claude_swarm 1.0.1 → 1.0.4

data/lib/swarm_memory/integration/sdk_plugin.rb

@@ -0,0 +1,531 @@
+# frozen_string_literal: true
+
+module SwarmMemory
+  module Integration
+    # SwarmSDK plugin implementation for SwarmMemory
+    #
+    # This plugin integrates SwarmMemory with SwarmSDK, providing:
+    # - Persistent memory storage for agents
+    # - Memory tools (MemoryWrite, MemoryRead, MemoryEdit, etc.)
+    # - LoadSkill tool for dynamic tool swapping
+    # - System prompt contributions for memory guidance
+    # - Semantic skill discovery on user messages
+    #
+    # The plugin automatically registers itself when SwarmMemory is loaded
+    # alongside SwarmSDK.
+    class SDKPlugin < SwarmSDK::Plugin
+      def initialize
+        super
+        # Track storages for each agent: { agent_name => storage }
+        # Needed for semantic skill discovery in on_user_message
+        @storages = {}
+        # Track memory mode for each agent: { agent_name => mode }
+        # Modes: :assistant (default), :retrieval, :researcher
+        @modes = {}
+      end
+
+      # Plugin identifier
+      #
+      # @return [Symbol] Plugin name
+      def name
+        :memory
+      end
+
+      # Tools provided by this plugin
+      #
+      # Returns all memory tools for PluginRegistry mapping.
+      # Tools are auto-registered by ToolConfigurator, then filtered
+      # by mode in on_agent_initialized using remove_tool.
+      #
+      # Note: LoadSkill is NOT included here because it requires special handling.
+      # It's registered separately in on_agent_initialized lifecycle hook because
+      # it needs chat, tool_configurator, and agent_definition parameters.
+      #
+      # @return [Array<Symbol>] All memory tool names
+      def tools
+        [
+          :MemoryRead,
+          :MemoryGlob,
+          :MemoryGrep,
+          :MemoryWrite,
+          :MemoryEdit,
+          :MemoryMultiEdit,
+          :MemoryDelete,
+          :MemoryDefrag,
+        ]
+      end
+
+      # Get tools for a specific mode
+      #
+      # @param mode [Symbol] Memory mode
+      # @return [Array<Symbol>] Tool names for this mode
+      def tools_for_mode(mode)
+        case mode
+        when :retrieval
+          # Read-only tools for Q&A agents
+          [:MemoryRead, :MemoryGlob, :MemoryGrep]
+        when :assistant
+          # Read + Write + Edit for learning assistants (need edit for corrections)
+          [:MemoryRead, :MemoryGlob, :MemoryGrep, :MemoryWrite, :MemoryEdit]
+        when :researcher
+          # All tools for knowledge extraction
+          [
+            :MemoryRead,
+            :MemoryGlob,
+            :MemoryGrep,
+            :MemoryWrite,
+            :MemoryEdit,
+            :MemoryMultiEdit,
+            :MemoryDelete,
+            :MemoryDefrag,
+          ]
+        else
+          # Default to assistant
+          [:MemoryRead, :MemoryGlob, :MemoryGrep, :MemoryWrite, :MemoryEdit]
+        end
+      end
+
+      # Create a tool instance
+      #
+      # @param tool_name [Symbol] Tool name
+      # @param context [Hash] Creation context with :storage, :agent_name, :chat, etc.
