType?) instead of calling res.end() and destroying the request. It preserves security headers, frames the complete error, then on Node and Node-compatible Bun HTTP transports closes the write side while keeping application body readers paused. Node's request backpressure bounds residual input buffering; cleanup allows at most one second, not another body-read timeout. A disconnected peer, malformed HTTP, or an exhausted cleanup budget can prevent delivery. Committed responses are closed without appending a replacement error or completing a partial successful body.

On these transports, rejections emit response close without finish. Use close for terminal cleanup or selected-error diagnostics; it does not prove delivery. Keep successful-response activity on finish, with the caller's success-status check, so an aborted request cannot report healthy activity.

Older Bun HTTP transports use native response completion because their raw socket operations do not flush the HTTP response. OpenClaw detects the native HTTP destroySoon implementation introduced by Bun's Node compatibility rework rather than relying on version labels shared by different canary builds. Queued HEAD rejections on newer Bun wait for response socket assignment, including builds without HTTP response-finish diagnostics. Older Bun can still report client connection resets during large outstanding uploads, even after delivering the complete error.

Gateway HTTP requests run in order on each connection, including their response lifetimes. A closing connection cannot admit later requests or upgrades. Queued requests apply input backpressure until earlier responses finish; finite pipelines drain in order. Use separate connections for concurrent requests. Keep the release hook returned by beginWebhookRequestPipelineOrReject in finally; it retains any selected rejection cleanup before releasing the in-flight slot.

Webhook transports can register their handler with registerPluginHttpRoute from openclaw/plugin-sdk/webhook-ingress. Gateway owns the listener, connection admission, request scope, and route lease handoff; the channel owns its signature verification and bounded body read.

For bundled callback setup and Doctor guidance, classifyGatewayProbePath(pathname) from the private openclaw/plugin-sdk/gateway-config-runtime facade identifies Gateway probe paths without loading webhook execution code. This facade is not part of the third-party SDK. Normalize callback input through new URL(rawPath, "http://localhost").pathname first. Results live, ready, and startup identify exact paths owned by probes on the Gateway port; choose a different webhook path. Results namespace and outside do not identify an exact probe route. The same private facade exports resolvePluginRoutePathContext and isProtectedPluginRoutePathFromContext for canonical protected-path checks. If the callback falls under a protected namespace, choose the channel's safe default path before moving the external callback or reverse proxy to the Gateway port. A legacy listener can still serve its old path during that migration.

For a shipped channel listener, registration can include legacyListener: { port, host? }. The Gateway forwards requests on that endpoint through the same plugin dispatch, preserving the original socket, URL, body, and response headers. The handler owns path and method rejection, including unknown paths. Core HTTP endpoints are never exposed on the compatibility port. Legacy listeners require auth: "plugin": the channel continues authenticating its old callback path, including paths under /api/channels. The Gateway port keeps its protected-path authentication policy. This exception applies only to requests received on the compatibility port; it grants no Gateway operator scopes and does not waive channel signature checks or work admission. getWebhookLegacyListener(req) returns its frozen configured { port, host? } endpoint, or undefined for an ordinary Gateway request; headers cannot set it. Filter account targets by this endpoint before signature resolution when old ports distinguished accounts sharing a path and secret. Ordinary Gateway requests still need an unambiguous account path or authentication identity.

The optional registration metadata health: { path, contentType? } preserves a shipped exact raw health target: 200 ok for ordinary HTTP methods, with Node's HEAD behavior and only the optional Content-Type. It applies only on the legacy port, including during route handoff, and does not expose Gateway probe details. Legacy ports retain native Node expectation handling, Upgrade fallback, header limits and timeout defaults. A shipped timeout profile can be preserved with timeouts: { headers, request, socket } in milliseconds. These are plugin registration contracts, not new operator configuration.

The channel owns effective listener resolution: preserve its shipped default when legacyWebhook is omitted, use an explicit endpoint object when configured, and register no legacy listener when it is false. Resolve the same endpoint for runtime routing and Doctor guidance. Plugin-owned Doctor contracts can compose createLegacyWebhookListenerDoctorContract from openclaw/plugin-sdk/runtime-doctor-migrations to preserve authored ports and inherited bind addresses through the normal backed-up config write. An explicit legacy host without a port uses the channel's shipped default port. Canonical false settings remain authoritative when Doctor removes retired keys. Return normal listener guidance in runConfigSequence().infoNotes so Doctor labels it as information. Keep actionable configuration problems in warningNotes; changeNotes describe applied repairs.

Account leases sharing a route can retain separate endpoints. Endpoints retained only by a restart handoff return retryable 503 responses; endpoints with live holders keep serving requests. A live holder at the same address takes precedence over a retained handoff. Live registrations sharing an endpoint must declare the same health and timeout profile; conflicting registrations are rejected without changing the listener. Bind failure warns without disabling the Gateway route. After the operator changes the provider callback or reverse proxy to reach the Gateway port, the plugin can stop registering the compatibility endpoint. Retiring an endpoint stops new connections while admitted responses finish. The Gateway keeps those closing sockets in its transport ownership and closes them on full shutdown; channels retain their own response-drain ordering before teardown.

