fix(client): support re-initializing after the server ends the session - #1018

* fix(client): don't write transport and trace headers into the caller's header map (mark3labs#1028)

* fix(client): don't write transport and trace headers into the caller's header map

* test(client): cover caller header isolation on streamable HTTP and tracing

* fix(server): preserve typed JSON-RPC handler errors (mark3labs#1027)

Co-authored-by: Fedor Bushlia <fedorbush@yandex-team.ru>

* fix(server): send empty arrays instead of null in results (mark3labs#989)

tools, prompts, tasks, completion values, prompt messages and resource
contents are required arrays, but a nil slice reaching them went out as
null: from a tool or prompt filter that hides everything, from tasks/list
before any task has run, and from completion providers and prompt or
resource handlers that return nil. The schema rejects null there, and so
do clients that validate responses.

listByPagination now returns an empty slice instead of nil, which covers
every list result and supersedes the resources/list guard from mark3labs#665.
completion/complete, prompts/get and resources/read fill in an empty
array the same way. The prompt handler's result is copied rather than
modified, and a result asking for input is left as it is.

* fix(server): reject missing client capabilities with invalid params (mark3labs#990)

A request declaring protocol version 2026-07-28 without
io.modelcontextprotocol/clientCapabilities in _meta was answered with
-32021 MissingRequiredClientCapability, and one with a malformed value
with -32020 HeaderMismatch. The spec treats a request missing a
required _meta field as malformed and requires -32602 Invalid params
(400 on HTTP, which is unchanged). -32021 is for a request that needs
a capability the client did not declare, and carries
data.requiredCapabilities; -32020 is for HTTP header mismatches.

* fix(server): send URLElicitationRequiredError from handlers as -32042 (mark3labs#992)

* fix(server): send URLElicitationRequiredError from handlers as -32042

The elicitation docs and example have a tool return
mcp.URLElicitationRequiredError to tell the client to complete a URL
elicitation first. Handler errors were all reported as INTERNAL_ERROR
with only the message, though, so the client got -32603 without the
elicitations, and mcp-go's own client could not decode it as a
URLElicitationRequiredError.

For clients before protocol version 2026-07-28, report it as -32042
with data.elicitations, as 2025-11-25 defines. 2026-07-28 reserves the
code and asks for URL elicitations through multi round-trip requests,
so modern clients still get an internal error. Only this error type is
recognized; any other handler error is an internal error as before.

* fix(server): also recognise a pointer to URLElicitationRequiredError

URLElicitationRequiredError has value receivers, so a handler can return
a pointer to one as well. errors.As with a value target does not match
the pointer, and the client got an internal error without the
elicitations.

* fix(streamable-http): deliver notifications on modern subscriptions/listen streams (mark3labs#993)

A 2026-07-28 client opens subscriptions/listen to receive notifications
such as tools/list_changed. Over Streamable HTTP each modern request is
served by an ephemeral, unregistered session, and broadcasts only reach
registered sessions, so a listen stream got its acknowledgement and then
nothing, while the acknowledgement promised the requested types.

Make the listen stream's session reachable by broadcasts while the
request runs, through a registry of its own so that no session hooks
fire. Delivering to it needs two more things the spec asks of a
subscription stream: the HTTP session now records the filter
subscriptions/listen sets, and until that happens (and after it is
cleared) it receives nothing; and every notification on the stream
carries the subscription ID in _meta.

* fix(mcp): parse an input_required tool result that has no content (mark3labs#994)

A tools/call that needs more input is answered, in protocol version
2026-07-28, with an InputRequiredResult: resultType, inputRequests and
requestState, and no content. ParseCallToolResult rejected any result
without content before looking at resultType, so against servers that
send that shape (the other SDKs, and the conformance reference server)
CallTool failed with "content is missing" and the multi round-trip retry
never ran. mcp-go's own server sends "content": [] alongside, which is
why this did not show between mcp-go peers.

Accept a missing content when resultType is input_required, as
ParseReadResourceResult already does for contents. A complete result
without content is still rejected.

* fix(client): exclude tools with invalid x-mcp-header annotations (mark3labs#996)

SEP-2243 requires a client using Streamable HTTP to reject a tool whose
x-mcp-header annotations break its constraints (empty, non-primitive or
number type, duplicate names, characters outside the header token set)
by leaving it out of the tools/list result, and SHOULD log a warning.
The client listed and cached every tool, so it would call such a tool
and mirror headers from it.

