ruby-mcp-client 1.1.0 → 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
@@ -111,12 +158,20 @@ module MCPClient
       @http_conn = nil
       @session_id = nil
       @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
@@ -196,11 +251,7 @@ module MCPClient
     # @raise [MCPClient::Errors::ToolCallError] for other errors during tool execution
     # @raise [MCPClient::Errors::ConnectionError] if server is disconnected
     def call_tool(tool_name, parameters)
-      rpc_request('tools/call', {
-                    name: tool_name,
-                    arguments: parameters.except(:_meta),
-                    **parameters.slice(:_meta)
-                  })
+      rpc_request('tools/call', build_named_request_params(tool_name, parameters))
     rescue MCPClient::Errors::ConnectionError, MCPClient::Errors::TransportError
       # Re-raise connection/transport errors directly to match test expectations
       raise
@@ -226,11 +277,14 @@ module MCPClient
     # @return [Hash] completion result with 'values', optional 'total', and 'hasMore' fields
     # @raise [MCPClient::Errors::ServerError] if server returns an error
     def complete(ref:, argument:, context: nil)
+      ensure_connected
+      require_capability!('completions', method: 'completion/complete')
       params = { ref: ref, argument: argument }
       params[:context] = context if context
       result = rpc_request('completion/complete', params)
       result['completion'] || { 'values' => [] }
-    rescue MCPClient::Errors::ConnectionError, MCPClient::Errors::TransportError
+    rescue MCPClient::Errors::ConnectionError, MCPClient::Errors::TransportError,
+           MCPClient::Errors::CapabilityError
       raise
     rescue StandardError => e
       raise MCPClient::Errors::ServerError, "Error requesting completion: #{e.message}"
@@ -242,8 +296,11 @@ module MCPClient
     # @return [Hash] empty result on success
     # @raise [MCPClient::Errors::ServerError] if server returns an error
     def log_level=(level)
+      ensure_connected
+      require_capability!('logging', method: 'logging/setLevel')
       rpc_request('logging/setLevel', { level: level })
-    rescue MCPClient::Errors::ConnectionError, MCPClient::Errors::TransportError
+    rescue MCPClient::Errors::ConnectionError, MCPClient::Errors::TransportError,
+           MCPClient::Errors::CapabilityError
       raise
     rescue StandardError => e
       raise MCPClient::Errors::ServerError, "Error setting log level: #{e.message}"
@@ -282,11 +339,7 @@ module MCPClient
     # @return [Object] the result of the prompt (with string keys for backward compatibility)
     # @raise [MCPClient::Errors::PromptGetError] if prompt retrieval fails
     def get_prompt(prompt_name, parameters)
-      rpc_request('prompts/get', {
-                    name: prompt_name,
-                    arguments: parameters.except(:_meta),
-                    **parameters.slice(:_meta)
-                  })
+      rpc_request('prompts/get', build_named_request_params(prompt_name, parameters))
     rescue MCPClient::Errors::ConnectionError, MCPClient::Errors::TransportError
       # Re-raise connection/transport errors directly
       raise
@@ -371,9 +424,12 @@ module MCPClient
     # @return [Boolean] true if subscription successful
     # @raise [MCPClient::Errors::ResourceReadError] for other errors during subscription
     def subscribe_resource(uri)
+      ensure_connected
+      require_capability!('resources', 'subscribe', method: 'resources/subscribe')
       rpc_request('resources/subscribe', { uri: uri })
       true
-    rescue MCPClient::Errors::ConnectionError, MCPClient::Errors::TransportError, MCPClient::Errors::ServerError
+    rescue MCPClient::Errors::ConnectionError, MCPClient::Errors::TransportError, MCPClient::Errors::ServerError,
+           MCPClient::Errors::CapabilityError
       raise
     rescue StandardError => e
       raise MCPClient::Errors::ResourceReadError, "Error subscribing to resource '#{uri}': #{e.message}"
@@ -384,9 +440,12 @@ module MCPClient
     # @return [Boolean] true if unsubscription successful
     # @raise [MCPClient::Errors::ResourceReadError] for other errors during unsubscription
     def unsubscribe_resource(uri)
+      ensure_connected
+      require_capability!('resources', 'subscribe', method: 'resources/unsubscribe')
       rpc_request('resources/unsubscribe', { uri: uri })
       true
-    rescue MCPClient::Errors::ConnectionError, MCPClient::Errors::TransportError, MCPClient::Errors::ServerError
+    rescue MCPClient::Errors::ConnectionError, MCPClient::Errors::TransportError, MCPClient::Errors::ServerError,
+           MCPClient::Errors::CapabilityError
       raise
     rescue StandardError => e
       raise MCPClient::Errors::ResourceReadError, "Error unsubscribing from resource '#{uri}': #{e.message}"
@@ -397,23 +456,20 @@ module MCPClient
       super
 
