* [Capability] Do not announce an externally loaded registry as changed
`Registry` suppresses its `*ListChangedEvent`s while it is loading — `dispatch()`
returns early on the `loading` guard, so the elements a loader registers are the
registry's initial contents rather than a change to them.
`testListChangedEventsAreSuppressedDuringTheDeferredLoad` pins that.
The guard only covers `Registry::load()`, which needs the loader the constructor
took. A registry the caller built cannot be given one that way, so
`Builder::resolve()` loads it from the outside instead — by calling
`$chainLoader->load($registry)` directly, which never sets the guard. Every
element then dispatches on the way in.
With a notification bus configured that is not quiet. `PublishingEventDispatcher`
turns each event into a published notification, so every `build()` puts one
`list_changed` per element on the bus for a registry that did not change. Under
PHP-FPM, where the server is built per request and `Psr16NotificationBus` is
shared and persistent, every request broadcasts its whole element list to every
open `subscriptions/listen` stream and consumes the 256-entry backlog — after
which a reader that fell behind silently skips real notifications.
Split the guarded body of `load()` into `loadFrom(LoaderInterface $loader)` and
route the custom-registry branch through it.
`loadFrom()` takes only the `loading` guard, not the `loaded` bookkeeping, which
stays with `load()`. The loader it runs belongs to the caller, so it cannot stand
in for the one the registry was constructed with: marking the registry loaded
would retire a constructor loader that never ran, and would make a second
`loadFrom()` — one registry handed to two builders — a silent no-op. Keeping
`loaded` out of it also keeps `load()`'s promise that a transient failure is
retried on the next read, including when the failing loader reads the registry
during its own run.
`RegistryInterface` declares no `load()`, so this stays on the concrete
`Registry`, and a third-party implementation keeps the path it has today.
* Apply batched suggestions from code review
Co-authored-by: Christopher Hertel <mail@christopher-hertel.de>
* Cast variadic tool parameters element-by-element
SchemaGenerator::buildVariadicParameterSchema() already advertises a
variadic parameter (e.g. `int ...$scores`) as a JSON "array" schema, so
a spec-conformant client sends an array for it. But
ReferenceHandler::prepareArguments() never special-cased
ReflectionParameter::isVariadic(): it fell through to the regular
single-value branch and passed the whole array to castArgumentType(),
which tried to cast it as one scalar and failed - e.g. "Cannot cast
value to integer" for `int ...$scores` given [1, 2, 3].
Any tool handler declaring a variadic parameter was therefore
unreachable through the advertised schema, confirmed by direct
reproduction against ReferenceHandler::handle() before this change.
Casts each array element individually via the existing
castArgumentType() (unchanged - it already inspects the variadic
parameter's element type, not "array", per PHP reflection semantics)
and appends to the result in order; variadic is always the last
parameter, so this preserves correct positional ordering. A non-array
value for a variadic argument now fails clearly with an invalid-params
error instead of the confusing cast message.
* Address chr-hertel's review: wrap scalars, reject objects, index cast errors
- Non-array variadic argument is now wrapped as a single-element list
instead of being rejected: MCP prompt arguments are always
Record<string,string> per protocol, so a variadic prompt parameter
legitimately receives a plain string, not an array. Rejecting it
regressed behaviour that worked on main.
- A JSON object (non-list array) is now explicitly rejected instead of
being silently flattened into positional args by array_values().
- Cast-failure errors now include the parameter name and element index,
so a failure inside a large array is actually actionable.
- Trimmed the inline rationale comment; the PR description already
covers it.
Tests updated: replaced the "throws on non-array" test (no longer
correct) with tests for the wrap-as-single-element, reject-object, and
indexed-error-message behaviours.
* Pin argument order for an injectable before a variadic
---------
Co-authored-by: Christopher Hertel <mail@christopher-hertel.de>
* [Capability] Suppress list-changed events during the initial registry load
* Shorten dispatch() comment to a one-liner
* Use multi-line docblock format for dispatch() comment
RequestMeta::parseTraceContext() now falls back to the native traceparent/
tracestate/baggage HTTP headers when _meta carries none, and HeaderFactory
mirrors an application-set _meta trace context onto those same headers on
outbound requests. _meta wins when both are present.
* [Server] Make resource subscribe/unsubscribe errors protocol-version-aware (SEP-2164)
ReadResourceHandler already picked -32602 vs -32002 based on the
negotiated revision; ResourceSubscribeHandler and
ResourceUnsubscribeHandler still hard-coded the retired -32002 for
every client. Bring them in line.
* Remove unnecessary comments
Co-authored-by: Christopher Hertel <mail@christopher-hertel.de>
One handler asks the user something on any revision now. Where the client
can be asked mid-request it still is; where it cannot, the ask becomes the
input_required result that revision carries and the same call returns the
answer once the client re-sends it.
Costs one handler entry per ask, so side effects belong after the last
question. Answers from earlier rounds travel in the requestState, which is
why asking more than once needs Builder::setRequestState().
sample() and listRoots() still raise a LogicException there: that revision
removed them outright.
docs/stateless-lifecycle.md walks the revision end to end - per-request
_meta, server/discover, multi round-trip requests, caching and subscriptions
- and a runnable example on each side shows what that looks like in code.
