ruby-mcp-client 2.0.0 → 2.1.0

data/lib/mcp_client/server_streamable_http.rb

@@ -33,6 +33,52 @@ module MCPClient
     SSE_MAX_RECONNECT_DELAY = 30   # Maximum reconnect delay in seconds
     THREAD_JOIN_TIMEOUT = 5 # Timeout for thread cleanup
 
+    # Floor for the delay between resumption GETs. SEP-1699's polling pattern
+    # wants fast reconnects, so this is far smaller than the events-stream
+    # floor — it only prevents a peer-supplied "retry: 0" from turning the
+    # deadline window into a back-to-back request loop.
+    MIN_RESUMPTION_RECONNECT_DELAY = 0.01
+
+    # Maximum length of a server-supplied SSE event id retained as the
+    # resumption cursor. The id is echoed in the Last-Event-ID header of
+    # subsequent requests, so an unbounded value means unbounded retained
+    # memory and oversized outbound headers.
+    MAX_EVENT_ID_LENGTH = 1024
+
+    # Characters allowed in a retained event id: printable ASCII, since the
+    # value becomes an HTTP header value. Notably excludes CR/LF. The range
+    # starts at 0x20 because a space is legal inside a field value, and
+    # rejecting ids like "cursor 42" would silently strand resumption on a
+    # stale cursor.
+    EVENT_ID_PATTERN = /\A[\x20-\x7E]+\z/
+
+    # Ceiling on concurrent threads POSTing server-initiated responses
+    # (pongs, roots/sampling/elicitation replies, error responses). Each
+    # server request on the events stream costs one blocking HTTP POST in
+    # its own thread; without a bound, a peer flooding requests could
+    # accumulate threads and connections until the host is exhausted.
+    # Responses beyond the budget are dropped, with saturation logged at most
+    # once per SATURATION_LOG_INTERVAL seconds.
+    MAX_CONCURRENT_RESPONSE_POSTS = 8
+
+    # Minimum gap between "response budget saturated" warnings
+    SATURATION_LOG_INTERVAL = 5
+
+    # Floor for server-supplied retry directives on the long-lived events
+    # stream. The directive is peer-controlled: honoring "retry: 0" literally
+    # would let a hostile server that closes every stream drive a tight
+    # reconnect loop (sustained CPU/TLS/connection churn). Waiting longer
+    # than the directive stays SEP-1699 compliant — the retry field is a
+    # lower bound on the reconnect delay, not an exact schedule.
+    MIN_EVENTS_RECONNECT_DELAY = 0.1
+
+    # Maximum bytes an SSE parse buffer (events stream or resumption GET) may
+    # hold while waiting for an event terminator. The stream is
+    # peer-controlled: without a cap, a hostile server could withhold the
+    # blank-line delimiter forever and grow the buffer until the host runs
+    # out of memory. Generous enough for any legitimate JSON-RPC event.
+    MAX_SSE_BUFFER_BYTES = 32 * 1024 * 1024
+
     # @!attribute [r] base_url
     #   @return [String] The base URL of the MCP server
     # @!attribute [r] endpoint
@@ -98,6 +144,7 @@ module MCPClient
 
       @read_timeout = opts[:read_timeout]
       @faraday_config = opts[:faraday_config]
+      @max_decompressed_body_bytes = validate_decompression_limit(opts[:max_decompressed_body_bytes])
       @tools = nil
       @tools_data = nil
       @prompts = nil
@@ -113,12 +160,18 @@ module MCPClient
       @last_event_id = nil
       @sse_retry_ms = nil
       @pending_stream_responses = {}
+      @response_post_count = 0
+      # Saturation bookkeeping for the response-POST budget
+      @dropped_response_posts = 0
+      @last_saturation_log_at = nil
       @oauth_provider = opts[:oauth_provider]
 