       # Add session and protocol version headers for non-initialize requests
-      if request['method'] != 'initialize'
-        if @session_id
-          req.headers['Mcp-Session-Id'] = @session_id
-          @logger.debug("Adding session header: Mcp-Session-Id: #{@session_id}")
-        end
+      return unless request['method'] != 'initialize'
 
-        if @protocol_version
-          req.headers['Mcp-Protocol-Version'] = @protocol_version
-          @logger.debug("Adding protocol version header: Mcp-Protocol-Version: #{@protocol_version}")
-        end
+      if @session_id
+        req.headers['Mcp-Session-Id'] = @session_id
+        @logger.debug("Adding session header: Mcp-Session-Id: #{@session_id}")
       end
 
-      # Add Last-Event-ID header for resumability (if available)
-      return unless @last_event_id
+      return unless @protocol_version
 
-      req.headers['Last-Event-ID'] = @last_event_id
-      @logger.debug("Adding Last-Event-ID header: #{@last_event_id}")
+      req.headers['Mcp-Protocol-Version'] = @protocol_version
+      @logger.debug("Adding protocol version header: Mcp-Protocol-Version: #{@protocol_version}")
+
+      # NOTE: Last-Event-ID is deliberately NOT sent on POSTs — per SEP-1699,
+      # resumption is always via HTTP GET with Last-Event-ID.
     end
 
     # Override handle_successful_response to capture session ID
@@ -478,6 +534,9 @@ module MCPClient
         @events_connection = nil
         @session_id = nil
         @last_event_id = nil
+        @sse_retry_ms = nil
+        @pending_stream_responses.each_value(&:close)
+        @pending_stream_responses.clear
 
         # Clear cached data
         @tools = nil
@@ -486,7 +545,8 @@ module MCPClient
         @prompts_data = nil
         @resources = nil
         @resources_data = nil
-        @buffer = ''
+        @buffer = +''
+        @buffer_scanned = 0
 
         @logger.info('Cleanup completed')
       end
@@ -528,6 +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',
@@ -538,7 +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
 
@@ -688,7 +760,7 @@ module MCPClient
         break unless @mutex.synchronize { @connection_established }
 
         @logger.info('Events connection closed, reconnecting...')
-        sleep reconnect_delay
+        sleep events_reconnect_delay(reconnect_delay)
         reconnect_delay = [reconnect_delay * 2, SSE_MAX_RECONNECT_DELAY].min
 
       # Intentional shutdown
@@ -697,53 +769,267 @@ module MCPClient
         break unless @mutex.synchronize { @connection_established }
 
         @logger.debug('Events connection timed out after inactivity, reconnecting...')
-        sleep reconnect_delay
+        sleep events_reconnect_delay(reconnect_delay)
       rescue Faraday::ConnectionFailed => e
         break unless @mutex.synchronize { @connection_established }
 
         @logger.warn("Events connection failed: #{e.message}, retrying in #{reconnect_delay}s...")
-        sleep reconnect_delay
+        sleep events_reconnect_delay(reconnect_delay)
         reconnect_delay = [reconnect_delay * 2, SSE_MAX_RECONNECT_DELAY].min
       rescue StandardError => e
         break unless @mutex.synchronize { @connection_established }
 