ListToolsByPage now drops those tools on modern HTTP connections,
including transports wrapped for logging, and logs the tool name and
reason. A definition cached from an earlier listing is forgotten, so
the tool is not called with headers mirrored from it. Other transports
may ignore the annotations, and connections before 2026-07-28 do not
mirror them, so both are unchanged.

* test(client): remove the mock server's build cache after building it (mark3labs#998)

compileTestServer gives each build of the mock stdio server a fresh
GOCACHE, which mark3labs#241 added to work around a flaky linker error in
parallel builds, but never removes it. Each build leaves about 120 MB in
the temp directory: one run of go test ./client/... left 16 of them,
1.9 GB.

Remove the cache once the build is done, and report a failure to create
it instead of quietly building in the shared cache.

* fix(server): report handlers that return no result as internal errors (mark3labs#999)

A handler returning nil, nil left the server with nothing to send:

- prompts/get dereferenced the nil result in the generated dispatcher
  and panicked. Over streamable HTTP the connection dropped with no
  response; with the in-process client the panic reached the caller.
- tools/call answered with a null result, which no client accepts.
- Run as a task, the tool completed with a nil result, and a later
  tasks/result panicked reading it.

Report an internal error for prompts and tools, and fail the task. The
checks run after any legacy multi round-trip retry, so a retry that
returns nil is caught too.

* fix(sse): send a recovered panic's error with the request's id (mark3labs#1000)

The SSE message handler recovers a panicking handler and queues an
INTERNAL_ERROR so the client doesn't hang, but it built the response
with a nil id. Clients match responses to pending requests by id; the
mcp-go SSE client treats an id-less message as a notification, so the
call still waited for its timeout. Use the id of the request that
panicked, as the stdio transport already does.

A notification gets no reply at all, so when its handler panics
nothing is queued. HandleMessage decides what is a notification the
same way, by the missing id.

* fix(mcp): accept an Mcp-Param header sent with an empty value (mark3labs#1001)

An empty string argument annotated with x-mcp-header is mirrored as a
header with an empty value; GenerateParamHeaders produces exactly that.
ValidateParamHeaders takes a getHeader func that returns "" for a
missing header, so it cannot tell the two apart and rejected the call
with "missing header for parameter". A 2026-07-28 client passing an
empty string could not call such a tool at all.

Add ValidateParamHeadersLookup, whose lookup also reports presence,
and use it in the server with http.Header.Values. ValidateParamHeaders
keeps its behavior and delegates to it. An empty header for an absent
argument is still ignored, as before.

* fix(stdio): recover handler panics on the read loop (mark3labs#1002)

The tool call workers and handleRequest recover a panicking handler and
answer with INTERNAL_ERROR, but two paths still call HandleMessage on the
read loop without that protection: notifications, and a tools/call run
synchronously because the worker queue is full. A panic in either
propagated out of Listen and ended the process, where every other
request just got an error.

Recover there too, with a handleMessage helper that handleRequest now
shares. A request still gets an INTERNAL_ERROR carrying its id; a
notification gets no response, as JSON-RPC requires.

* fix(oauth): reject authorization server metadata for another issuer (mark3labs#1004)

After protected resource metadata names an authorization server, the
client builds the well-known URLs from that issuer and uses the first
document it gets, whatever issuer the document declares. RFC 8414 §3.3
and OpenID Connect Discovery §4.3 say the issuer in the document must be
identical to the one the URL was built from, and the 2026-07-28 MCP
authorization spec repeats it: if they differ, the client must not use
the metadata. Otherwise a server could send the client to one
authorization server's endpoints under another's name.

Reject such metadata with an error instead of using it or falling back
to default endpoints. A final trailing slash is ignored, as in the
official Go SDK. OAuthConfig.SkipIssuerMetadataValidation turns the
check off for authorization servers known to publish a mismatched
issuer, like the TypeScript SDK's skipIssuerMetadataValidation.

Metadata from an explicit AuthServerMetadataURL, which the caller chose,
and from the legacy discovery without protected resource metadata is
left unchecked, so as not to break servers that rely on it.

