@modelcontextprotocol/client

Upgrade notes

• One server per request. Server.connect() now rejects while the instance is already connected, and a stateless Streamable HTTP transport handles one request. Create the McpServer and the transport inside the request handler (or in the createMcpHandler factory) instead of sharing one instance across requests. Creating a server is cheap since #2889.

• Redirects stay on the same origin. The HTTP client transports now follow a redirect only when it stays on the same origin (same scheme, host and port; http to https on the same host is allowed). A deployment whose endpoint redirects to another host or port either configures the final URL or sets redirectPolicy: 'follow' on the transport. In browsers, a redirected request fails unless that option is set.

• New options, both off unless you set them. maxToolInputElements on McpServer limits the number of array elements and object members in a tool call's arguments. expectedResource on requireBearerAuth / verifyBearerToken accepts only tokens issued for this server (the token's audience); with Express, upgrade @modelcontextprotocol/express together with @modelcontextprotocol/server. (#2926, #2929)

• prompts/get without arguments is validated as {}, as tools/call already is. A top-level .optional() or .default(...) on a prompt's argsSchema no longer sees undefined.

• The client requires eventsource-parser 3.0.8 or later. It cuts the time and memory needed to receive a large message sent as a single SSE event: in our test a 100 MB tool result went from about a minute and 1.7 GB of peak memory to under a second and about 0.5 GB.

New

• validateOriginHeader and the allowedOrigins option of the Express, Fastify and Hono app helpers accept <scheme>://* entries such as moz-extension://*, so a server can admit MCP clients that run as a browser extension.

• tasks/get and tasks/cancel of the Tasks extension can be served and called on a 2026-07-28 connection.

Fixes

• A large message received as a single SSE event over Streamable HTTP is fast again: a 50 MB tool result that took about 13 seconds arrives in under a second.

• prompts/get without arguments no longer fails when every argument of the prompt is optional.

• SSEClientTransport refreshes once after a 401 on connect instead of retrying without limit. (#2905, #2934)

• registerTool no longer converts tool schemas up front, so a server built per request stops converting every tool on every request.

• hono is a regular dependency of @modelcontextprotocol/node, so installs with strict peer-dependency checking no longer fail.

• A server/discover probe answered with an unusable 2xx reply now says so instead of reading like a network failure.

• McpServer.registerPrompt() types the callback correctly when no argsSchema is given.

• The license field of the package manifests is Apache-2.0.

Upgrade notes

• deprecated

  Pass expectedIssuer to the machine-to-machine OAuth providers. Constructing ClientCredentialsProvider, PrivateKeyJwtProvider, StaticPrivateKeyJwtProvider or CrossAppAccessProvider without it is deprecated and logs a warning. Set it to the issuer of the authorization server the credentials were registered with.

• fetchToken() checks which authorization server the client information belongs to. It throws AuthorizationServerMismatchError before sending anything when the provider's client information is bound to a different authorization server.

• List calls return the whole list. listTools(), listPrompts(), listResources() and listResourceTemplates() called without a cursor now follow nextCursor until the server stops sending one. listMaxPages still caps the walk.

Fixes

• CommonJS TypeScript projects type-check again: the jose types used by the DPoP API are inlined into the declaration files (regression in 2.1.0).

• Client.listen() no longer lets a rejection escape as a process-level unhandled rejection, and no longer hangs when a send never settles.

• _meta is preserved on input_required results.

• Hostnames ending in .localhost count as loopback for OAuth token endpoints, so host-based multi-tenant local setups work.

• OAuthTokensSchema and OAuthClientInformationSchema accept the optional issuer field, so a provider that reads its storage back through them keeps it.

• createMcpHandler no longer overflows the stack when the factory returns the same server instance for more than one request. Returning a fresh instance per request is still required.

• createMcpHandler ends a subscriptions/listen stream right after the acknowledgement when it honored none of the requested notification types.

• Sending a notification on a closed connection no longer produces a briefly unhandled promise rejection.

• The v1-to-v2 codemod keeps a file's leading comment block and directives above the rewritten imports.

• Corrected citations in the JSDoc for OAuth token endpoints and for the registerClient deprecation notice. (#2768, #2729)

Per-package changelogs

client · server · core · server-legacy · codemod

Minor Changes

• Add DPoP (RFC 9449 / SEP-1932) sender-constrained access token support to the client.