       # SSE events connection state
       @events_connection = nil
       @events_thread = nil
-      @buffer = '' # Buffer for partial SSE event data
+      @buffer = +'' # Buffer for partial SSE event data
+      # How much of @buffer has already been searched for an event terminator
+      @buffer_scanned = 0
       @elicitation_request_callback = nil # MCP 2025-06-18
       @roots_list_request_callback = nil # MCP 2025-06-18
       @sampling_request_callback = nil # MCP 2025-11-25
@@ -492,7 +545,8 @@ module MCPClient
         @prompts_data = nil
         @resources = nil
         @resources_data = nil
-        @buffer = ''
+        @buffer = +''
+        @buffer_scanned = 0
 
         @logger.info('Cleanup completed')
       end
@@ -534,6 +588,17 @@ module MCPClient
 
     # Default options for server initialization
     # @return [Hash] Default options
+    # Validate the configured decompression ceiling.
+    # @param value [Object] the max_decompressed_body_bytes option
+    # @return [Integer] the validated positive byte limit
+    # @raise [ArgumentError] if the value is not a positive Integer
+    def validate_decompression_limit(value)
+      return value if value.is_a?(Integer) && value.positive?
+
+      raise ArgumentError,
+            "max_decompressed_body_bytes must be a positive Integer, got #{value.inspect}"
+    end
+
     def default_options
       {
         endpoint: '/rpc',
@@ -544,7 +609,8 @@ module MCPClient
         name: nil,
         logger: nil,
         oauth_provider: nil,
-        faraday_config: nil
+        faraday_config: nil,
+        max_decompressed_body_bytes: JsonRpcTransport::MAX_DECOMPRESSED_BODY_BYTES
       }
     end
 
@@ -723,12 +789,17 @@ module MCPClient
     end
 
     # Reconnect delay for the events stream: the server's SSE retry directive
-    # (in ms) when present, otherwise the caller's backoff value.
+    # (in ms) when present, otherwise the caller's backoff value. The
+    # peer-controlled directive is floored at MIN_EVENTS_RECONNECT_DELAY so a
+    # zero/near-zero value cannot drive a tight reconnect loop; the
+    # deadline-bounded resumption loop intentionally keeps honoring zero.
     # @param fallback_seconds [Numeric] exponential-backoff fallback
     # @return [Numeric] delay in seconds
     def events_reconnect_delay(fallback_seconds)
       retry_ms = @sse_retry_ms
-      retry_ms ? retry_ms / 1000.0 : fallback_seconds
+      return fallback_seconds unless retry_ms
+
+      [retry_ms / 1000.0, MIN_EVENTS_RECONNECT_DELAY].max
     end
 
     # Wait for a response replayed after the POST stream was closed before
@@ -770,8 +841,11 @@ module MCPClient
       deadline = Process.clock_gettime(Process::CLOCK_MONOTONIC) + @read_timeout
       state = { cursor: cursor, retry_ms: retry_ms }
       # SEP-1699: the client MUST respect the server's retry directive before
-      # attempting to reconnect.
-      delay = retry_ms ? retry_ms / 1000.0 : 0
+      # attempting to reconnect. With no directive the FIRST GET goes out
+      # immediately, as before this change: delaying it adds latency to every
+      # resumption and, on a short read_timeout, can burn the whole budget
+      # before any I/O happens.
+      delay = retry_ms ? resumption_delay(retry_ms) : 0
 
       loop do
         sleep(delay) if delay.positive?
@@ -779,10 +853,21 @@ module MCPClient
         break unless @mutex.synchronize { @pending_stream_responses.key?(request_id) }
         break if Process.clock_gettime(Process::CLOCK_MONOTONIC) >= deadline
 
-        delay = state[:retry_ms] ? state[:retry_ms] / 1000.0 : SSE_RECONNECT_DELAY
+        delay = resumption_delay(state[:retry_ms])
       end
     end
 