* fix(oauth): authenticate at the token endpoint the way the server supports (mark3labs#1005)

A confidential client always sent its client_id and client_secret in the
token request body (client_secret_post), whatever the authorization
server said. A server that only accepts HTTP Basic (client_secret_basic)
rejected every code exchange and refresh, and the method returned by
dynamic client registration was ignored.

Pick the method per request, for the code exchange and the refresh:

1. OAuthConfig.TokenEndpointAuthMethod, new, for pre-registered or saved
   credentials on servers that support more than one method.
2. The token_endpoint_auth_method from the registration response
   (RFC 7591 §3.2.1). GetTokenEndpointAuthMethod returns it, so it can be
   saved with the client ID and secret.
3. HTTP Basic when the server's metadata lists client_secret_basic but
   not client_secret_post.
4. Otherwise the secret in the body, as before, which also covers servers
   that list nothing.

A client registered with "none" sends only its client_id. Basic
credentials are form-encoded first, as RFC 6749 §2.3.1 requires, and are
not repeated in the body.

* fix(client): resume an SSE response stream the server ends early (mark3labs#1010)

Before protocol version 2026-07-28, a server that answers a POST with an
SSE stream may end the response once it has sent an event ID, before
the JSON-RPC response, and the client should poll by reconnecting. It
must wait the time given by the server's retry field, and it resumes
with a GET that carries Last-Event-ID (SEP-1699). The client ignored
the id and retry fields and failed the call with "unexpected nil
response".

Record the last dispatched event ID and the retry time while reading
the stream. When a POST stream ends without the response, on a session
before 2026-07-28 and after an event ID, wait the retry time (one
second without one, and at least 10ms) and reopen the stream with a GET
carrying Last-Event-ID, until the response arrives or the context ends.
Each resumed connection is closed as soon as it ends. A GET the server
refuses fails the call with the reason, and a stream that breaks off
instead of ending still fails it, as before. Protocol version 2026-07-28
removed this mechanism, so modern sessions are unchanged.

* fix(client): don't answer a server's request with null (mark3labs#1011)

The client answers sampling, elicitation and roots requests with what
its handler returns. A handler that returned no result was answered
with a null result, and a roots handler with no roots, returning
&mcp.ListRootsResult{}, with "roots": null. The schema requires an
object and an array, so servers that validate the answer reject it; the
TypeScript SDK fails with "expected object, received null" and
"expected array, received null". mcp-go's own Streamable HTTP server
takes a null result for an empty response, answers 202 and drops it, so
its RequestSampling or RequestElicitation waits until its context ends.

Report a missing result as an error, as the multi round-trip path
already does, and send an empty roots array, there too. The in-process
client passed a nil result straight to server code, which now gets the
same error.

* fix(server): run a session's own tool when it is called as a task (mark3labs#1012)

handleToolCall looks a tool up in the session's tools before the
server's, and sends a call with task params to handleTaskAugmentedToolCall.
That looked the tool up again in the server's tools only. A session tool
that supports tasks was reported as not found when called as one. A
session tool that shadows a server tool of the same name ran the server
tool's handler instead when the server tool supports tasks too, and
failed with -32601 when it doesn't.

Look up the session's tools first there too.

* fix(oauth): share one token refresh among concurrent callers (mark3labs#1013)

getValidToken refreshed the token for every caller that found it
expired. Callers that did so at the same time, such as requests a
client sends in parallel after the token ran out, all sent the same
refresh token. The MCP spec requires authorization servers to rotate
refresh tokens for public clients, so a server accepts only the first
of those requests: every other caller got ErrOAuthAuthorizationRequired
and was sent back to authorize, although a new token was stored a
moment later. A server that detects the reuse may also revoke the
tokens the first request got.

Let callers share the refresh in flight, and start one only if none is.
The refresh runs on its own context with a timeout, not the context of
the caller that started it, so that a caller giving up after the server
spent the refresh token doesn't leave the others to send it again. It
reads the store first, so a caller that saw the expired token just
before another refresh finished uses the new token. Callers waiting on
a refresh stop when their context ends, and all of them get its result,
failure included, instead of retrying one after another.

* fix(oauth): discover the authorization server again after a failure (mark3labs#1014)

getServerMetadata runs discovery through a sync.Once and keeps whatever
it ends with, an error included. Discovery runs with the context of the
first caller that needs it, so when it ended with an error, because
that caller's context ended, a metadata request failed on the network,
or a configured AuthServerMetadataURL answered 5xx, every later call
got the same error back. Refreshing a token and starting authorization
both need the metadata, so the handler stayed unusable until a 401
carrying a resource metadata URL reset it, which a server that doesn't
send one never does.