CI matrixes both conformance roles over both revisions. --spec-version is
cumulative across the dated revisions, so 2025-11-25 covers the two before
it; 2026-07-28 is not cumulative with them and needs its own run against its
own baseline. The runner is pinned so a PR only goes red for reasons in the
PR, and conformance-weekly tracks the moving target instead - on both
revisions, since the draft scenarios ship only on the alpha dist-tag.
The Inspector cannot reach a modern-lifecycle server (it opens with
initialize), so the example is covered by integration tests instead: one
drives it with hand-built HTTP the way a conforming client would, one drives
it with this SDK's client, and one drives the single endpoint from both eras.
* [Client] Speak the 2026-07-28 lifecycle
On that revision the client opens with server/discover instead of
initialize, stamps each request's _meta with the protocol version, its own
capabilities and client info, and sends the standard Mcp-Method / Mcp-Name /
Mcp-Param-* headers an intermediary routes on. ToolCatalog is what knows,
from the tool list, which arguments a call has to mirror.
An input_required result is answered rather than surfaced: InputRequestResolver
asks the host's elicitation, sampling and roots handlers and retries the same
request carrying inputResponses and the requestState the server sent.
McpHeader holds the header names and the =?base64?…?= sentinel, so the two
sides no longer keep their own copies of the same rules.
* Fix CI baseline drift and address Copilot review findings
Drop stale expected-failure entries the new client fixture now passes.
Copilot review: fix McpHeader round-trip for a wrapper-shaped literal,
encode empty/numeric-keyed inputResponses as a JSON object, stop the
incident example repeating side effects on MRTR retry, and reset
ToolCatalog on reconnect instead of carrying the previous server's
verdicts.
* Skip DualEraExampleTestCase on PHP 8.1
PHP_CLI_SERVER_WORKERS does not reliably fork multiple php -S
workers there (php/php-src#9400), reproducing the single-worker
deadlock the extra workers exist to avoid.
* [Server] Serve both protocol eras from one endpoint
The two lifecycles share a transport, not a dispatcher. StreamableHttpTransport
classifies each request - a 2026-07-28 envelope, an initialize handshake, or a
session-bound follow-up - through InboundClassifier and routes it to the
dispatcher that owns it, so one URL answers a modern client and a handshake-era
one alike. Builder::build() carries both; withoutModernEra() opts out and
setModernVersions() narrows what the modern leg answers for.
InputRequiredShim lets a handler written for multi round-trip requests also
serve a handshake-era client, by turning each ask into the request/response
exchange that era has - so a handler is written once rather than twice.
The conformance fixture collapses into one server for the same reason: both
legs now hit the same URL.
* Trim redundant comments, fix stale in-memory claim in conformance fixture
* Drop stale json-schema-2020-12 baseline entries
The new fixture tool makes this scenario pass in both eras now.
* Reject handshake revisions in setModernVersions(), warn on unsafe custom middleware
Copilot review: setModernVersions() accepted revisions InboundClassifier
would never route to the modern leg. StreamableHttpTransport now warns
when a custom middleware list carries ProtocolVersionMiddleware, since
it runs before era classification and rejects modern-era traffic by
default. Also fixed a misleading docblock claim on the same middleware.
* [Server] Serve an extension's methods under the modern lifecycle
The modern dispatcher takes the method-to-extension map the builder collects,
so a method belonging to an extension this server does not serve is answered
-32601 naming the extension instead of a bare "no handler found". It is still
an unknown method - the server genuinely does not implement it - but the
caller can now act on the answer.
* [Server] Reject two extensions claiming the same RPC method
The message factory resolves a contested method to whichever class
registered first, while extensionMethods kept the last one — so error
messages could name the wrong extension. Also fixes a test that
claimed to prove a method gets named by its extension while asserting
the opposite; the case it meant to cover (an enabled extension with no
handler for one of its methods) had no coverage at all.
* [Server] Enforce x-mcp-header where a request uses it
Only top-level properties were inspected, so an annotation on a nested
property was silently unenforced - the header could say anything. The walk
now follows a chain of properties keys to any depth and reads the argument
at that exact path.
Integer values compare numerically, so a client writing 42.0 for a body
value of 42 is no longer rejected.
* [Server] Stop comparing numeric-looking strings as numbers
is_numeric() and loose == both treat a string like "042" as equal to
"42", so a string argument could pass header validation against a
header that doesn't match it exactly. Gate the numeric comparison on
the argument's actual type and a decimal-only header instead.
traceparent, tracestate and baggage reach handlers through RequestContext
and are echoed onto the notifications the request causes, so a span is not
broken by the response stream. Passed through verbatim: what they mean is
the tracing library's business, and no dependency is taken for it.
The fixture returned the literal template pattern test://template/{id}/data
instead of substituting id, failing the wire-schema uri format check. Also
switches the json-schema-2020-12 baseline entry to a bare scenario name: the
latest (0.1.x) conformance CLI doesn't parse scenario:check-id syntax, only
alpha (0.2.x) does, and the two need to agree since conformance-weekly runs
2025-11-25 against latest while pipeline.yaml pins alpha.