  • Opt in by implementing OAuthClientProvider.dpop() returning a DpopSession (new, along with generateDpopKeyPair, accessTokenHash, isDpopNonceChallenge). auth() / exchangeAuthorization / refreshAuthorization / fetchToken then sign a DPoP proof into token requests (retrying once on an authorization-server use_dpop_nonce challenge, with client authentication re-applied per attempt), and StreamableHTTPClientTransport, SSEClientTransport and withOAuth present a token_type: "DPoP" access token as Authorization: DPoP <token> plus a fresh per-request proof, retry a resource-server use_dpop_nonce challenge once, and pick up a DPoP-Nonce delivered on any response. Tokens the AS issued as Bearer are still presented as Bearer.

  • DPoP is applied at the fetch layer: the transports wrap their resource-server fetch (including a caller-supplied fetch / eventSourceInit.fetch) with the new withDpopFromProvider(provider) middleware, so proofs are always bound to the request actually sent. withDpop(session, getToken) is exported for callers that manage tokens themselves (e.g. alongside a minimal AuthProvider); the AuthProvider interface itself is unchanged.

  • auth() now recovers from invalid_dpop_proof on refresh (e.g. a refresh token bound to a key that is no longer held) by discarding the tokens and re-authorizing, like invalid_grant. OAuthErrorCode gains InvalidDpopProof and UseDpopNonce; extractWWWAuthenticateParams recognizes the DPoP challenge scheme; OAuthMetadataSchema gains dpop_signing_alg_values_supported.

Patch Changes

• deprecated

  SdkError and SdkHttpError accept standard ErrorOptions as an optional fourth constructor argument and forward it to Error, so a wrapped error is reachable through the standard Error.cause chain. Version-negotiation probe failures (SdkErrorCode.EraNegotiationFailed) now use it: the underlying TypeError: fetch failed and the DNS or socket error beneath it surface via error.cause, so pino, Sentry, and util.inspect render ENOTFOUND / ECONNREFUSED / ETIMEDOUT instead of stopping at the SdkError (#2657). The previous error.data.cause slot is still populated for compatibility but is deprecated and slated for removal; read error.cause instead.

• Treat request id 0 as a real id. Two guards tested a RequestId for truthiness, so the legal JSON-RPC ids 0 and '' were read as absent. Id 0 is not a corner case: the outbound request counter is zero-based, so it is the first id every peer assigns, which on the server→client leg is the first sampling/createMessage, elicitation/create, or roots/list a server sends.

  • notifications/cancelled carrying id 0 was ignored, and the in-flight handler ran to completion with its AbortSignal never fired.

  • A notification sent with relatedRequestId: 0 wrongly passed the debounce gate (for methods opted into debouncedNotificationMethods). Because the pending set is keyed by method alone, a second such notification in the same tick was silently dropped rather than sent.

  Absent is now the only value that means "no id".

• On Windows, stdio servers spawned by StdioClientTransport now also inherit COMSPEC, PATHEXT, PROGRAMDATA, PROGRAMFILES(X86), PROGRAMW6432, and WINDIR (added to DEFAULT_INHERITED_ENV_VARS). Programs a server launches can depend on them: PowerShell finds no native executables without PATHEXT, and Windows OpenSSH exits 255 without ProgramData.

• Stop sending notifications/cancelled for the initialize handshake. The spec is explicit that a client MUST NOT attempt to cancel its initialize request, but the outbound cancel path fired for any in-flight request: aborting the AbortSignal passed to connect(), or letting the handshake hit its timeout, put a forbidden cancellation on the wire naming the initialize request id.

  The local behaviour is unchanged — the caller's promise still rejects with the same abort/timeout error, and connect() still tears the connection down. Only the wire notification is suppressed. Every other method keeps the existing cancellation path.

• StreamableHTTPClientTransport and SSEClientTransport now give their transport-managed headers precedence over same-named entries in requestInit.headers: Authorization when authProvider yields a token, mcp-protocol-version, and (Streamable HTTP) mcp-session-id. Header names compare case-insensitively and every HeadersInit form is covered (plain object, tuple array, Headers instance). Previously the caller-supplied value won, so a static Authorization placeholder (e.g. an env-var API key) kept overriding the OAuth token even after the provider obtained one and the fallback-to-OAuth flow never completed; a Headers instance or lowercase key produced a combined Bearer <fresh>, Bearer <stale> value instead. A configured Authorization is still sent while the provider has no token, and other configured headers pass through unchanged. Closes #2208.