Channel webhook listeners that own their createServer admission serialize each connection with runHttpConnectionRequest(req, run, res?) from openclaw/plugin-sdk/webhook-request-guards. Pass the ServerResponse as the third argument: the shared owner waits for response completion (finish or close) before admitting the connection's next request, so a close-aware rejection — whose cleanup may destroy the socket within one second — can never overtake an earlier queued acknowledgement. Omitting the response argument releases the next request before the current response finishes and loses that guarantee; omit it only for dispatch that writes no response on the shared connection. Already admitted work always finishes; queued work is never dispatched after closure, and a closing connection cannot admit later requests.

Post-ack webhook work¶

Webhook routes that acknowledge a request before processing finishes must move that detached work onto its own tracked admission root:

import { runDetachedWebhookWork } from "openclaw/plugin-sdk/webhook-request-guards";

// processWebhookEvent, event, and runtime are your plugin's own handler, payload, and logger.
void runDetachedWebhookWork(() => processWebhookEvent(event)).catch((error) => {
  runtime.error?.(`webhook dispatch failed: ${String(error)}`);
});

Call runDetachedWebhookWork(...) synchronously while the HTTP request is still admitted. The helper reserves an independent root immediately, then starts the callback in the next microtask so the request handler can write its acknowledgement first. The returned promise adopts the callback result; callers still own rejection handling. This keeps post-ack queue work accepted and makes restart or suspension drains wait for it. Handlers that await all processing before returning do not need this helper.

Requester-scoped MCP connections¶

Keep the MCP server identity static (name, tool filter) in mcp.servers, a native plugin's mcpServers manifest field, or a bundle manifest. Optionally register a connection resolver so each trusted message requester gets their own transport:

api.registerMcpServerConnectionResolver({
  serverName: "user-email",
  resolve: async (ctx) => {
    // ctx.requesterSenderId is host-trusted; never invent sender identity here.
    // lookupUserToken is your plugin's own credential-store lookup, not an SDK export.
    const token = await lookupUserToken(ctx.requesterSenderId);
    if (!token) {
      return null; // omit this server for the current run
    }
    return {
      url: "https://mcp.example.com/email",
      headers: { Authorization: `Bearer ${token}` },
    };
  },
});

Contract notes:

• Resolver context carries trusted host identity only (requesterSenderId, optional agentAccountId / messageChannel). Future trusted fields (for example cron/subagent user context) can be added additively.
• One plugin owns one server name: a duplicate registerMcpServerConnectionResolver for the same serverName from another plugin is rejected with an error diagnostic (first registration wins), so connection ownership never depends on plugin load order.
• Tool names are derived from the full declared server set so partial resolution never changes safe server names between requesters or turns. Core does not verify that different requester endpoints serve identical tool schemas; a resolver must point every requester at the same logical service, or tool schemas (and prompt-cache stability) diverge per requester.
• Runs without a trusted requesterSenderId (cron, subagent, heartbeat, public gateway) never materialize requester-scoped servers. There is no shared fallback connection.
• resolve is bounded at 10 seconds per server; a timeout or throw omits that server for the run without failing static MCP.
• Resolved connections are revalidated at most every 5 minutes per requester: rotation rebuilds the transport with fresh credentials, and a null result revokes it (the cached runtime is disposed even mid-session). A revoked or rotated credential can therefore stay in use for up to 5 minutes.
• Resolved headers are never logged or persisted; core keeps only an ephemeral in-memory keyed digest (process-local HMAC) to detect credential rotation, and registers resolved header/URL credential values with the log/debug-capture redaction registry.
• Requester-scoped servers do not mint MCP App views: a view outlives the requester-authenticated run and the gateway view boundary has no requester identity, so app previews stay fail-closed for these servers. Tool results are unaffected.
• Static servers without a resolver keep the existing session-scoped lifecycle.
• Harness delivery rule: requester-scoped servers never enter harness-native MCP client config (Codex thread mcp_servers, CLI -c mcp_servers=…, or any other session-shared MCP projection). Harnesses deliver them as run-scoped tools instead:
• Embedded runner: session MCP runtime + bundle tools (static + scoped).
• Codex app-server: dynamic tools via materializeRequesterScopedMcpToolsForHarnessRun (scoped-only; static servers stay on Codex's native MCP client).
• Scoped tool specs are session-stable after the first successful resolve in that session, so shared-thread harnesses (Codex) do not rotate threads when senders change. Before any requester resolves, no scoped specs are advertised.
• Unauthenticated requesters on a shared-thread harness still see the advertised scoped tools; calling one returns a clean not-connected tool error for that requester. OpenClaw never falls back to another requester's credentials.

Memory prompt supplement builders receive optional agentId, agentSessionKey, and sandboxed context. Memory corpus supplement search and get calls receive optional agentId and sandboxed context. Plugins with agent-owned storage should resolve that storage for each call instead of capturing one global path during registration. If an agent id is required but missing in a multi-agent operation, fail closed rather than choosing an arbitrary agent.

Use registerMemoryPromptPreparation(...) when prompt text depends on async plugin state. The callback runs once before each full agent prompt and receives the same tool, agent, session, and sandbox context as synchronous memory prompt builders. Validate the current storage-owner instance before loading persisted state, then return only lines for that run. OpenClaw freezes those lines and hands the immutable result to synchronous prompt assembly. Keep persistence, atomic replacement, and owner-removal deletion inside the owning plugin; do not poll or read files from a prompt builder.

Telegram interactive handlers can return { submitText } to route text through Telegram's normal inbound agent path after the handler succeeds. OpenClaw keeps the callback button when inbound policy skips the text or processing fails, so the user can retry after the blocking condition changes. This result field is Telegram-specific; other channels keep their own interactive result contracts.