Verified locally against both dist-tags and both spec revisions.
The stream acknowledged, slept thirty seconds and closed, so tools and
resource changes never reached a client — and with resources/subscribe
gone, that left the revision with no server-push at all. A notification bus
carries them: in-memory where publisher and stream share a process, PSR-16
where php-fpm puts them in different ones. The conformance server proves
the latter — a tools/call in one worker reaches a stream held in another.
Every cacheable result went out as ttlMs 0, private — the letter of
SEP-2549 and none of its point, with no way to change it. CachePolicy adds
a default and per-method overrides; a resource can override per read. The
conservative default stands, because public means a shared cache may serve
one caller's answer to another.
- Rename conformance-baseline.yml to conformance-baseline-2025-11-25.yml
so both baseline files carry their spec revision symmetrically, and
bring conformance-server/conformance-client up to the same rigor
(--suite all, pinned CONFORMANCE, baseline check) as the draft targets
- score.php now takes a required spec-version and writes
<suite>-conformance-<spec-version>.json, so 2025-11-25 and 2026-07-28
each get their own badge instead of only the dated one
- conformance-weekly.yaml generates and publishes badges for both
matrix entries instead of gating on 2025-11-25 only, and drops the
now-stale unversioned badge files from the badges branch
- README shows all four badges alongside the existing packagist/CI/PHP
version/license row
- Trimmed workflow comments to the minimum needed
* [Server] Let resources/read ask for input, and stop caching retries
The MRTR pattern names three methods and resources/read was the one whose
handler dropped an InputRequiredResult into the resource formatter. A
retry's answer also carried ttlMs and cacheScope, which the caching rules
forbid: its inputs are not part of any cache key.
* [Server] Refuse an ask the client cannot answer
A handler could put an elicitation into inputRequests for a client that
never declared one, and the only symptom was a retry that never came.
Checked against the request's own capabilities now, and reported as -32021
with the missing set — which is the code the spec defines for exactly this.
* [Server] Hand multi round-trip answers back typed
response() returned whatever JSON arrived, so every handler re-implemented
the same array poking. A malformed answer now reads as absent, which sends
the handler back to asking rather than to failing.
* [Server] Gate streamed input_required asks by capability too
checkInputRequests() only ran on the non-streaming path, so a handler
that emits a notification before returning InputRequiredResult could
still leak an undeclared-capability ask under HTTP 200.
* [CI] Matrix the conformance jobs over both spec revisions
Runs server/client conformance against both 2025-11-25 and 2026-07-28,
each against its own endpoint and baseline. PR pipeline pins an exact
conformance package version; conformance-weekly tracks the moving
target on both dist-tags.
* [CI] Run the draft conformance suite from the alpha package
The conformance repo now ships 2026-07-28 scenarios on its alpha
dist-tag, so the draft targets no longer need a local checkout.
Pinned to the same version the pipeline runs.
* [CI] Baseline the failures --suite all newly exposes
The old workflow ran without --suite all/--spec-version, so several
pre-existing gaps never surfaced in CI. Now that both revisions run
the full suite:
- server: json-schema-2020-12 and resources-templates-read are
fixture bugs shared by both revisions; 2026-07-28 also needs
resources/subscribe and resources/unsubscribe declared removed
under the stateless wire, which the server doesn't yet do
- client: the example client doesn't call add_numbers or echo back
a tool's inputSchema; against 2026-07-28 it doesn't speak the
lifecycle or OAuth at all yet, so every such scenario fails
Verified locally: all four conformance jobs (server/client x
2025-11-25/2026-07-28) pass against these baselines.
tools/call answered with one JSON object, so a handler's progress and log
notifications had nowhere to go — and reached for a fiber that was not
there, coming back as -32603. Handlers now run in a fiber and the stream
opens after the first suspension, late enough that an error still gets the
status the spec fixes for it. logLevel finally gates the log messages it
was already being parsed for, and a server-initiated request is refused by
name rather than as a FiberError.
* [Server] Let the client gateway see modern-era capabilities
The six supports*() probes read session state only InitializeHandler wrote,
so every request served statelessly reported no client capabilities at all —
a tool guarding on them took the unsupported branch without a signal.
StatelessProtocol now writes the request's declaration under the same keys.
* [Server] Reject the RPCs 2026-07-28 removed
resources/subscribe and resources/unsubscribe were still dispatched and
answered 200 OK, recording subscriptions the modern era never reads. They
join initialize, ping and logging/setLevel in the era guard; the handlers
stay registered because the handshake era still serves them.
* [Server] Decode a wrapped Mcp-Name before comparing it
The Base64 sentinel was only unwrapped for Mcp-Param-*, so any resource URI
or tool name outside the header-safe ASCII set — which the spec explicitly
tells clients to wrap — was compared encoded and refused with -32020.
