swarm_sdk 2.7.9 → 2.7.11

data/lib/swarm_sdk/ruby_llm_patches/responses_api_patch.rb

@@ -0,0 +1,599 @@
+# frozen_string_literal: true
+
+# Adds OpenAI Responses API support to RubyLLM
+# - ResponsesSession class for session management
+# - OpenAIResponses provider (uses v1/responses endpoint)
+# - Chat methods: with_responses_api, restore_responses_session, responses_api_enabled?
+#
+# Fork Reference: Commit 5f77903
+
+module RubyLLM
+  # Responses API specific errors
+  class ResponsesApiError < Error; end
+  class ResponseIdNotFoundError < ResponsesApiError; end
+  class ResponseFailedError < ResponsesApiError; end
+  class ResponseInProgressError < ResponsesApiError; end
+  class ResponseCancelledError < ResponsesApiError; end
+  class ResponseIncompleteError < ResponsesApiError; end
+
+  # Manages state for OpenAI Responses API stateful conversations.
+  # Tracks response IDs, session validity, and failure recovery.
+  class ResponsesSession
+    RESPONSE_ID_TTL = 300 # 5 minutes
+    MAX_FAILURES = 2
+
+    attr_reader :response_id, :last_activity, :failure_count
+
+    def initialize(response_id: nil, last_activity: nil, failure_count: 0, disabled: false)
+      @response_id = response_id
+      @last_activity = last_activity
+      @failure_count = failure_count
+      @disabled = disabled
+    end
+
+    def reset!
+      @response_id = nil
+      @last_activity = nil
+      @failure_count = 0
+      @disabled = false
+    end
+
+    def update(new_response_id)
+      @response_id = new_response_id
+      @last_activity = Time.now
+      @failure_count = 0
+    end
+
+    def valid?
+      return false if @disabled
+      return false unless @response_id
+      return false unless @last_activity
+
+      (Time.now - @last_activity) < RESPONSE_ID_TTL
+    end
+
+    def record_failure!
+      @failure_count += 1
+
+      if @failure_count >= MAX_FAILURES
+        @disabled = true
+      else
+        @response_id = nil
+        @last_activity = nil
+      end
+    end
+
+    def disabled?
+      @disabled
+    end
+
+    def to_h
+      {
+        response_id: @response_id,
+        last_activity: @last_activity&.iso8601,
+        failure_count: @failure_count,
+        disabled: @disabled,
+      }
+    end
+
+    def self.from_h(hash) # rubocop:disable Style/ClassMethodsDefinitions
+      hash = hash.transform_keys(&:to_sym)
+      last_activity = hash[:last_activity] ? Time.parse(hash[:last_activity]) : nil
+
+      new(
+        response_id: hash[:response_id],
+        last_activity: last_activity,
+        failure_count: hash[:failure_count] || 0,
+        disabled: hash[:disabled] || false,
+      )
+    end
+  end
+
+  module Providers
+    # OpenAI Responses API provider.
+    # Uses v1/responses endpoint instead of v1/chat/completions.
+    # Inherits from OpenAI and overrides only what differs.
+    class OpenAIResponses < OpenAI
+      attr_reader :responses_session, :responses_config
+
+      def initialize(config, responses_session = nil, responses_config = {})
+        @responses_session = responses_session || ResponsesSession.new
+        @responses_config = {
+          stateful: false,
+          store: true,
+          truncation: :disabled,
+          include: [],
+        }.merge(responses_config)
+
+        super(config)
+      end
+
+      # Override endpoint URL
+      def completion_url
+        "responses"
+      end
+
+      # Override complete to handle response ID failures
+      def complete(messages, tools:, temperature:, model:, params: {}, headers: {}, schema: nil, thinking: nil, &block)
+        super
+      rescue BadRequestError => e
+        raise unless response_id_not_found_error?(e)
+
+        handle_response_id_failure
+        retry
+      end
+
+      # Override render_payload for Responses API format
+      def render_payload(messages, tools:, temperature:, model:, stream: false, schema: nil, thinking: nil)
+        system_msgs, other_msgs = partition_messages(messages)
+
+        payload = build_base_payload(model, stream)
+        add_instructions(payload, system_msgs)
+        add_input(payload, other_msgs)
+        add_temperature(payload, temperature)
+        add_tools(payload, tools)
+        add_schema(payload, schema)
+        add_optional_parameters(payload)
+        add_stream_options(payload, stream)
+
+        payload
+      end
+
+      # Override parse_completion_response for Responses API format
+      def parse_completion_response(response)
+        data = response.body
+        return if data.nil? || !data.is_a?(Hash) || data.empty?