         @logger.error("Unexpected error in events connection: #{e.class} - #{e.message}")
         @logger.debug(e.backtrace.join("\n")) if @logger.level <= Logger::DEBUG
-        sleep reconnect_delay
+        sleep events_reconnect_delay(reconnect_delay)
         reconnect_delay = [reconnect_delay * 2, SSE_MAX_RECONNECT_DELAY].min
       end
     ensure
       @logger.info('Events connection thread terminated')
     end
 
+    # Reconnect delay for the events stream: the server's SSE retry directive
+    # (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
+      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
+    # delivering it (SEP-1699 polling pattern). A dedicated worker issues
+    # HTTP GETs carrying the disconnected stream's Last-Event-ID cursor — the
+    # general events stream (opened without that cursor) cannot receive the
+    # replay.
+    # @param request_id [Integer, String] id of the outstanding request
+    # @param cursor [String] last event id received on the closed stream
+    # @param retry_ms [Integer, nil] retry directive received on the closed
+    #   stream itself (not the shared events-stream directive)
+    # @return [Hash, nil] the replayed JSON-RPC response, or nil on timeout
+    def resume_response_via_get(request_id, cursor, retry_ms = nil)
+      queue = Thread::Queue.new
+      @mutex.synchronize { @pending_stream_responses[request_id] = queue }
+
+      resume_thread = Thread.new { run_resumption_loop(request_id, cursor, retry_ms) }
+      begin
+        queue.pop(timeout: @read_timeout)
+      ensure
+        @mutex.synchronize { @pending_stream_responses.delete(request_id) }
+        resume_thread.kill
+        resume_thread.join(1)
+      end
+    end
+
+    # Reconnecting worker for SEP-1699 resumption. The server MAY close a
+    # resumed stream again before returning the response (polling pattern),
+    # so each iteration issues a GET with the CURRENT cursor until the
+    # overall read timeout elapses or the waiter has been served. `id:`
+    # fields received on the resumed stream advance the cursor and `retry:`
+    # fields update the delay honored before the next reconnect (zero is a
+    # valid immediate-reconnect directive).
+    # @param request_id [Integer, String] id of the outstanding request
+    # @param cursor [String] Last-Event-ID cursor from the closed stream
+    # @param retry_ms [Integer, nil] retry directive from the closed stream
+    # @return [void]
+    def run_resumption_loop(request_id, cursor, retry_ms)
+      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. 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?
+        issue_resumption_get(state)
+        break unless @mutex.synchronize { @pending_stream_responses.key?(request_id) }
+        break if Process.clock_gettime(Process::CLOCK_MONOTONIC) >= deadline
+
+        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
+    # received on this stream update the resumption state live.
+    # @param state [Hash] mutable resumption state (:cursor, :retry_ms)
+    # @return [void]
+    def issue_resumption_get(state)
+      conn = Faraday.new(url: @base_url) do |f|
+        f.options.open_timeout = 10
+        f.options.timeout = @read_timeout
+        f.adapter :net_http
+      end
+
+      buffer = +''
+      conn.get(@endpoint) do |req|
+        apply_events_headers(req)
+        req.headers['Last-Event-ID'] = state[:cursor]
+        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
+      @logger.debug("Resumption GET failed: #{e.message}")
+    end
+
+    # Extract complete SSE events (terminated by a blank line, LF or CRLF)
+    # from the resumption buffer and handle each of them.
+    # @param buffer [String] mutable stream buffer
+    # @param state [Hash] mutable resumption state (:cursor, :retry_ms)
+    # @return [void]
+    def process_resumption_buffer(buffer, state)
+      # 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
+    # (advancing the cursor used by the next reconnect) and `retry:` lines
+    # (updating the reconnect delay; zero is valid), then dispatch any data.
+    # @param event_text [String] one complete SSE event, separator included
+    # @param state [Hash] mutable resumption state (:cursor, :retry_ms)
+    # @return [void]
+    def handle_resumption_event(event_text, state)
+      data_lines = []
+      event_text.each_line do |raw_line|
+        line = raw_line.chomp
+        if line.start_with?('data:')
+          data_lines << line.sub(/\Adata:\s*/, '')
+        elsif line.start_with?('id:')
+          id = line.sub(/\Aid:\s*/, '').strip
+          # 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/)
+        end
+      end
+      handle_server_message(data_lines.join("\n")) unless data_lines.empty?
+    end
+
+    # Deliver a response replayed on the events stream to a waiting request.
+    # The pending entry is removed on delivery so resumption workers can see
+    # that their waiter has been served.
+    # @param message [Hash] a JSON-RPC response message
+    # @return [void]
+    def deliver_stream_response(message)
+      queue = @mutex.synchronize do
+        key = [message['id'], message['id'].to_s].find { |k| @pending_stream_responses.key?(k) }
+        @pending_stream_responses.delete(key) unless key.nil?
+      end
+      queue << message if queue
+    end
+
     # Apply headers for events connection
     # @param req [Faraday::Request] HTTP request
     def apply_events_headers(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)
+      # SEP-1699: resumption is via GET with the Last-Event-ID cursor, so the
+      # server can replay messages missed since the last received event.
+      last_event_id = @mutex.synchronize { @last_event_id }
+      req.headers['Last-Event-ID'] = last_event_id if last_event_id
     end
 