* [Server] Require the MCP-Protocol-Version header
It was only checked for contradicting the body, so omitting it entirely
passed — leaving the value intermediaries route on unenforced. Tied to the
header validator's presence, since that is what a header-bearing transport
installs and stdio carries its metadata inline.
* [Server] Read the modern-era request body whole
A single read() takes whatever the stream cares to give — 64 bytes from a
chunked transfer is legal — so an oversized-but-valid POST came back as a
parse error. Shares the legacy transport's incremental read as a trait.
* [Schema][Server] Answer a notification with a status, not a response
A body with no id was dispatched as a request and came back as a JSON-RPC
error carrying "id": "" — an id nobody issued, and on a path a client uses
to tell a modern server from a legacy one. Notifications now get 202 and no
body, and Error omits an id it could not read instead of emptying it.
* [Server] Stop emitting the error codes 2026-07-28 reserved
resources/read answered -32002, which this revision forbids; prompts/get
and completion/complete answered it for an unknown *name*, which it never
meant; tools/call answered -32601, which belongs to an unknown method.
Only the resources/read code is revision-gated — older peers expect -32002
there, and nowhere else. Also routes the -32021 capability exception out of
the prompt and resource handlers, which were swallowing it into -32603.
The revision has no way for a server to put a request on a response stream,
so a handler that needs input from the user says so in its result instead: a
handler returning InputRequiredResult comes back as
resultType: "input_required" with the asks and an opaque requestState, and
the client retries the same request carrying inputResponses.
RequestContext::getInputContext() is what the handler reads them through.
RequestStateCodec signs and time-bounds the state, so a tampered or expired
one is refused rather than trusted - the state travels through the client,
which is not a place to keep something the server later believes.
The revision puts the method, the subject name and the mirrored arguments in
headers so an intermediary can route and cache without parsing the body. A
header that contradicts the body it travels with therefore has to be
refused, not ignored: StandardHeaderValidator compares each against the
request and answers -32020 on a mismatch.
Set with Builder::setHeaderValidator(); a transport with no header layer is
unaffected.
* [Server][Tests] Serve the modern (SEP-2575) lifecycle
The revision removed the initialize handshake: every request carries its own
_meta - protocol version, client capabilities, client info - and answers
server/discover for what initialize used to return. There is no session to
keep, so there is nothing for the handshake-era dispatcher to do.
Builder::buildStateless() produces StatelessProtocol, and
StatelessHttpTransport serves it over one POST-only endpoint - GET and
DELETE are 405, and no Mcp-Session-Id exists to hand out.
Results go through a per-era wire codec rather than the handlers, so
resultType and the caching hints are stamped for 2026-07-28 and left off the
revisions that do not define them.
Covered by the modern conformance suite (make conformance-draft-server)
against its own baseline; unit coverage arrives with the fixes stacked on
this.
* Apply suggestions from code review
Co-authored-by: Christopher Hertel <mail@christopher-hertel.de>
* Fix critical stateless-lifecycle bugs from review
- Don't leak internal exception messages to clients on uncaught
\Throwable; use a generic message like the handshake-era protocol
and every built-in handler already do.
- Default a missing/invalid request id to null instead of "", and
stop answering notifications and malformed messages as if they
were "method not found" 404s.
- Read the POST body in a loop instead of a single read() call,
which PSR-7 allows to return short.
* [Schema][Server] Add the extensions framework SEP-2133 defines
enableExtension() took any string and advertised it, with no way for an
extension to add a method: MessageFactory could not decode one, so nothing
downstream ever saw it.
Identifiers are now checked against the _meta key naming rules through
ExtensionIdentifier - a prefix is mandatory, since an unprefixed name has no
owner, and the modelcontextprotocol/mcp second labels the specification
reserves are recognised. An extension implementing
MethodProvidingExtensionInterface contributes both its message classes and
the handlers serving them.
MessageFactory::make() takes an $additional list of message classes, and
RequestHandlerInterface's result template is covariant so a handler
declaring a concrete result satisfies a collection typed by the interface.
* Fix stale @see reference in MethodProvidingExtensionInterface docblock
* [Schema][Server][Client] Make ExtensionIdentifier a value object
ExtensionInterface::getId() now returns ExtensionIdentifier, which
validates the SEP-2133 naming rules at construction instead of the
callers checking it. getMessages()/getRequestHandlers() move into
ExtensionInterface itself; AbstractExtension gives them empty
defaults for extensions that only announce a capability.
SEP-2207 says a resource server must not advertise `offline_access` as a
required scope. Nothing in src/ writes it - `scopesSupported` is entirely
operator-supplied - so there was nothing to fix, but also nothing stopping a
future default from quietly starting to advertise one.
* [Schema] Refuse an invalid x-mcp-header annotation at definition time
SEP-2243 calls a tool definition invalid when its x-mcp-header annotation is
empty, is not an HTTP field name, collides case-insensitively with another,
or sits on a number, array or object property - none of which can be
mirrored onto a header.
Nothing said so. The earliest place to is where the tool is defined, rather
than on the request that trips over it.