+      # @return [RubyLLM::Tool] Tool instance
+      def create_tool(tool_name, context)
+        storage = context[:storage]
+        agent_name = context[:agent_name]
+
+        # Delegate to SwarmMemory's tool factory
+        SwarmMemory.create_tool(tool_name, storage: storage, agent_name: agent_name)
+      end
+
+      # Create plugin storage for an agent
+      #
+      # @param agent_name [Symbol] Agent identifier
+      # @param config [Object] Memory configuration (MemoryConfig or Hash)
+      # @return [Core::Storage] Storage instance with embeddings enabled
+      def create_storage(agent_name:, config:)
+        # Extract adapter type and options from config
+        adapter_type, adapter_options = if config.respond_to?(:adapter_type)
+          # MemoryConfig object (from DSL)
+          [config.adapter_type, config.adapter_options]
+        elsif config.is_a?(Hash)
+          # Hash (from YAML)
+          adapter = (config[:adapter] || config["adapter"] || :filesystem).to_sym
+          options = config.reject { |k, _v| k == :adapter || k == "adapter" || k == :mode || k == "mode" }
+          [adapter, options]
+        else
+          raise SwarmSDK::ConfigurationError, "Invalid memory configuration for #{agent_name}"
+        end
+
+        # Get adapter class from registry
+        begin
+          adapter_class = SwarmMemory.adapter_for(adapter_type)
+        rescue ArgumentError => e
+          raise SwarmSDK::ConfigurationError, "#{e.message} for agent #{agent_name}"
+        end
+
+        # Instantiate adapter with options
+        # Note: Adapter is responsible for validating its own requirements
+        begin
+          adapter = adapter_class.new(**adapter_options)
+        rescue ArgumentError => e
+          raise SwarmSDK::ConfigurationError,
+            "Failed to initialize #{adapter_type} adapter for #{agent_name}: #{e.message}"
+        end
+
+        # Create embedder for semantic search
+        embedder = Embeddings::InformersEmbedder.new
+
+        # Create storage with embedder (enables semantic features)
+        Core::Storage.new(adapter: adapter, embedder: embedder)
+      end
+
+      # Parse memory configuration
+      #
+      # @param raw_config [Object] Raw config (MemoryConfig or Hash)
+      # @return [Object] Parsed configuration
+      def parse_config(raw_config)
+        # Already parsed by Agent::Definition, just return as-is
+        raw_config
+      end
+
+      # Contribute to agent system prompt
+      #
+      # @param agent_definition [Agent::Definition] Agent definition
+      # @param storage [Core::Storage, nil] Storage instance (may be nil during prompt building)
+      # @return [String] Memory prompt contribution
+      def system_prompt_contribution(agent_definition:, storage:)
+        # Extract mode from memory config
+        memory_config = agent_definition.memory
+        mode = if memory_config.is_a?(SwarmMemory::DSL::MemoryConfig)
+          memory_config.mode # MemoryConfig object from DSL
+        elsif memory_config.respond_to?(:mode)
+          memory_config.mode # Other object with mode method
+        elsif memory_config.is_a?(Hash)
+          (memory_config[:mode] || memory_config["mode"] || :assistant).to_sym
+        else
+          :assistant # Default mode
+        end
+
+        # Select prompt template based on mode
+        prompt_filename = case mode
+        when :retrieval then "memory_retrieval.md.erb"
+        when :researcher then "memory_researcher.md.erb"
+        else "memory_assistant.md.erb" # Default
+        end
+
+        memory_prompt_path = File.expand_path("../prompts/#{prompt_filename}", __dir__)
+        template_content = File.read(memory_prompt_path)
+
+        # Render with agent_definition binding
+        ERB.new(template_content).result(agent_definition.instance_eval { binding })
+      end
+
+      # Tools that should be marked immutable (mode-aware)
+      #
+      # Memory tools for the current mode plus LoadSkill (if applicable) are immutable.
+      # This prevents LoadSkill from accidentally removing memory tools.
+      #
+      # @param mode [Symbol] Memory mode
+      # @return [Array<Symbol>] Immutable tool names for this mode
+      def immutable_tools_for_mode(mode)
+        base_tools = tools_for_mode(mode)
+
+        # LoadSkill only for assistant and researcher modes (not retrieval)
+        if mode == :retrieval
+          base_tools
+        else
+          base_tools + [:LoadSkill]
+        end
+      end
+
+      # Check if storage should be created for this agent
+      #
+      # @param agent_definition [Agent::Definition] Agent definition
+      # @return [Boolean] True if agent has memory configuration
+      def storage_enabled?(agent_definition)
+        agent_definition.memory_enabled?
+      end
+
+      # Lifecycle: Agent initialized
+      #
+      # Filters tools by mode (removing non-mode tools), registers LoadSkill,
+      # and marks memory tools as immutable.
+      #
+      # LoadSkill needs special handling because it requires chat, tool_configurator,
+      # and agent_definition to perform dynamic tool swapping.