     # Process event chunks from the server
     # 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
@@ -775,13 +1061,18 @@ 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:')
-          # Server can suggest reconnection delay (in milliseconds)
-          retry_ms = line[6..].strip.to_i
-          @logger.debug("Server suggested retry delay: #{retry_ms}ms") if @logger.level <= Logger::DEBUG
+          # SEP-1699: the client MUST respect the server's retry directive
+          # (milliseconds) when reconnecting; zero is a valid directive.
+          raw = line[6..].strip
+          @sse_retry_ms = raw.to_i if raw.match?(/\A\d+\z/)
+          @logger.debug("Server suggested retry delay: #{raw}ms") if @logger.level <= Logger::DEBUG
         end
       end
 
@@ -799,20 +1090,39 @@ module MCPClient
 
       begin
         message = JSON.parse(data)
-
-        # Handle ping requests from server (keepalive mechanism)
-        if message['method'] == 'ping' && message.key?('id')
-          handle_ping_request(message['id'])
-        elsif message['method'] && message.key?('id')
-          # Handle server-to-client requests (MCP 2025-06-18)
-          handle_server_request(message)
-        elsif message['method'] && !message.key?('id')
-          # Handle server notifications (messages without id)
-          @notification_callback&.call(message['method'], message['params'])
+        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
+
+    # Dispatch a parsed server message (request, ping, or notification).
+    # Used for messages arriving on the GET events stream and for messages
+    # interleaved on a POST SSE response stream.
+    # @param message [Hash] the parsed JSON-RPC message
+    def dispatch_server_message(message)
+      # Handle ping requests from server (keepalive mechanism)
+      if message['method'] == 'ping' && message.key?('id')
+        handle_ping_request(message['id'])
+      elsif message['method'] && message.key?('id')
+        # Handle server-to-client requests (MCP 2025-06-18)
+        handle_server_request(message)
+      elsif message['method'] && !message.key?('id')
+        # Handle server notifications (messages without id)
+        @notification_callback&.call(message['method'], message['params'])
+      elsif message.key?('id')
+        # A response replayed on the events stream after its POST stream was
+        # closed before delivery (SEP-1699 resumption)
+        deliver_stream_response(message)
       end
     end
 