* Clarify x-mcp-header validation only reaches properties
* Make checkHeaderAnnotations() a private instance method
* Make headerAnnotations() a private instance method
* Apply suggestion from @chr-hertel
* [Schema] Stop rejecting elicitation schemas the specification allows
The specification makes `title` optional on every elicitation schema
definition. AbstractSchemaDefinition required it, so a client reading a
conformant server's elicitation/create refused the request - and a server
building one could not omit it either.
`title` is now nullable on all seven definitions, omitted from the
serialized form when absent, and rejected only when present and not a
string.
* Fix stale title docblocks; allow omitting title where safe
* Give Number's title param a default too
* Apply suggestions from code review
Co-authored-by: Christopher Hertel <mail@christopher-hertel.de>
* Preserve request id in JSON-RPC error responses instead of fabricating id:""
When the server received input it could not turn into a valid message, the
error response carried a fabricated empty-string id that was never in the
request, breaking JSON-RPC client correlation.
The id default for the error factories was '' and that default reached the
wire unchanged: unrecoverable parse errors emitted id:"", and for invalid-
but-parseable messages MessageFactory discarded the decoded id so it also
fell back to "".
Now unrecoverable parse errors (-32700) return id: null per JSON-RPC 2.0,
and invalid-but-parseable requests (-32600) preserve the original request
id. Error::$id / getId() and the $id parameter of forParseError() /
forInvalidRequest() are widened to string|int|null, and the recoverable id
is threaded through InvalidInputMessageException.
Fixes#333
* Preserve nullable request id consistently across Error factories and fromArray
- Widen the remaining for*() factories (forMethodNotFound, forInvalidParams,
forInternalError, forServerError, forResourceNotFound) to string|int|null
defaulting to null, so no error path can fabricate id:"" again. The only
defaulting call site (ProtocolVersionMiddleware) now emits id:null instead
of id:"" for a pre-parse rejection, which is the spec-correct value.
- Accept a null id in Error::fromArray(): require the id key to be present
(JSON-RPC responses must carry it) but allow its value to be null, so a
spec-compliant id:null error response round-trips and incoming error
responses with id:null are accepted instead of rejected.
- Widen the @phpstan-type ErrorData id to string|int|null to match the ctor,
getId(), and jsonSerialize() shape.
* Document intentional id recovery guard in MessageFactory
* Cover id: 0 and string ids in recoverable-id preservation test
* Fixed client handling of null-id error responses
* [Schema] Omit the id of an error response that never had one
An error response whose id could not be read sent `"id": ""`, which claims
the peer issued a request with an empty-string id - a different statement
from "the id could not be read". JSON-RPC spells the latter as a null id,
and a receiver that cannot read the id is exactly the case the code covers.
[BC Break] Error::$id accepts null and getId() may return it. All the for*()
factories default to null, fromArray() accepts a missing or explicitly-null
id, and the member is omitted from the serialized form when absent.
* Guard null response id in Client/Server Protocol handlers
Also fix Error::jsonSerialize()'s phpdoc: data belongs inside error, not beside it.
* [Capability] Apply a parameter's complete schema definition
Schema(definition:) is documented as taking precedence over the individual
keywords, and generate() unwraps it that way at method level. At parameter
level it was merged as an ordinary key instead, in both the plain and the
variadic path, so the property reached clients carrying a literal
"definition" member - not JSON Schema, read by nothing, and everything
inside it lost.
Both paths now replace the schema with the definition. The variadic one
still forces type: array afterwards, since a variadic arrives as an array
however the definition describes it, and the signature default survives a
definition that states none - it is what the handler receives when the
argument is omitted, and default is an annotation rather than a constraint.
* Fix docblock description overwriting definition on variadic params
* Apply suggestion from @chr-hertel
* [Schema] Serialize a settingless extension as an object
An extension declaring support with no settings was advertised as `[]`, an
empty JSON array, which the wire schema rejects - `extensions` maps an id to
a settings *object*. Both ServerCapabilities and ClientCapabilities now cast
each entry, so it goes out as `{}`.
* Apply suggestion from @chr-hertel
* [Capability] Stop a union type losing a branch
mapPhpTypeToJsonSchemaType() tested for array syntax before splitting the
union, so string[]|int mapped to array alone; applyArrayConstraints() then
overwrote the type with array whenever the union contained one, dropping the
scalar branch a second time. A tool declaring such a parameter rejected a
valid integer argument.
The split now happens first and only at the top level, so array<int|string>
and array{a: int|string} stay one type rather than two that do not exist, and
the array branch keeps its items alongside the others - which is what 2020-12
means by items: a constraint that applies only when the instance is an array.
* Apply suggestion from @chr-hertel
* Isolate the array branch before inferring its element type
* Don't apply items to arrays whose union branches disagree on element type
* [Examples][Tests] Make the MCP Apps example work over HTTP, and pin it
It was the only example without a session store, so every request under
`php -S` or php-fpm - each its own process - was answered "Session not found
or has expired". It now uses a FileSessionStore like the others.
Inspector snapshots pin what the extension actually puts on the wire: the
ui:// resource's `text/html;profile=mcp-app` MIME type and `ui` marker, and
the `_meta.ui` on the tool that links the two.