• Preserve the exact OAuth resource indicator from protected resource metadata when building authorization and token requests. Previously a pathless resource such as https://example.com was normalized to https://example.com/ via URL.href, which breaks authorization servers that require the resource parameter to match the published value exactly (Microsoft Entra ID rejects it with AADSTS9010010). The exported OAuth helpers (startAuthorization, exchangeAuthorization, refreshAuthorization, fetchToken, executeTokenRequest) now also accept a string for resource; selectResourceURL still returns a URL, and a provider's validateResourceURL result is used unchanged. Fixes #1968.

• Let saveTokens failures surface after a successful token refresh. In auth(), one try
  wrapped both refreshAuthorization() and the provider.saveTokens() that persists its
  result, and the catch deliberately swallows anything that is not an OAuthError — plus
  ServerError — so that a failed refresh falls through to a fresh authorization request.
  A persistence error thrown by the provider landed in that same branch: it was discarded
  with no log and no rethrow, and auth() continued to startAuthorization() and returned
  'REDIRECT'.

  Against an authorization server that rotates refresh tokens (the OAuth 2.1 default, and
  Keycloak's) this loses credentials rather than merely hiding an error. The exchange has
  already succeeded server-side, so the old refresh token is invalidated at the moment the
  new one is issued; dropping the new token set leaves nothing usable on either side. On a
  headless or CLI client, where redirectToAuthorization is typically a no-op, the fallthrough
  is silent and the client is left with stale tokens and no indication of why.

  The try/catch now covers only refreshAuthorization(). Persisting the result happens
  after it, on an unguarded path, so a provider's I/O error propagates to the caller.

  Refresh-request failures keep their existing control flow exactly: a ServerError or an
  unknown error still falls through to a new authorization flow, a non-ServerError
OAuthError is still rethrown, and InsecureTokenEndpointError is still surfaced. The
  SEP-2352 issuer stamp written with the refreshed tokens is unchanged.

  Those fallbacks no longer happen in silence, though. Both routes to an unexplained
  re-authorization now emit a console.warn naming the cause: the in-place fallthrough in
  the refresh block, and auth()'s outer recovery for invalid_grant, invalid_client,
  and unauthorized_client, which discards stored credentials and retries. The second one
  matters most in practice — an expired, revoked, or rotation-reuse-detected refresh token
  is reported as invalid_grant, which is precisely the state a dropped token set leaves
  behind for the next call.

  Consumers whose OAuthClientProvider.saveTokens can reject should note that auth() may
  now reject where it previously returned 'REDIRECT' — that rejection is the failure that
  was being discarded.

• Emit and validate the Mcp-Name header for tasks requests per SEP-2663's Streamable HTTP binding: the client transport now mirrors params.taskId into Mcp-Name on tasks/get / tasks/update / tasks/cancel (previously omitted, causing conforming servers to reject every task poll with -32020 HeaderMismatch), and the server-side standard-header validation cross-checks it via the same shared MCP_NAME_HEADER_SOURCE table.

  On the server, createMcpHandler now answers a modern (2026-07-28) tasks/get / tasks/update / tasks/cancel POST that omits Mcp-Name, or whose header disagrees with params.taskId, with 400 / -32020 (HeaderMismatch) at the standard-header-validation rung, the same treatment tools/call / prompts/get / resources/read already get. Legacy-era (2025-11-25) tasks traffic is unaffected. Clients built with this SDK release send the header; hand-rolled clients that omitted it must add it.

Minor Changes

• Export the Protocol base class and mergeCapabilities from the @modelcontextprotocol/client and @modelcontextprotocol/server package roots, restoring the v1 import for consumers that subclass Protocol (e.g. the MCP Apps SDK). The client and server packages each bundle their own compiled copy of the class, so import it from one package consistently within a process.

  The codemod now rewrites Protocol and mergeCapabilities imports from shared/protocol.js to the client or server package root, like the module's other symbols, instead of dropping them with an action-required marker.

• ConnectOptions.prior accepts a cached era verdict — the new exported type PriorDiscovery. { kind: 'modern', discover } adopts a previously obtained DiscoverResult with zero round trips; { kind: 'legacy' } skips the server/discover probe and runs the plain initialize handshake directly, for servers known out-of-band to be legacy — without pinning the client to mode: 'legacy': stop supplying the verdict and connect() falls back to the configured versionNegotiation mode (under 'auto', it re-probes and rediscovers an upgraded server). Freshness is the supplying host's responsibility — a stale legacy verdict succeeds silently against an upgraded server, so hosts must date cached legacy verdicts in their own storage and stop supplying them past their policy horizon. Persisted-blob plumbing is hardened: prior: null is treated as absent, the modern arm's discover payload is schema-validated before any connection state changes, and an unrecognized shape rejects with a typed SdkError(EraNegotiationFailed) instead of a TypeError.