+    # Delay before the next resumption GET: the peer's retry directive when
+    # present, floored so "retry: 0" cannot turn the deadline window into a
+    # back-to-back request loop, otherwise the standard reconnect delay.
+    # @param retry_ms [Integer, nil] retry directive in milliseconds
+    # @return [Numeric] delay in seconds
+    def resumption_delay(retry_ms)
+      return SSE_RECONNECT_DELAY unless retry_ms
+
+      [retry_ms / 1000.0, MIN_RESUMPTION_RECONNECT_DELAY].max
+    end
+
     # One GET with the current cursor; complete SSE events are dispatched
     # through the standard server-message path, which routes replayed
     # responses to their registered waiters. `id:` and `retry:` fields
@@ -803,6 +888,10 @@ module MCPClient
         req.options.on_data = proc do |chunk, _bytes|
           buffer << chunk
           process_resumption_buffer(buffer, state)
+          # A peer replaying delimiter-free data must not grow this buffer
+          # without bound: abort the GET (rescued below); the deadline-bounded
+          # resumption loop decides whether to retry with a fresh buffer.
+          enforce_sse_buffer_cap!(buffer)
         end
       end
     rescue StandardError => e
@@ -815,10 +904,16 @@ module MCPClient
     # @param state [Hash] mutable resumption state (:cursor, :retry_ms)
     # @return [void]
     def process_resumption_buffer(buffer, state)
-      while (separator = buffer.match(/\r\n\r\n|\n\n/))
+      # Same incremental scan as the events stream: without a cursor every
+      # chunk re-matched the whole accumulated buffer from byte zero.
+      scan_from = [state[:scanned].to_i - 3, 0].max
+      while (separator = buffer.match(/\r\n\r\n|\n\n/, scan_from))
         event_text = buffer.slice!(0, separator.end(0))
         handle_resumption_event(event_text, state)
+        scan_from = 0
+        state[:scanned] = 0
       end
+      state[:scanned] = buffer.length
     end
 
     # Parse one SSE event received on a resumption stream: track `id:` lines
@@ -835,7 +930,9 @@ module MCPClient
           data_lines << line.sub(/\Adata:\s*/, '')
         elsif line.start_with?('id:')
           id = line.sub(/\Aid:\s*/, '').strip
-          state[:cursor] = id unless id.empty?
+          # Same validation as the events stream: this cursor is echoed in
+          # the Last-Event-ID header of the next resumption GET.
+          state[:cursor] = id if retainable_event_id?(id)
         elsif line.start_with?('retry:')
           raw = line.sub(/\Aretry:\s*/, '').strip
           state[:retry_ms] = raw.to_i if raw.match?(/\A\d+\z/)
@@ -875,22 +972,64 @@ module MCPClient
     # Buffers partial chunks and processes complete SSE events
     # @param chunk [String] the chunk to process
     def process_event_chunk(chunk)
-      @logger.debug("Processing event chunk: #{chunk.inspect}") if @logger.level <= Logger::DEBUG
+      # Size only: the chunk is raw wire data carrying sampling prompts,
+      # elicitation content and tool results.
+      @logger.debug("Processing event chunk (#{describe_body_size(chunk)})") if @logger.level <= Logger::DEBUG
 
       @mutex.synchronize do
-        @buffer += chunk
-
-        # Extract complete events (SSE format: events end with double newline)
-        while (event_end = @buffer.index("\n\n") || @buffer.index("\r\n\r\n"))
+        # Append in place and scan only the newly arrived bytes. `+= chunk`
+        # reallocated and copied the whole buffer per callback, and each
+        # search restarted at index 0, so an unterminated event delivered in
+        # N chunks cost O(N^2) work: capped memory, uncapped CPU.
+        @buffer << chunk
+
+        scan_from = [@buffer_scanned - 3, 0].max
+        while (event_end = next_buffer_event_end(scan_from))
           event_data = extract_event(event_end)
           parse_and_handle_event(event_data)
+          scan_from = 0
+          @buffer_scanned = 0
+        end
+        @buffer_scanned = @buffer.length
+
+        # Everything left is a partial event awaiting its terminator; cap how
+        # much a peer may accumulate. The raise propagates (unlike processing
+        # errors below) so the events loop drops this connection and its
+        # backoff takes over — memory stays bounded either way.
+        begin
+          enforce_sse_buffer_cap!(@buffer)
+        rescue MCPClient::Errors::ConnectionError
+          @buffer = +''
+          @buffer_scanned = 0
+          raise
         end
       end
+    rescue MCPClient::Errors::ConnectionError
+      raise
     rescue StandardError => e
       @logger.error("Error processing event chunk: #{e.message}")
       @logger.debug(e.backtrace.join("\n")) if @logger.level <= Logger::DEBUG
     end
 