+      #
+      # @param agent_name [Symbol] Agent identifier
+      # @param agent [Agent::Chat] Chat instance
+      # @param context [Hash] Initialization context
+      def on_agent_initialized(agent_name:, agent:, context:)
+        storage = context[:storage]
+        agent_definition = context[:agent_definition]
+        tool_configurator = context[:tool_configurator]
+
+        return unless storage # Only proceed if memory is enabled for this agent
+
+        # Extract mode from memory config
+        memory_config = agent_definition.memory
+        mode = if memory_config.is_a?(SwarmMemory::DSL::MemoryConfig)
+          memory_config.mode # MemoryConfig object from DSL
+        elsif memory_config.respond_to?(:mode)
+          memory_config.mode # Other object with mode method
+        elsif memory_config.is_a?(Hash)
+          (memory_config[:mode] || memory_config["mode"] || :interactive).to_sym
+        else
+          :interactive # Default
+        end
+
+        # Store storage and mode for this agent
+        @storages[agent_name] = storage
+        @modes[agent_name] = mode
+
+        # Get mode-specific tools
+        allowed_tools = tools_for_mode(mode)
+
+        # Get all registered memory tool names
+        all_memory_tools = tools # Returns all possible memory tools
+
+        # Remove tools not allowed in this mode
+        tools_to_remove = all_memory_tools - allowed_tools
+
+        tools_to_remove.each do |tool_name|
+          agent.remove_tool(tool_name)
+        end
+
+        # Create and register LoadSkill tool (NOT for retrieval mode - read-only)
+        unless mode == :retrieval
+          load_skill_tool = SwarmMemory.create_tool(
+            :LoadSkill,
+            storage: storage,
+            agent_name: agent_name,
+            chat: agent,
+            tool_configurator: tool_configurator,
+            agent_definition: agent_definition,
+          )
+
+          agent.with_tool(load_skill_tool)
+        end
+
+        # Mark mode-specific memory tools + LoadSkill as immutable
+        agent.mark_tools_immutable(immutable_tools_for_mode(mode).map(&:to_s))
+      end
+
+      # Lifecycle: User message
+      #
+      # Performs TWO semantic searches:
+      # 1. Skills - For loadable procedures with LoadSkill
+      # 2. Memories - For concepts/facts/experiences that provide context
+      #
+      # Returns system reminders for both if high-confidence matches found.
+      #
+      # @param agent_name [Symbol] Agent identifier
+      # @param prompt [String] User's message
+      # @param is_first_message [Boolean] True if first message
+      # @return [Array<String>] System reminders (0-2 reminders)
+      def on_user_message(agent_name:, prompt:, is_first_message:)
+        storage = @storages[agent_name]
+        return [] unless storage&.semantic_index
+
+        # Adaptive threshold based on query length
+        # Short queries use lower threshold as they have less semantic richness
+        # Optimal: cutoff=10 words, short=0.25, normal=0.35 (discovered via systematic evaluation)
+        word_count = prompt.split.size
+        word_cutoff = (ENV["SWARM_MEMORY_ADAPTIVE_WORD_CUTOFF"] || "10").to_i
+
+        threshold = if word_count < word_cutoff
+          (ENV["SWARM_MEMORY_DISCOVERY_THRESHOLD_SHORT"] || "0.25").to_f
+        else
+          (ENV["SWARM_MEMORY_DISCOVERY_THRESHOLD"] || "0.35").to_f
+        end
+        reminders = []
+
+        # Run both searches in parallel with Async
+        Async do |task|
+          # Search 1: Skills (type = "skill")
+          skills_task = task.async do
+            storage.semantic_index.search(
+              query: prompt,
+              top_k: 3,
+              threshold: threshold,
+              filter: { "type" => "skill" },
+            )
+          end
+
+          # Search 2: All results (for memories + logging)
+          all_results_task = task.async do
+            storage.semantic_index.search(
+              query: prompt,
+              top_k: 10,
+              threshold: 0.0, # Get all for logging
+              filter: nil,
+            )
+          end
+
+          # Wait for both searches to complete
+          skills = skills_task.wait
+          all_results = all_results_task.wait
+
+          # Filter to concepts, facts, experiences (not skills)
+          memories = all_results
+            .select { |r| ["concept", "fact", "experience"].include?(r.dig(:metadata, "type")) }
+            .select { |r| r[:similarity] >= threshold }
+            .take(3)
+
+          # Emit log events (include word count for adaptive threshold analysis)
+          search_context = { threshold: threshold, word_count: word_count, word_cutoff: word_cutoff }
+          emit_skill_search_log(agent_name, prompt, skills, all_results, search_context)
+          emit_memory_search_log(agent_name, prompt, memories, all_results, search_context)
+
+          # Build skill reminder if found
+          if skills.any?
+            reminders << build_skill_discovery_reminder(skills)
+          end
+
+          # Build memory reminder if found
+          if memories.any?
+            reminders << build_memory_discovery_reminder(memories)
+          end
+        end.wait
+
+        reminders
+      end
+
+      private
+
+      # Emit log event for semantic skill search
+      #
+      # @param agent_name [Symbol] Agent identifier
+      # @param prompt [String] User's message
+      # @param skills [Array<Hash>] Found skills (filtered)
+      # @param all_results [Array<Hash>] All search results (unfiltered)
+      # @param search_context [Hash] Search context with :threshold and :word_count
+      # @return [void]
+      def emit_skill_search_log(agent_name, prompt, skills, all_results, search_context)
+        return unless SwarmSDK::LogStream.enabled?