@@ -826,24 +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
-          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)
@@ -868,30 +1163,31 @@ 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 2025-06-18)
-    # @param request_id [String, Integer] the JSON-RPC request ID (used as elicitationId)
+    # Handle elicitation/create request from server (MCP 2025-11-25)
+    # @param request_id [String, Integer] the JSON-RPC request ID
     # @param params [Hash] the elicitation parameters
     # @return [void]
     def handle_elicitation_create(request_id, params)
-      # The request_id is the elicitationId per MCP spec
-      elicitation_id = request_id
-
-      # If no callback is registered, decline the request
+      # Without a callback there is no user to interact with: answer with a
+      # JSON-RPC error rather than fabricating a user "decline".
       unless @elicitation_request_callback
-        @logger.warn('Received elicitation request but no callback registered, declining')
-        send_elicitation_response(elicitation_id, { 'action' => 'decline' })
+        @logger.warn('Received elicitation request but no callback registered')
+        send_error_response(request_id, -32_601, 'Elicitation not supported: no handler configured')
         return
       end
 
       # Call the registered callback
       result = @elicitation_request_callback.call(request_id, params)
 
-      # Send the response back to the server
-      send_elicitation_response(elicitation_id, result)
+      # Send the response back to the server (echoing related-task _meta)
+      send_elicitation_response(request_id, merge_related_task_meta(result, params))
     end
 
     # Handle roots/list request from server (MCP 2025-06-18)
@@ -909,8 +1205,8 @@ module MCPClient
       # Call the registered callback
       result = @roots_list_request_callback.call(request_id, params)
 
-      # Send the response back to the server
-      send_roots_list_response(request_id, result)
+      # Send the response back to the server (echoing related-task _meta)
+      send_roots_list_response(request_id, merge_related_task_meta(result, params))
     end
 
     # Handle sampling/createMessage request from server (MCP 2025-11-25)
@@ -928,8 +1224,8 @@ module MCPClient
       # Call the registered callback
       result = @sampling_request_callback.call(request_id, params)
 
-      # Send the response back to the server
-      send_sampling_response(request_id, result)
+      # Send the response back to the server (echoing related-task _meta)
+      send_sampling_response(request_id, merge_related_task_meta(result, params))
     end
 
     # Send roots/list response back to server via HTTP POST (MCP 2025-06-18)
@@ -972,29 +1268,29 @@ module MCPClient
       @logger.error("Error sending sampling response: #{e.message}")
     end
 
-    # Send elicitation response back to server via HTTP POST (MCP 2025-06-18)
-    # For streamable HTTP, this is sent as a JSON-RPC request (not response)
-    # because HTTP is unidirectional.
-    # @param elicitation_id [String] the elicitation ID from the server
+    # Send elicitation response back to server via HTTP POST (MCP 2025-11-25)
+    # The reply to a server's elicitation/create request is a standard
+    # JSON-RPC response echoing the request id, POSTed like every other
+    # response on this transport.
+    # @param request_id [String, Integer] the JSON-RPC request ID
     # @param result [Hash] the elicitation result (action and optional content)
     # @return [void]
-    def send_elicitation_response(elicitation_id, result)
-      params = {
-        'elicitationId' => elicitation_id,
-        'action' => result['action']
-      }
-
-      # Only include content if present (typically for 'accept' action)
-      params['content'] = result['content'] if result['content']
+    def send_elicitation_response(request_id, result)
+      # Error-shaped results become JSON-RPC error responses (e.g. -32602 for
+      # an undeclared elicitation mode), mirroring the sampling error path.
+      if result.is_a?(Hash) && result['error']
+        send_error_response(request_id, result['error']['code'] || -32_603,
+                            result['error']['message'] || 'Elicitation error')
+        return
+      end
 
-      request = {
+      response = {
         'jsonrpc' => '2.0',
-        'method' => 'elicitation/response',
-        'params' => params
+        'id' => request_id,
+        'result' => result
       }
 
-      # Send as a JSON-RPC request via HTTP POST
-      post_jsonrpc_response(request)
+      post_jsonrpc_response(response)
     rescue StandardError => e
       @logger.error("Error sending elicitation response: #{e.message}")
     end
@@ -1025,7 +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)
@@ -1034,17 +1354,61 @@ module MCPClient
           @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 = json_body
         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