+    # Index of the earliest event terminator at or after an offset.
+    # @param offset [Integer] character offset to start searching from
+    # @return [Integer, nil] index of the terminator, or nil if none yet
+    def next_buffer_event_end(offset)
+      lf = @buffer.index("\n\n", offset)
+      crlf = @buffer.index("\r\n\r\n", offset)
+      [lf, crlf].compact.min
+    end
+
+    # @param buffer [String] a partial-event SSE buffer
+    # @raise [MCPClient::Errors::ConnectionError] when the buffer exceeds MAX_SSE_BUFFER_BYTES
+    def enforce_sse_buffer_cap!(buffer)
+      return if buffer.bytesize <= MAX_SSE_BUFFER_BYTES
+
+      raise MCPClient::Errors::ConnectionError,
+            "SSE event exceeded the maximum buffered size (#{MAX_SSE_BUFFER_BYTES} bytes) " \
+            'without a terminator'
+    end
+
     # Extract a single event from the buffer
     # @param event_end [Integer] the position where the event ends
     # @return [String] the extracted event data
@@ -922,9 +1061,12 @@ module MCPClient
           # SSE allows multiple data lines that should be joined with newlines
           data_lines << line[5..].strip
         elsif line.start_with?('id:')
-          # Track event ID for resumability (MCP future enhancement)
+          # Track event ID for resumability (MCP future enhancement). The id
+          # is peer-controlled and gets echoed in the Last-Event-ID header, so
+          # only a bounded, header-safe value is retained; the previous valid
+          # cursor is kept when one is rejected.
           event[:id] = line[3..].strip
-          @last_event_id = event[:id]
+          @last_event_id = event[:id] if retainable_event_id?(event[:id])
         elsif line.start_with?('retry:')
           # SEP-1699: the client MUST respect the server's retry directive
           # (milliseconds) when reconnecting; zero is a valid directive.
@@ -947,10 +1089,19 @@ module MCPClient
       return if data.empty?
 
       begin
-        dispatch_server_message(JSON.parse(data))
+        message = JSON.parse(data)
+        unless message.is_a?(Hash)
+          # A JSON-parseable scalar/array is not a JSON-RPC message; dispatch
+          # would raise on it. Log the type, never the value.
+          @logger.warn("Skipping non-object JSON-RPC message on the events stream (#{message.class})")
+          return
+        end
+
+        dispatch_server_message(message)
       rescue JSON::ParserError => e
-        @logger.error("Invalid JSON in server message: #{e.message}")
-        @logger.debug("Raw data: #{data.inspect}") if @logger.level <= Logger::DEBUG
+        # The parser message names the failure position, not the payload; the
+        # payload itself stays out of the log.
+        @logger.error("Invalid JSON in server message: #{describe_parse_error(e, data)}")
       end
     end
 
@@ -985,26 +1136,9 @@ module MCPClient
         result: {}
       }
 
-      # Send pong response in a separate thread to avoid blocking event processing
-      Thread.new do
-        conn = http_connection
-        response = conn.post(@endpoint) do |req|
-          @headers.each { |k, v| req.headers[k] = v }
-          req.headers['Mcp-Session-Id'] = @session_id if @session_id
-          req.headers['Mcp-Protocol-Version'] = @protocol_version if @protocol_version
-          # MCP: authorization MUST be included in every HTTP request
-          @oauth_provider&.apply_authorization(req)
-          req.body = pong_response.to_json
-        end
-
-        if response.success?
-          @logger.debug("Sent pong response for ping ID: #{ping_id}") if @logger.level <= Logger::DEBUG
-        else
-          @logger.warn("Failed to send pong response: HTTP #{response.status}")
-        end
-      rescue StandardError => e
-        @logger.error("Failed to send pong response: #{e.message}")
-      end
+      # Pongs go through the same bounded response path as every other
+      # server-initiated reply, so a ping flood cannot fan out threads.
+      post_jsonrpc_response(pong_response)
     end
 