* Apply suggestion from @chr-hertel
* Apply suggestion from @chr-hertel
* [Schema] Drop SDK-only name pattern on resources and resource templates
* Update src/Schema/ResourceDefinition.php
* Update src/Schema/ResourceTemplate.php
* [Schema][Server] Close the 2025-11-25 schema gaps and add the 2026-07-28 surface
Ports the type definitions this SDK still misses outside of sampling tool use,
which #409 and #420 already cover. Every addition is optional and defaults to
current behaviour, so a connection negotiated on an older revision is
unaffected.
From 2025-11-25, elicitation gains modes. ElicitationMode splits `form` —
build a form from the requested schema — from `url`, which sends the user out
of band and returns only the accept/decline/cancel outcome. That is why
requestedSchema becomes optional and `url` appears beside it.
ClientCapabilities learns the matching sub-capabilities, where an
`elicitation` naming no mode declares form, the only shape that existed
before. Icon gains `theme` from the same revision, and Implementation gains
the `title` BaseMetadata has carried since 2025-06-18.
From 2026-07-28, schemas loosen where the revision loosens them: SEP-2106
drops the object-only restriction, so Tool::outputSchema may describe any JSON
value and CallToolResult::structuredContent follows.
The same revision defines three error codes (-32020 header mismatch, -32021
missing required client capability, -32022 unsupported protocol version).
ProtocolVersionMiddleware switches to the last of them, so a rejected version
carries the supported set as structured data the client can retry from rather
than only as prose.
* [Schema] Reject what the 2026-07-28 types cannot represent
`Implementation::title` reached the typed constructor unchecked, so malformed
wire data raised a TypeError instead of InvalidArgumentException.
`ToolUseContent::input` accepted a list and serialized it as a JSON array,
where the protocol requires an object. The empty array stays exempt: it is
also an empty map and still emits `{}`.
`ToolChoice` and `ElicitRequest` read their mode with isset(), which is false
for an explicit null, so `{"mode": null}` silently became the default instead
of being rejected. Both use array_key_exists() now, letting the existing type
check refuse null.
* [Client][Server] Advertise the implementation title through both builders
`Implementation::title` could be parsed but never sent: neither
`Client\Builder::setClientInfo()` nor `Server\Builder::setServerInfo()`
accepted one, so every SDK user emitted null.
Both gain a trailing optional `$title`. On the server it sits where the
Implementation constructor already puts it, so existing positional calls keep
their meaning; the client builder forwards it by name, leaving the icons and
websiteUrl slots defaulted.
* [Schema] Serialize every non-null structuredContent
The object-only hydration guard was never that: `!is_array()` admitted
`[1, 2, 3]` and `[]`, which serialize to JSON arrays, while rejecting the
scalars 2026-07-28 permits. The truthiness emission gate was backwards in the
same way — it dropped `[]`, `0`, `false` and `""`, yet emitted lists, strings
and an empty stdClass.
Hydration now accepts any JSON value, and `null` alone means absent, matching
`ToolResultContent` which already carries this field that way. Which values a
given revision permits is a question for version-aware serialization, which
results cannot answer yet.
* [Schema][Server] Make url elicitation reachable
`ElicitRequest::forUrl()` built a request no SDK user could send: the only
public gateway method always constructed form mode from an ElicitationSchema,
and `request()` is private.
`elicitUrl()` joins `elicit()`, and both funnel through one send path that
hydrates the result with the request's own mode. Without that, a url-mode
accept — contentless by design — threw, because ElicitResult requires content
whenever the action is accept. The result carries no discriminator of its own,
so the mode has to come from the request it answers.
`supportsElicitationUrl()` reports whether the client named the mode, reusing
the sub-capability reader the sampling checks already use.
* [Schema] Add CHANGELOG entry for the 2026-07-28 surface
* [Capability][Schema][Server] Carry the widened structuredContent end to end
Review follow-up on the SEP-2106 widening, which stopped at the type.
`ToolReference::extractStructuredContent()` returned `?array`, so a tool
declaring a scalar `outputSchema` had its result dropped and logged as
unsendable. It returns `mixed` now and keeps a scalar — but only from
2026-07-28 on, and only when the tool declared an outputSchema: without one
the value is already carried in `content`, and advertising it twice is not an
improvement.
`CallToolHandler` only warned about list-shaped `structuredContent` on a
self-built `CallToolResult`, so a scalar reached revisions that require an
object unremarked. The check is on the shape now, not on the list case alone.
`ElicitResult` retained a `content` the spec says is absent from url-mode
results, leaving a malformed response indistinguishable from a valid one.
`Tool::jsonSerialize()` emitted the empty root schema as `[]`, which is not a
schema at all. Sub-schemas already had this treatment.
* harden sampling tools against the spec
Report tool-flow violations as -32602 instead of dropping the request,
accept resource_link in tool results, gate tools on sampling.tools, and
reject empty content.
* normalize keyed content arrays to lists
array_filter() and friends preserve keys, and a keyed array serializes as a
JSON object instead of the content-block array the schema requires.