• The response cache now stores results as JSON-serialized documents (serialize on write, parse on read) instead of live object graphs isolated with structuredClone. Same mutation isolation, but no dependency on the structuredClone global — whose absence (jest+jsdom, Node < 17) previously made every cache write throw into the store-error swallow, silently disabling caching and output-schema lookups for the session. A value without a JSON representation now fails the write loudly to the error sink, and an undecodable document in an external store is reported, dropped, and read as a miss.

  Migration for custom ResponseCacheStore implementations: CacheEntry.value (and the set() entry value) is now string — persist and return it verbatim, JSON.parse to inspect. Entries persisted by a previous SDK version fail decode once (reported, dropped) and are rewritten on the next fetch.

• Move the schema source modules (spec schemas, OAuth schemas, protocol constants) into @modelcontextprotocol/core and resolve them from there as a regular runtime dependency instead of bundling a private copy into each package. An application importing more than one of the packages now evaluates a single shared schema graph with shared object identity. @modelcontextprotocol/core gains a ./internal subpath (SDK-internal contract; may change in any release) and the four packages now version together.

• Align the 2026-07-28 wire with the final revision (spec PR #3002): serverInfo moves from the DiscoverResult body to the result _meta, and the per-request envelope's clientInfo demotes from required to SHOULD.

  Before this change the SDK shipped the pre-#3002 shape in both directions: the client hard-rejected a conforming server's DiscoverResult (missing body serverInfo failed parse, so the probe misclassified the server as legacy and attempted an initialize handshake against it — a hard connect failure against a modern-only server such as go-sdk v1.7.0-pre.3), and the server rejected conforming clients that omit clientInfo.

  Now:

  • The 2026 wire schemas are the final revision exactly: no body serverInfo on DiscoverResult, envelope clientInfo optional (a present-but-malformed value still fails validation).

  • Servers stamp _meta['io.modelcontextprotocol/serverInfo'] on every 2026-era response (spec SHOULD; a handler-authored value wins, the 2025-era wire is untouched). This includes the entry-built subscriptions/listen graceful-close results — the spec's SubscriptionsListenResultMeta extends ResultMetaObject.

  • Clients keep sending clientInfo and read server identity from the discover result's _meta only. A server that stamps no identity is anonymous: getServerVersion() is undefined and the response cache partitions under a per-connection surrogate. A malformed _meta serverInfo value is treated as absent on receive (the spec marks the field self-reported, unverified, and display-only).

  • Breaking type changes: DiscoverResult no longer declares serverInfo; RequestMetaEnvelope's clientInfo is optional. New public constant SERVER_INFO_META_KEY ('io.modelcontextprotocol/serverInfo').

• Add preloadSchemas(), an explicit opt-in to eager wire-schema construction, and call it automatically in the Cloudflare Workers builds. The wire schemas are built lazily by default, which is the right trade on process-per-invocation runtimes — but on isolate platforms that bill request CPU while module evaluation runs during isolate warm-up, laziness moves construction into the first request each fresh isolate serves. Calling preloadSchemas() at module scope (it is synchronous and idempotent) moves that one-time cost back to module evaluation; the packages' workerd export condition now does this automatically, while the Node and browser builds stay lazy. The server package gains a dedicated browser shim for this (its browser condition previously reused the workerd shim), so browser bundles keep lazy construction.

Patch Changes

• First beta release of SDK v2 with support for the MCP 2026-07-28 specification
  revision. See the migration guides for upgrading from v1
  (docs/migration/upgrade-to-v2.md) and adopting the 2026-07-28 revision
  (docs/migration/support-2026-07-28.md).

• Restore the v1 parse tolerance for CallToolResult.content: an inbound legacy-era tools/call result without content defaults to `` instead of failing validation. Deployed servers — accepted by SDK v1 for years — return structuredContent-only (or otherwise content-less) results, and the strict parse turned every such call into an INVALID_RESULT error before application code could run.

  The silent-empty-success hazard the strictness guarded is preserved where it matters: the 2025 era's wire-seam schema refuses to default content for a body carrying another result family's vocabulary (task, inputRequests, requestState — the era is frozen, so the list is complete), and the 2026-era wire schemas stay strict — modern-revision servers have no legacy excuse. Task interop through an explicit result schema is untouched (including bodies that also stamp a foreign resultType), and the server-side authoring normalization refuses the same foreign-family vocabulary.