     # Handle incoming JSON-RPC request from server (MCP 2025-06-18)
@@ -1029,8 +1163,11 @@ module MCPClient
         send_error_response(request_id, -32_601, "Method not found: #{method}")
       end
     rescue StandardError => e
+      # The exception message is host-internal (file paths, connection
+      # strings, library internals): log it locally, but answer the peer with
+      # a constant message so failures cannot be used to probe the host.
       @logger.error("Error handling server request: #{e.message}")
-      send_error_response(request_id, -32_603, "Internal error: #{e.message}")
+      send_error_response(request_id, -32_603, 'Internal error')
     end
 
     # Handle elicitation/create request from server (MCP 2025-11-25)
@@ -1184,7 +1321,31 @@ module MCPClient
     # @return [void]
     # @private
     def post_jsonrpc_response(response)
-      # Send response in a separate thread to avoid blocking event processing
+      # Send the response in a separate thread to avoid blocking event
+      # processing, but never beyond the concurrency budget: server requests
+      # arrive at the peer's rate, and each unanswered POST would otherwise
+      # pin one thread and one connection.
+      unless acquire_response_post_slot
+        log_response_post_saturation
+        return
+      end
+
+      begin
+        start_response_post_thread(response)
+      rescue ThreadError => e
+        # Thread.new failed, so the ensure inside the block never runs and the
+        # reservation would leak. Eight such failures would silently mute this
+        # instance's replies for the rest of its life, including after
+        # reconnect (cleanup does not reset the counter).
+        release_response_post_slot
+        @logger.error("Failed to start response POST thread: #{e.message}")
+      end
+    end
+
+    # Spawn the worker that POSTs one server-initiated response.
+    # @param response [Hash] the JSON-RPC response
+    # @return [Thread]
+    def start_response_post_thread(response)
       Thread.new do
         conn = http_connection
         json_body = JSON.generate(response)
@@ -1199,13 +1360,55 @@ module MCPClient
         end
 
         if resp.success?
-          @logger.debug("Sent JSON-RPC response: #{json_body}")
+          @logger.debug("Sent JSON-RPC response: #{describe_jsonrpc_message(response)}")
         else
           @logger.warn("Failed to send JSON-RPC response: HTTP #{resp.status}")
         end
       rescue StandardError => e
         @logger.error("Failed to send JSON-RPC response: #{e.message}")
+      ensure
+        release_response_post_slot
       end
     end
+
+    # Warn that the response-POST budget is saturated, at most once per
+    # SATURATION_LOG_INTERVAL.
+    #
+    # The peer controls how often this path is reached, so logging every
+    # rejection (at WARN, which the default logger emits) would just trade a
+    # thread-exhaustion vector for a log-volume one. The peer-supplied request
+    # id is deliberately omitted for the same reason.
+    # @return [void]
+    def log_response_post_saturation
+      now = Process.clock_gettime(Process::CLOCK_MONOTONIC)
+      dropped = @mutex.synchronize do
+        @dropped_response_posts += 1
+        next nil if @last_saturation_log_at && now - @last_saturation_log_at < SATURATION_LOG_INTERVAL
+
+        @last_saturation_log_at = now
+        @dropped_response_posts
+      end
+      return unless dropped
+
+      @logger.warn("Dropping server-initiated responses: #{MAX_CONCURRENT_RESPONSE_POSTS} " \
+                   "response POSTs already in flight (#{dropped} dropped so far)")
+    end
+
+    # Reserve one slot of the response-POST concurrency budget.
+    # @return [Boolean] whether a slot was available
+    def acquire_response_post_slot
+      @mutex.synchronize do
+        return false if @response_post_count >= MAX_CONCURRENT_RESPONSE_POSTS
+
+        @response_post_count += 1
+        true
+      end
+    end
+
+    # Release a slot reserved by acquire_response_post_slot.
+    # @return [void]
+    def release_response_post_slot
+      @mutex.synchronize { @response_post_count -= 1 }
+    end
   end
 end

data/lib/mcp_client/tool.rb

@@ -38,7 +38,6 @@ module MCPClient
     # @param task_support [String, nil] execution.taskSupport value (MCP 2025-11-25)
     # @param icons [Array<Hash>, nil] optional icons for display in user interfaces (MCP 2025-11-25)
     # @param meta [Hash, nil] optional `_meta` metadata attached to the tool (MCP 2025-11-25)
-    # rubocop:disable Metrics/ParameterLists
     def initialize(name:, description:, schema:, title: nil, output_schema: nil, annotations: nil, server: nil,
                    task_support: nil, icons: nil, meta: nil)
       @name = name
@@ -52,7 +51,6 @@ module MCPClient
       @icons = icons
       @meta = meta
     end
-    # rubocop:enable Metrics/ParameterLists
 