* Address review feedback on resource_link
Three follow-ups from review of the resource_link work:
* PromptResultFormatter: add a regression test proving the optional
fields of a typed resource_link block (title, description, mimeType,
size, annotations, _meta) survive formatting. The delegation to
ResourceLink::fromArray() already landed while rebasing onto the
PromptResultFormatter refactor, but the existing test only supplied
uri/name and so could not catch a reintroduced field drop.
* ResourceLink::fromArray(): validate optional fields consistently, the
way the equivalent ResourceDefinition::fromArray() already does.
description, mimeType and size now raise InvalidArgumentException
instead of surfacing a TypeError (or silently coercing, for size's
(int) cast); annotations goes through Annotations::tryFromArray(),
which carries the is_array() guard that icons already had; icons uses
Icon::listFromArray() so a non-array entry is reported in context.
* ToolReference::extractStructuredContent(): never emit a list as
structuredContent. The Content guard added earlier fixed only one
instance of the real invariant - structuredContent must be a JSON
object, and a PHP list can never be one. Tool::fromArray() already
enforces the matching rule by rejecting an outputSchema whose type is
not "object", so this aligns the runtime path with it. The test
asserting the opposite for an array-typed outputSchema contradicted
the phpstan type and the fromArray() check introduced alongside it in
the same commit, so it is replaced by tests for the two list shapes.
* Apply the structuredContent object rule to object results too
`extractStructuredContent()` guards raw array results against being emitted
as a JSON array, but the object branch handed back whatever `json_decode()`
produced. A `JsonSerializable` returning a list or a scalar slipped straight
through, producing exactly the `structuredContent` the array guard exists to
prevent — and a return value that contradicts the method's own
`array<string, mixed>|null` signature.
Check the decoded value before returning it, and cover the object branch,
which had no tests at all.
* Document structured tool output
`outputSchema` and `structuredContent` were undocumented: the tool return
value docs covered only the `content` side, and the schema generation
section is about tool parameters.
Add a "Structured Output" subsection covering how to declare the schema,
which return values populate `structuredContent`, and why a list has to be
wrapped in a key to get structured output at all.
* Gate structuredContent on the negotiated protocol revision
SEP-2106 (revision 2026-07-28) widens structuredContent to any JSON value;
earlier revisions require an object. Resolve the revision per request and
apply the matching rule, and warn when a declared outputSchema yields none.
* Resolve the protocol revision on RequestContext
Keeps the `_meta`-then-session lookup in one place instead of per handler,
and exposes the negotiated revision to tool handlers.
* Document reading the negotiated revision from RequestContext
* Warn when a self-built CallToolResult carries an invalid structuredContent
Returning a CallToolResult opts out of the extraction rules, so the value is
still sent unchanged — but a JSON array is not valid before SEP-2106.
* feat: support sampling with tools
* address sampling tools review feedback
* tighten sampling tools spec compliance
* drop SamplingStopReason enum in favor of open string
The spec leaves stopReason open for provider-specific values, so upcasting
the four known ones to enum cases makes every future enum addition a silent
BC break for string comparisons.
---------
Co-authored-by: Christopher Hertel <mail@christopher-hertel.de>
* [Client] Implement setMaxRetries connection retries
* [Client] Reject a negative retry count in Configuration
* [Client] Use a positive init timeout in the retry test
* [Client] Drop redundant docblock note on setMaxRetries
* [Client] Clear the initialized flag on a failed connection
* [Client] Trim comments to the non-obvious ones
* [Client] Cover retries and timeout cleanup with integration tests
The retry fixture counts the processes it was started as, which is what
separates a real retry from a second call into the same server. The timeout
fixture outlives the client's patience and then goes idle, so the next
request runs on a connection the client already gave up on once.
* [Tests] Introduce client-server integration tests
* Add an in-process loopback that connects a real client transport to a real
server transport, draining both Fiber loops in one synchronous pass instead
of polling. A stalled exchange fails where it happened rather than hanging.
* Cover protocol version negotiation across both implementations, where the
unit tests only ever drive one side against a canned counter-offer.
* Cover the server-to-client round-trips end to end: elicitation, sampling,
roots, plus progress and logging notifications.
* Add the `integration` test suite and a `make integration-tests` target.
* [Tests] Run the integration suite in CI
The unit job pins `--testsuite=unit`, so the new suite would never have run on
GitHub. It goes in that job rather than its own: the loopback leans on Fibers,
and the matrix is what covers them down to the PHP 8.1 floor.
* [Tests] Run integration tests against a real server process
Replaces the in-process loopback with the setup from examples/client: the
client spawns a fixture server over the real StdioTransport, so the tests
exercise the shipped transports and no longer nest both sides' Fibers.
Moves the suite into its own CI job, matrixed over PHP 8.1-8.5.
Sampling was the one client capability a tool could not check before asking.
`supportsRoots()` and `supportsElicitation()` both read the capabilities the
client advertised during the handshake; sampling is stored in that same session
data but had no accessor, so a tool had to issue the request and catch the
refusal.