  Server-side authoring is era-independent: a handler result without content (dynamic/JS callers — the TypeScript surface requires it) is normalized to content: before era validation on every leg, reaching the wire spec-valid.

  Conscious call: the nested sampling ToolResultContentSchema stays spec-strict — v1 had defaulted its content too, but it is params-side (tool results a caller authors into a sampling message), deliberately not restored.

• Fix the CommonJS validators/ajv subpath so reading the exported Ajv class no longer throws ReferenceError: import_ajv is not defined. The subpath now re-exports the bundled provider's concrete Ajv value in CJS output, matching the existing ESM behavior.

• Ship CommonJS builds alongside ESM. Each package now emits both .mjs/.d.mts
  and .cjs/.d.cts (via tsdown format: ['esm', 'cjs']), and its exports map
  adds a require condition so require('@modelcontextprotocol/…') works from
  CommonJS consumers. Output extensions are normalized across all packages
  (@modelcontextprotocol/core moves from .js/.d.ts to .mjs/.d.mts); the
  public import paths are unchanged.

• POSTs whose Content-Type media type is not application/json are now
  rejected with 415 Unsupported Media Type; the header is parsed instead of
  substring-matched. Previously any value merely containing the substring
  passed the check (for example text/plain; a=application/json), case
  variants were wrongly rejected, and the 2026-07-28 entry did not inspect
  Content-Type at all — requests with a missing or non-JSON header that used
  to be served on that path now also answer 415. Values with parameters
  (application/json; charset=utf-8, including malformed parameter sections
  like application/json;) continue to work. SDK clients always send the
  correct header and are unaffected.

  The new isJsonContentType(header) helper is exported for transport and
  framework-adapter authors — custom entries composing the exported building
  blocks (classifyInboundRequest, PerRequestHTTPServerTransport) must apply
  it themselves. The hono adapter's JSON body pre-parse and the client's
  response dispatch now use the same parsed-media-type comparison.

• instanceof on the SDK error classes (ProtocolError and its typed subclasses, SdkError/SdkHttpError, OAuthError, and the client's SseError, UnauthorizedError, and OAuth-client-flow error family — OAuthClientFlowError and its subclasses) now works across separately bundled copies of the SDK. The classes match by a stable brand (via Symbol.hasInstance and a registry symbol) instead of prototype identity, so a process that uses both @modelcontextprotocol/client and @modelcontextprotocol/server - a gateway, host, or in-process test - can check errors constructed by either package against the class re-exported by the other. Ordinary prototype-based instanceof is preserved as a fallback; user-defined subclasses keep plain prototype semantics. Notes: cross-bundle matching requires both copies to be at or after this release; brands assert identity, not field shape, across versions - keep reading fields defensively. As a side effect, a foreign-bundle SdkError used as an abort reason is now rethrown as-is instead of being wrapped as a RequestTimeout. Branded hierarchies additionally expose an explicit static guard, X.isInstance(value), that reads the same brand and narrows in TypeScript — an alternative for codebases that prefer predicate-style checks over instanceof. Also: UnauthorizedError now sets error.name to 'UnauthorizedError' (previously 'Error'), and per-package conformance tests enforce that every exported error class participates in branding. Version-negotiation probing now recognizes UnauthorizedError (previously a dead name-string check) and propagates it unchanged, so connect() on an auth-gated server rejects with the original UnauthorizedError (previously wrapped as the cause of an SdkError(EraNegotiationFailed)) — run finishAuth() and reconnect, and the retry probes with the token.

• The Streamable HTTP client transport no longer attaches a session ID to a POST containing an initialize request — a new session starts "without a session ID attached" (2025-11-25 transports §Session Management) — and it only captures the mcp-session-id response header from a successful initialize response, since the spec assigns the session ID "at initialization time … on the HTTP response containing the InitializeResult". Previously the transport stored the header from any response, so a legacy server answering a protocol-version probe with an error that happened to carry a session ID would poison the fallback initialize, which then went out with a session ID it should not have had. A stale session ID from a previous connection is likewise no longer leaked onto the initialize handshake, and a successful initialize response that carries no session ID now clears any stale ID the transport was holding — clients include only an ID "returned by the server during initialization", so an ID the server never returned this session is outside the session model. Ignoring mcp-session-id headers mid-session is the complement of the spec's one actual rotation mechanism: a server that wants a new session terminates the old one (it "MAY terminate the session at any time") and answers 404, after which the client "MUST start a new session by sending a new InitializeRequest without a session ID attached". Rotation exists as session replacement via 404 + re-initialize, never as a header swap on a live session, so a server that rotates per the spec's own flow is handled correctly by this transport.