+
+        case data["status"]
+        when "completed"
+          parse_completed_response(data, response)
+        when "failed"
+          raise ResponseFailedError.new(response, data.dig("error", "message") || "Response failed")
+        when "in_progress", "queued"
+          raise ResponseInProgressError.new(response, "Response still processing: #{data["id"]}")
+        when "cancelled"
+          raise ResponseCancelledError.new(response, "Response was cancelled: #{data["id"]}")
+        when "incomplete"
+          parse_incomplete_response(data, response)
+        else
+          raise Error.new(response, data.dig("error", "message")) if data.dig("error", "message")
+
+          parse_completed_response(data, response)
+        end
+      end
+
+      # Override tool_for for flat format (not nested under 'function')
+      def tool_for(tool)
+        parameters_schema = parameters_schema_for(tool)
+
+        definition = {
+          type: "function",
+          name: tool.name,
+          description: tool.description,
+          parameters: parameters_schema,
+        }
+
+        return definition if tool.provider_params.empty?
+
+        RubyLLM::Utils.deep_merge(definition, tool.provider_params)
+      end
+
+      # Override build_chunk for Responses API streaming events
+      def build_chunk(data)
+        if responses_api_event?(data)
+          build_responses_chunk(data)
+        else
+          super
+        end
+      end
+
+      private
+
+      def stateful_mode?
+        @responses_config[:stateful] == true
+      end
+
+      def partition_messages(messages)
+        system_msgs = messages.select { |m| m.role == :system }
+        other_msgs = messages.reject { |m| m.role == :system }
+        [system_msgs, other_msgs]
+      end
+
+      def build_base_payload(model, stream)
+        {
+          model: model.id,
+          stream: stream,
+          store: @responses_config[:store],
+        }
+      end
+
+      def add_instructions(payload, system_msgs)
+        payload[:instructions] = format_instructions(system_msgs) if system_msgs.any?
+      end
+
+      def add_input(payload, other_msgs)
+        if stateful_mode? && @responses_session.valid?
+          payload[:previous_response_id] = @responses_session.response_id
+          payload[:input] = format_new_input_only(other_msgs)
+        else
+          payload[:input] = format_responses_input(other_msgs)
+        end
+      end
+
+      def add_temperature(payload, temperature)
+        payload[:temperature] = temperature unless temperature.nil?
+      end
+
+      def add_tools(payload, tools)
+        payload[:tools] = tools.map { |_, tool| tool_for(tool) } if tools.any?
+      end
+
+      def add_schema(payload, schema)
+        return unless schema
+
+        payload[:text] = {
+          format: {
+            type: "json_schema",
+            name: "response",
+            schema: schema,
+            strict: schema[:strict] != false,
+          },
+        }
+      end
+
+      # Responses API does not support stream_options (Chat Completions API only).
+      # Usage is included automatically in the response.completed streaming event.
+      def add_stream_options(payload, stream)
+      end
+
+      def response_id_not_found_error?(error)
+        return false unless @responses_session.response_id
+
+        error.message.include?("not found")
+      end
+
+      def handle_response_id_failure
+        @responses_session.record_failure!
+
+        if @responses_session.disabled?
+          RubyLLM.logger.warn("Responses API: Disabling stateful mode after repeated failures")
+        else
+          RubyLLM.logger.debug("Responses API: Response ID not found, retrying fresh")
+        end
+      end
+
+      def format_instructions(system_messages)
+        system_messages.map { |m| m.content.to_s }.join("\n\n")
+      end
+
+      def format_responses_input(messages)
+        messages.each_with_object([]) do |msg, items|
+          case msg.role
+          when :user
+            items << {
+              type: "message",
+              role: "user",
+              content: format_input_content(msg.content),
+            }
+          when :assistant
+            format_assistant_input(msg, items)
+          when :tool
+            items << {
+              type: "function_call_output",
+              call_id: msg.tool_call_id,
+              output: msg.content.to_s,
+            }
+          end
+        end
+      end
+
+      def format_assistant_input(msg, items)
+        # Emit function_call items for each tool call (required before function_call_output)
+        if msg.tool_calls && !msg.tool_calls.empty?
+          msg.tool_calls.each_value do |tc|
+            items << {
+              type: "function_call",
+              call_id: tc.id,
+              name: tc.name,
+              arguments: tc.arguments.is_a?(String) ? tc.arguments : tc.arguments.to_json,
+            }
+          end
+        end
+
+        # Emit assistant message content if present
+        return if msg.content.nil? || msg.content.to_s.strip.empty?