     # Create a Tool instance from JSON data
     # @param data [Hash] JSON data from MCP server

data/lib/mcp_client/version.rb

@@ -2,7 +2,7 @@
 
 module MCPClient
   # Current version of the MCP client gem
-  VERSION = '2.0.0'
+  VERSION = '2.1.0'
 
   # MCP protocol version (date-based) - unified across all transports
   PROTOCOL_VERSION = '2025-11-25'

data/lib/mcp_client.rb

@@ -453,11 +453,17 @@ module MCPClient
   # @param retry_backoff [Integer] Backoff delay in seconds (default: 1)
   # @param name [String, nil] Optional name for this server
   # @param logger [Logger, nil] Optional logger for server operations
+  # @param max_decompressed_body_bytes [Integer] ceiling on how far a gzip-encoded
+  #   response may expand before it is rejected, guarding against a small highly
+  #   compressed body exhausting memory (default: 64 MiB)
   # @yieldparam faraday [Faraday::Connection] the configured connection instance for additional customization
   #   (e.g., SSL settings, custom middleware). The block is called after default configuration is applied.
   # @return [Hash] server configuration
   def self.streamable_http_config(base_url:, endpoint: '/rpc', headers: {}, read_timeout: 30, retries: 3,
-                                  retry_backoff: 1, name: nil, logger: nil, &faraday_config)
+                                  retry_backoff: 1, name: nil, logger: nil,
+                                  max_decompressed_body_bytes:
+                                    MCPClient::ServerStreamableHTTP::JsonRpcTransport::MAX_DECOMPRESSED_BODY_BYTES,
+                                  &faraday_config)
     {
       type: 'streamable_http',
       base_url: base_url,
@@ -468,6 +474,7 @@ module MCPClient
       retry_backoff: retry_backoff,
       name: name,
       logger: logger,
+      max_decompressed_body_bytes: max_decompressed_body_bytes,
       faraday_config: faraday_config
     }
   end

metadata

@@ -1,14 +1,13 @@
 --- !ruby/object:Gem::Specification
 name: ruby-mcp-client
 version: !ruby/object:Gem::Version
-  version: 2.0.0
+  version: 2.1.0
 platform: ruby
 authors:
 - Szymon Kurcab
-autorequire:
 bindir: bin
 cert_chain: []
-date: 2026-07-21 00:00:00.000000000 Z
+date: 1980-01-02 00:00:00.000000000 Z
 dependencies:
 - !ruby/object:Gem::Dependency
   name: base64
@@ -157,6 +156,7 @@ files:
 - lib/mcp_client/server_http/json_rpc_transport.rb
 - lib/mcp_client/server_sse.rb
 - lib/mcp_client/server_sse/json_rpc_transport.rb
+- lib/mcp_client/server_sse/origin_policy.rb
 - lib/mcp_client/server_sse/reconnect_monitor.rb
 - lib/mcp_client/server_sse/sse_parser.rb
 - lib/mcp_client/server_stdio.rb
@@ -171,7 +171,6 @@ licenses:
 - MIT
 metadata:
   rubygems_mfa_required: 'true'
-post_install_message:
 rdoc_options: []
 require_paths:
 - lib
@@ -186,8 +185,7 @@ required_rubygems_version: !ruby/object:Gem::Requirement
     - !ruby/object:Gem::Version
       version: '0'
 requirements: []
-rubygems_version: 3.5.16
-signing_key:
+rubygems_version: 4.0.17
 specification_version: 4
 summary: A Ruby client for the Model Context Protocol (MCP)
 test_files: []