Keep the metadata a discovery found, but not a failure: report it to
the callers that waited on that discovery, then let the next call
discover again, with no error left over from the last one.

* fix(oauth): send a token without a type as a bearer token (mark3labs#1015)

GetAuthorizationHeader builds the header from the token's type and
access token, normalizing any case of "bearer" to "Bearer". A token
with no type produced " <token>", which net/http sends as the bare
token, without the Bearer scheme servers expect. That happens when an
application puts a token in the store itself, a custom TokenStore
doesn't keep the type, or an authorization server leaves token_type out
of a response.

Send such a token as a bearer token, which is how MCP authorization
says access tokens are sent (Authorization: Bearer <access-token>).

* fix(client): cancel requests over stdio when their context ends (mark3labs#1016)

When a request's context ended before the response, the client gave up
waiting and returned the context's error, but told the server nothing.
Over stdio there is no per-request stream to close, so the server kept
working on the request, a tool call for example, for as long as it
took. Protocol version 2026-07-28 says a client on stdio must send
notifications/cancelled referencing the request to cancel it, and
earlier versions offer the same notification for it.

Send notifications/cancelled for a request whose context ends after it
was sent, on transports that aren't HTTP, once the connection is
initialized. It isn't sent for initialize or the server/discover probe,
which must not be cancelled, for task-augmented requests, which
tasks/cancel is for, or over HTTP, where 2026-07-28 makes the end of the
request's stream the signal and earlier versions keep their current
behaviour. A request whose context has already ended isn't sent at
all. The notification goes out from its own goroutine with a timeout,
so a peer that doesn't read can't hold up the caller.

* fix(sse): wait for room in the event queue instead of dropping a response (mark3labs#1017)

When a session's event queue was full, handleMessage dropped the response
to the message it had handled, and the client waited until its timeout. The
queue fills when events are produced faster than the client reads them, such
as a tool sending notifications in a loop. The error sent after a recovered
panic had the same fallback.

Both now wait for room, or for the session to end. handleSSE closes the
session's done channel on every exit, including a panic, so a waiting
response can't outlive its stream.

* fix(client): support re-initializing after the server ends the session (mark3labs#1018)

When the server answers 404 for the session, the transport returns
ErrSessionTerminated and the caller is expected to call Initialize again.
Two things got in the way. The GET stream stayed on the old session, or
stopped for good, so the client got no notifications for the new one.
And when the GET stream saw the 404 first, the transport dropped the
session ID and later requests went out without one, which some servers
reject or treat as the start of another session, so the caller never
learned that it had to re-initialize.

After a 404 for the session, requests, notifications and responses now
fail with ErrSessionTerminated until a request that starts a new
connection succeeds: initialize, or server/discover on 2026-07-28. The
GET stream waits for the new session and moves to it. A 404 for a GET
stream that never opened on the current session means the server
doesn't offer the stream: the listener stops, as before, but the
session goes on.

* fix(client): follow pagination cursors when listing tasks (mark3labs#1020)

* fix(client): follow pagination cursors when listing tasks

ListTasks issued one tasks/list request and returned that page. The other
list helpers follow NextCursor, and the server paginates tasks when a
limit is set, so every task after the first page was dropped.

ListTasksByPage keeps the single-page call. ListTasks now walks the
cursor the same way ListTools does.

* fix(client): stop ListTasks when a cursor repeats

A server that returns a cursor already requested would keep ListTasks
issuing tasks/list until the caller canceled the context.

* fix(mcp): omit empty completion context arguments (mark3labs#1024)

CompleteContext.Arguments had no omitempty, so a completion request without
context arguments was sent as "context":{"arguments":null}. The spec
defines arguments as an optional object, and servers that validate it, such
as those built on the TypeScript SDK, reject null with -32602.

Omit the field when it is empty, as GetPromptParams.Arguments already does.

* perf(server): touch a known session without allocating (mark3labs#1026)

With WithSessionIdleTTL set, touchSession runs on every request and
called sync.Map.LoadOrStore with a freshly allocated atomic.Int64, which
also boxed the string key: 2 allocs (24 B) per request even when the
session was already tracked.