+
+        items << {
+          type: "message",
+          role: "assistant",
+          content: format_output_content(msg.content),
+        }
+      end
+
+      def format_new_input_only(messages)
+        formatted = []
+        last_assistant_idx = messages.rindex { |msg| msg.role == :assistant }
+
+        if last_assistant_idx
+          new_messages = messages[(last_assistant_idx + 1)..]
+          new_messages.each do |msg|
+            case msg.role
+            when :tool
+              formatted << {
+                type: "function_call_output",
+                call_id: msg.tool_call_id,
+                output: msg.content.to_s,
+              }
+            when :user
+              formatted << {
+                type: "message",
+                role: "user",
+                content: format_input_content(msg.content),
+              }
+            end
+          end
+        else
+          messages.each do |msg|
+            next unless msg.role == :user
+
+            formatted << {
+              type: "message",
+              role: "user",
+              content: format_input_content(msg.content),
+            }
+          end
+        end
+
+        formatted
+      end
+
+      def format_input_content(content)
+        case content
+        when String
+          [{ type: "input_text", text: content }]
+        when Content
+          parts = []
+          parts << { type: "input_text", text: content.text } if content.text && !content.text.empty?
+          content.attachments.each do |attachment|
+            parts << format_input_attachment(attachment)
+          end
+          parts
+        when Content::Raw
+          content.value
+        else
+          [{ type: "input_text", text: content.to_s }]
+        end
+      end
+
+      def format_output_content(content)
+        if content.is_a?(String)
+          [{ type: "output_text", text: content }]
+        elsif content.is_a?(Content)
+          [{ type: "output_text", text: content.text || "" }]
+        else
+          [{ type: "output_text", text: content.to_s }]
+        end
+      end
+
+      def format_input_attachment(attachment)
+        case attachment.type
+        when :image
+          if attachment.url?
+            { type: "input_image", image_url: attachment.source.to_s }
+          else
+            { type: "input_image", image_url: attachment.for_llm }
+          end
+        when :file, :pdf
+          { type: "input_file", file_data: attachment.encoded, filename: attachment.filename }
+        else
+          { type: "input_text", text: "[Unsupported attachment: #{attachment.type}]" }
+        end
+      end
+
+      def add_optional_parameters(payload)
+        if @responses_config[:truncation] && @responses_config[:truncation] != :disabled
+          payload[:truncation] = @responses_config[:truncation].to_s
+        end
+
+        if @responses_config[:include] && !@responses_config[:include].empty?
+          payload[:include] = @responses_config[:include].map { |i| i.to_s.tr("_", ".") }
+        end
+
+        payload[:service_tier] = @responses_config[:service_tier].to_s if @responses_config[:service_tier]
+        payload[:max_tool_calls] = @responses_config[:max_tool_calls] if @responses_config[:max_tool_calls]
+      end
+
+      def parse_completed_response(data, response)
+        output = data["output"] || []
+        content_parts = []
+        tool_calls = {}
+
+        output.each do |item|
+          case item["type"]
+          when "message"
+            content_parts << extract_message_content(item)
+          when "function_call"
+            tool_calls[item["call_id"]] = ToolCall.new(
+              id: item["call_id"],
+              name: item["name"],
+              arguments: parse_tool_arguments(item["arguments"]),
+            )
+          end
+        end
+
+        usage = data["usage"] || {}
+
+        Message.new(
+          role: :assistant,
+          content: content_parts.join("\n"),
+          tool_calls: tool_calls.empty? ? nil : tool_calls,
+          response_id: data["id"],
+          reasoning_summary: data.dig("reasoning", "summary"),
+          reasoning_tokens: usage.dig("output_tokens_details", "reasoning_tokens"),
+          input_tokens: usage["input_tokens"] || 0,
+          output_tokens: usage["output_tokens"] || 0,
+          cached_tokens: usage.dig("prompt_tokens_details", "cached_tokens"),
+          cache_creation_tokens: 0,
+          model_id: data["model"],
+          raw: response,
+        )
+      end
+
+      def parse_tool_arguments(arguments)
+        if arguments.nil? || arguments.empty?