* [Client][Docs][Schema][Server] Implement protocol version negotiation
The `initialize` handler never looked at the version the client asked for:
it always answered with the server's configured version (or the SDK default),
so a client offering a revision the server supports could still be told to
speak a different one. The spec requires echoing back a supported request and
only counter-offering when the requested version cannot be honoured.
Turns ProtocolVersion into the registry the other SDKs converged on
(typescript-sdk's protocolEras.ts, python-sdk's mcp_types/version.py):
`latestHandshake()`, `handshakeVersions()`, `modernVersions()` and ordered
comparison through `isAtLeast()`. Comparison goes through declaration order
rather than string collation, so an unrecognized peer string compares
conservatively instead of accidentally sorting above a known revision. The
one throw the enum can produce is the SDK's own LogicException, so a consumer
catching `Mcp\Exception\ExceptionInterface` does not miss it.
Adds 2026-07-28 as a known revision and splits the set into two eras:
handshake (2025-11-25 and earlier, negotiated via `initialize`) and modern
(2026-07-28+, per-request `_meta`). The lists are kept apart so a modern
revision can never leak into an `initialize` exchange, which has no
`initialize` at all. Serving the modern era is follow-up work; this only
teaches the SDK that the revision exists and must not be mis-negotiated.
Deliberately absent is any "newest revision in any era" accessor, which
would hand back the one revision the handshake must never offer.
Server: negotiates and stores the result on the session. A version pinned in
Configuration still pins the handshake to exactly that revision.
Client: validates the counter-offer and fails the handshake instead of
continuing on a revision it cannot speak; records the negotiated version on
the client state, readable through the new Client::getProtocolVersion().
The HTTP middleware now defaults to the handshake versions rather than every
declared case, so it stops accepting a modern version the server cannot yet
serve. When the header is absent it falls back to
ProtocolVersion::DEFAULT_NEGOTIATED_VERSION (2025-03-26), the revision that
introduced both Streamable HTTP and the header itself.
The negotiation behaviour is one provideNegotiationTable() data provider
whose rows are the documented table -- supported revision echoed back,
unknown or malformed counter-offered, modern counter-offered -- so the docs
and the suite cannot drift apart. The supported-revision rows iterate
handshakeVersions() rather than a literal list, so declaring a new handshake
revision extends the table automatically, and the "never answer with a modern
revision" assertion runs on every row.
Documented in server-builder.md (the enum registry, how the server answers a
handshake, what pinning changes), transports.md (the header check) and
client.md (the client's side), all linking the spec's versioning section.
* [Client][Docs][Schema][Server] Warn when the configured protocol version cannot be negotiated
* Log a warning instead of silently falling back when a modern revision is
configured on the client or pinned on the server builder.
* Document that setProtocolVersion() does not restrict the MCP-Protocol-Version
header check, which validates against a fixed set.
* Pin every revision's era in a test, so a new case declared past the modern
boundary fails instead of dropping out of the handshake set.
* Add CHANGELOG entry.
* [Client][Docs][Schema][Server] Rename version fallback constant and trim comments
* Add resource_link content block
The spec's `resource_link` content type (protocol revision 2025-06-18,
still present in 2025-11-25) had no corresponding schema class: it lets a
tool result or prompt message reference a resource by URI/name without
embedding its full contents, the way `EmbeddedResource` does. Without it,
tools returning many or large resource references (e.g. search results
over a DAM) are stuck smuggling URIs through text/structured content.
Add `Mcp\Schema\Content\ResourceLink`, mirroring the spec's `Resource`
shape (uri, name, title, description, mimeType, annotations, size, icons,
_meta) with a fixed `type: "resource_link"`, following the conventions of
the sibling `EmbeddedResource`/`ResourceDefinition` classes.
Register it everywhere `resource` content blocks are (de)serialized by
type per the spec's `ContentBlock` union:
- `CallToolResult::fromArray()`
- `PromptMessage` (constructor type, `fromArray()`, phpstan types)
- `PromptResultFormatter` (typed-array and direct-instance paths)
`SamplingMessage`/`CreateSamplingMessageResult` are intentionally left
unchanged: the spec's `SamplingMessageContentBlock` does not include
`ResourceLink`.
Fixes#399
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* Add resource_link example coverage to discovery-userprofile
Add a lookup_user tool to the discovery-userprofile example: given a
userId, it returns a short text summary plus a resource_link pointing
at the existing user://{userId}/profile resource template, instead of
embedding the profile. This mirrors the PR's motivating use case (a
tool referencing many/large resources by URI rather than inlining
them) and exercises resource_link through the example's own inspector
snapshot tests (tools_call-lookup_user[_not_found], tools_list).
Wiring this up surfaced a real bug: ToolReference::extractStructuredContent()
passed any raw array return value straight through as structuredContent,
even when it held Content instances (e.g. the TextContent/ResourceLink
pair above). That serializes to a JSON list, which fails the spec's
requirement that structuredContent be an object - the inspector's MCP
client rejected the tool call outright until fixed. Skip structuredContent
extraction whenever the raw result contains a Content item, since those
are already represented in `content` and were never meant to double as
structured data.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---------
Co-authored-by: Alex Rothberg <51059+cancan101@users.noreply.github.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