Look the session up with Load first and fall back to LoadOrStore only on
a session's first touch. A known session now costs 0 allocs; a first
touch is unchanged. A new test pins 0 allocs on a known session.

BenchmarkTouchSessionKnown, AMD Ryzen 9 9950X, go1.27.0 windows/amd64,
default GC, -count 1 run six times per side, old and new alternated:

                       old            new
  sec/op               29.10n +- 1%   12.20n +- 1%  -58.09% (p=0.002 n=6)
  B/op                 24.00 +- 0%    0.00 +- 0%    -100.00% (p=0.002 n=6)
  allocs/op            2.000 +- 0%    0.000 +- 0%   -100.00% (p=0.002 n=6)

* fix: parse sampling content sent as an array (mark3labs#1029)

Since 2025-11-25 the content of a sampling message may be a single block or
an array of blocks. The client parsed the messages of a sampling request,
and the stdio and Streamable HTTP sessions parsed the result, only when the
content was a single object. An array, which is what a model returns when it
calls more than one tool, reached the handler or RequestSampling as []any of
maps.

Add mcp.ParseSamplingContent, which parses a single block into a Content
and an array into a []Content, and use it in all three places.

* fix(server): reject an unsupported MCP-Protocol-Version header (mark3labs#1030)

* fix(server): reject an unsupported MCP-Protocol-Version header

The Streamable HTTP transport requires a server that receives an invalid or
unsupported MCP-Protocol-Version header to respond with 400 Bad Request.
Requests from modern clients are already validated, but on the legacy path
the header was never read: a POST, GET or DELETE with an unknown version
such as 2025-01-01 was served as usual.

Reject such requests with 400 when the header names a version this SDK does
not implement. Initialize is not checked, since it negotiates the version,
and a missing header is still accepted for clients on 2025-03-26.

* fix(server): check the protocol version before accepting a response

A POST carrying a ping response or an empty response returned 202 before
the MCP-Protocol-Version check ran, so an unsupported header was accepted
on that path. Detect the era and run the check right after parsing the
body, ahead of those early returns.

* fix(mcp): keep structuredContent on tool_result content (mark3labs#1031)

* fix(mcp): keep structuredContent on tool_result content

The 2025-11-25 schema gives ToolResultContent an optional structuredContent
object next to content, like CallToolResult has. ToolResultContent had no
such field, so a server could not send it in a sampling request, and
UnmarshalContent and ParseContent dropped it when decoding one.

Add the field and read it in both decoders.

* fix(mcp): keep the original structuredContent bytes on tool_result

Decoding structuredContent into an any turns numbers into float64, so an
integer above 2^53 changed value on a decode and encode round trip.
CallToolResult avoids this with RawStructuredContent; do the same for
ToolResultContent and write the raw bytes back when marshaling.

* fix(client): reject repeated pagination cursors on list APIs (mark3labs#1033)

Servers that echo the same nextCursor would otherwise loop forever when
aggregating tools, resources, prompts, and related iterators.

* test(conformance): add official MCP server suite to CI (mark3labs#1038)

* test(conformance): add official MCP server suite to CI

* fix(conformance): harden suite validation

* fix(conformance): parse timestamp suffixes by format

---------

Co-authored-by: Omid Mirzaei <omidomirzaei@gmail.com>

* fix: restore SSE resume and protocol-aware handler errors

- resume SSE streams without starting a new session
- honor protocol-aware handler errors and preserve wrapped context

* fmt

* conformance: the line passes the notification scenarios upstream expects to fail

* conformance: accept an empty server baseline

---------

Co-authored-by: Serhii Zghama <20826225+serhiizghama@users.noreply.github.com>
Co-authored-by: FlameHost10 <147077735+FlameHost10@users.noreply.github.com>
Co-authored-by: Fedor Bushlia <fedorbush@yandex-team.ru>
Co-authored-by: Captain <42566883+po-et@users.noreply.github.com>
Co-authored-by: Omid <86771298+o-mid@users.noreply.github.com>
Co-authored-by: Hamed Yousefi <hdyousefi@gmail.com>
Co-authored-by: Peter Bednarčík <29061766+pbednarcik@users.noreply.github.com>
Co-authored-by: mika <spacexstarship01@outlook.com>
Co-authored-by: Omid Mirzaei <omidomirzaei@gmail.com>
Co-authored-by: Ed Zynda <ezynda3@gmail.com>