+          {}
+        elsif arguments.is_a?(String)
+          JSON.parse(arguments)
+        else
+          arguments
+        end
+      rescue JSON::ParserError
+        {}
+      end
+
+      def parse_incomplete_response(data, response)
+        message = parse_completed_response(data, response)
+        RubyLLM.logger.warn("Responses API: Incomplete response: #{data["incomplete_details"]}")
+        message
+      end
+
+      def extract_message_content(item)
+        return "" unless item["content"].is_a?(Array)
+
+        item["content"].filter_map do |content_item|
+          content_item["text"] if content_item["type"] == "output_text"
+        end.join
+      end
+
+      def responses_api_event?(data)
+        data.is_a?(Hash) && data["type"]&.start_with?("response.")
+      end
+
+      def build_responses_chunk(data)
+        case data["type"]
+        when "response.output_text.delta"
+          # delta is a plain string in the Responses API
+          Chunk.new(
+            role: :assistant,
+            content: data["delta"] || "",
+            model_id: nil,
+            input_tokens: nil,
+            output_tokens: nil,
+          )
+        when "response.completed"
+          usage = data.dig("response", "usage") || {}
+          Chunk.new(
+            role: :assistant,
+            content: nil,
+            model_id: data.dig("response", "model"),
+            input_tokens: usage["input_tokens"],
+            output_tokens: usage["output_tokens"],
+            cached_tokens: usage.dig("prompt_tokens_details", "cached_tokens"),
+            cache_creation_tokens: 0,
+          )
+        when "response.function_call_arguments.delta"
+          # Tool call argument streaming - accumulate via tool_calls
+          Chunk.new(
+            role: :assistant,
+            content: nil,
+            model_id: nil,
+            input_tokens: nil,
+            output_tokens: nil,
+            tool_calls: parse_function_call_delta(data),
+          )
+        when "response.output_item.done"
+          build_output_item_done_chunk(data)
+        else
+          # No-op chunk for unrecognized events (response.created, response.in_progress, etc.)
+          Chunk.new(role: :assistant, content: nil, model_id: nil, input_tokens: nil, output_tokens: nil)
+        end
+      end
+
+      def parse_function_call_delta(data)
+        call_id = data["call_id"] || data["item_id"]
+        return {} unless call_id
+
+        { call_id => ToolCall.new(id: nil, name: nil, arguments: data["delta"] || "") }
+      end
+
+      def build_output_item_done_chunk(data)
+        item = data["item"] || {}
+        if item["type"] == "function_call"
+          tool_calls = {
+            item["call_id"] => ToolCall.new(
+              id: item["call_id"],
+              name: item["name"],
+              arguments: parse_tool_arguments(item["arguments"]),
+            ),
+          }
+          Chunk.new(
+            role: :assistant,
+            content: nil,
+            model_id: nil,
+            input_tokens: nil,
+            output_tokens: nil,
+            tool_calls: tool_calls,
+          )
+        else
+          Chunk.new(role: :assistant, content: nil, model_id: nil, input_tokens: nil, output_tokens: nil)
+        end
+      end
+    end
+  end
+
+  class Chat
+    # Enable OpenAI Responses API for this chat
+    #
+    # @param stateful [Boolean] Enable stateful mode (uses previous_response_id)
+    # @param store [Boolean] Store responses for retrieval
+    # @param truncation [Symbol] Truncation strategy (:disabled, :auto)
+    # @param include [Array] Additional data to include in responses
+    # @return [self] for chaining
+    def with_responses_api(stateful: false, store: true, truncation: :disabled, include: [])
+      responses_config = {
+        stateful: stateful,
+        store: store,
+        truncation: truncation,
+        include: include,
+      }
+
+      @responses_session ||= ResponsesSession.new
+      @provider = Providers::OpenAIResponses.new(@config, @responses_session, responses_config)
+      @connection = @provider.connection
+      @responses_api_enabled = true
+      self
+    end
+
+    # Restore a Responses API session from saved state
+    #
+    # @param session_hash [Hash] Session state from ResponsesSession#to_h
+    # @return [self] for chaining
+    def restore_responses_session(session_hash)
+      @responses_session = ResponsesSession.from_h(session_hash)
+
+      if @provider.is_a?(Providers::OpenAIResponses)
+        # Re-create provider with restored session
+        @provider = Providers::OpenAIResponses.new(
+          @config,
+          @responses_session,
+          @provider.responses_config,
+        )
+        @connection = @provider.connection
+      end
+
+      self
+    end
+
+    # Check if Responses API is enabled
+    #
+    # @return [Boolean] true if Responses API is enabled
+    def responses_api_enabled?
+      @responses_api_enabled == true
+    end
+
+    # Get the current Responses API session
+    #
+    # @return [ResponsesSession, nil] The session if Responses API is enabled
+    attr_reader :responses_session
+  end
+end