• Construct the default Ajv validation engine lazily on first validation. Creating a Client or Server no longer pays the ajv + ajv-formats instantiation cost at startup when no JSON Schema validation ever runs.

• Build protocol-revision wire schemas lazily on first validation instead of at import. Each revision's schema set is now constructed by a module-level memoized factory, so importing the client or server package no longer pays the construction cost of both frozen wire-schema graphs up front. Method membership in the revision registries stays static, the schemas themselves are unchanged, and registry lookups keep returning reference-identical schema objects.

• The version-negotiation probe no longer misclassifies auth-protected or
  failing servers as legacy. Auth status is never era evidence: a 401 or 403
  rejection of the server/discover probe now surfaces as a typed
  authorization failure — an SdkHttpError with code ClientHttpAuthentication
  (401) or ClientHttpForbidden (403), carrying the HTTP status, reason
  phrase, and response text — instead of triggering the legacy initialize
  fallback (which put a doomed initialize on the wire) or, under pin mode,
  the false "server did not offer pinned protocol version" diagnostic. The
  codes are deliberately not EraNegotiationFailed, so era-recovery flows
  keyed on that code cannot persist a verdict for an unauthorized exchange. A
  5xx rejecting the probe is a server failure and now also rejects typed
  (SdkHttpError(EraNegotiationFailed)) instead of demoting a mid-deploy
  modern server to legacy — the legacy fallback now fires only on the 4xx
  shapes the spec licenses.

  With an authProvider, a 401 (and a 403 insufficient_scope challenge) runs the transport's auth flow first — a plain 403 rejects the same as without a provider — and whatever
  escapes it propagates unchanged, identity intact: the HTTP transports stamp
  errors at their auth seams (the token() read, onUnauthorized including
  custom callbacks, the 403 step-up flow, and their own auth-failure
  constructions), so UnauthorizedError for finishAuth(), the flow's typed
  failures (OAuthError, InsufficientScopeError, the
  401-after-re-authentication diagnostic), and even an untyped TypeError
  thrown inside the flow all reach the caller as thrown — never rewrapped,
  never consumed by the probe's browser CORS heuristic as legacy-era evidence.

• Probe stdio servers on a disposable sibling process. Some stdio servers exit on any pre-initialize request (servers built on the official Rust SDK, rmcp, behave this way), so under versionNegotiation: { mode: 'auto' } the server/discover probe previously killed the server and connect() hard-failed. The probe now runs on a short-lived sibling spawned from the same parameters — its stderr is discarded and it is reaped once the era is known — and the caller's transport spawns exactly once, afterwards: a legacy verdict connects with the plain initialize handshake (byte-identical to mode: 'legacy'), a modern verdict is adopted directly, and the session wire never carries server/discover. Closing the caller's transport during the probe aborts connect() with the typed SdkError(EraNegotiationFailed) and the session child is never spawned. On HTTP — and on custom stdio-shaped transports, which probe in place — a mid-probe connection close keeps rejecting with the typed connect error, now naming the close in pin-mode and modern-only diagnostics.

• Version negotiation no longer discards transport handlers the caller set before connect(). The probe window now saves any pre-set onmessage/onerror/onclose, forwards error and close events to them while the probe is in flight, and restores them when the window closes — so Protocol.connect() chains them exactly as it does on a plain connect. Previously, connecting with versionNegotiation silently cleared pre-set handlers (e.g. an onerror used to detect session-expiry auth failures), leaving them permanently detached for the life of the connection.

• Stop advertising validator provider classes from the root client/server type declarations. The provider classes remain available from the explicit validator subpaths.

• The default validator now honors declared 2019-09 and draft-07/06 dialects instead of rejecting them: a schema stamped "$schema": "http://json-schema.org/draft-07/schema#" (zod-to-json-schema's default output) validates with draft-07 semantics, and a 2019-09 stamp (zod-to-json-schema's 2019-09/openAi targets) with 2019-09 semantics, on both the Ajv and Cloudflare Workers providers (with known engine differences documented in the migration guide). Schemas with no $schema still validate as 2020-12, and unknown dialects still produce the typed error (now listing the supported dialects: 2020-12, 2019-09, draft-07, draft-06).