+
+        threshold = search_context[:threshold]
+        word_count = search_context[:word_count]
+        word_cutoff = search_context[:word_cutoff]
+
+        # Include top 5 results for debugging (even if below threshold or wrong type)
+        all_entries_debug = all_results.take(5).map do |result|
+          {
+            path: result[:path],
+            title: result[:title],
+            hybrid_score: result[:similarity].round(3),
+            semantic_score: result[:semantic_score]&.round(3),
+            keyword_score: result[:keyword_score]&.round(3),
+            type: result.dig(:metadata, "type"),
+            tags: result.dig(:metadata, "tags"),
+          }
+        end
+
+        # Get actual weights being used (from ENV or defaults)
+        semantic_weight = (ENV["SWARM_MEMORY_SEMANTIC_WEIGHT"] || "0.5").to_f
+        keyword_weight = (ENV["SWARM_MEMORY_KEYWORD_WEIGHT"] || "0.5").to_f
+
+        SwarmSDK::LogStream.emit(
+          type: "semantic_skill_search",
+          agent: agent_name,
+          query: prompt,
+          query_word_count: word_count,
+          threshold: threshold,
+          threshold_type: word_count < word_cutoff ? "short_query" : "normal_query",
+          adaptive_cutoff: word_cutoff,
+          skills_found: skills.size,
+          total_entries_searched: all_results.size,
+          search_mode: "hybrid",
+          weights: { semantic: semantic_weight, keyword: keyword_weight },
+          skills: skills.map do |skill|
+            {
+              path: skill[:path],
+              title: skill[:title],
+              hybrid_score: skill[:similarity].round(3),
+              semantic_score: skill[:semantic_score]&.round(3),
+              keyword_score: skill[:keyword_score]&.round(3),
+            }
+          end,
+          debug_top_results: all_entries_debug,
+        )
+      end
+
+      # Emit log event for semantic memory search
+      #
+      # @param agent_name [Symbol] Agent identifier
+      # @param prompt [String] User's message
+      # @param memories [Array<Hash>] Found memories (concepts/facts/experiences)
+      # @param all_results [Array<Hash>] All search results (unfiltered)
+      # @param search_context [Hash] Search context with :threshold and :word_count
+      # @return [void]
+      def emit_memory_search_log(agent_name, prompt, memories, all_results, search_context)
+        return unless SwarmSDK::LogStream.enabled?
+
+        threshold = search_context[:threshold]
+        word_count = search_context[:word_count]
+        word_cutoff = search_context[:word_cutoff]
+
+        # Filter all_results to only concept/fact/experience types for debug output
+        memory_entries = all_results.select do |r|
+          ["concept", "fact", "experience"].include?(r.dig(:metadata, "type"))
+        end
+
+        # Include top 10 memory entries for debugging (even if below threshold)
+        debug_all_memories = memory_entries.take(10).map do |result|
+          {
+            path: result[:path],
+            title: result[:title],
+            hybrid_score: result[:similarity].round(3),
+            semantic_score: result[:semantic_score]&.round(3),
+            keyword_score: result[:keyword_score]&.round(3),
+            type: result.dig(:metadata, "type"),
+            tags: result.dig(:metadata, "tags"),
+            domain: result.dig(:metadata, "domain"),
+          }
+        end
+
+        # Get actual weights being used (from ENV or defaults)
+        semantic_weight = (ENV["SWARM_MEMORY_SEMANTIC_WEIGHT"] || "0.5").to_f
+        keyword_weight = (ENV["SWARM_MEMORY_KEYWORD_WEIGHT"] || "0.5").to_f
+
+        SwarmSDK::LogStream.emit(
+          type: "semantic_memory_search",
+          agent: agent_name,
+          query: prompt,
+          query_word_count: word_count,
+          threshold: threshold,
+          threshold_type: word_count < word_cutoff ? "short_query" : "normal_query",
+          adaptive_cutoff: word_cutoff,
+          memories_found: memories.size,
+          total_memory_entries_searched: memory_entries.size,
+          search_mode: "hybrid",
+          weights: { semantic: semantic_weight, keyword: keyword_weight },
+          memories: memories.map do |memory|
+            {
+              path: memory[:path],
+              title: memory[:title],
+              type: memory.dig(:metadata, "type"),
+              hybrid_score: memory[:similarity].round(3),
+              semantic_score: memory[:semantic_score]&.round(3),
+              keyword_score: memory[:keyword_score]&.round(3),
+            }
+          end,
+          debug_top_results: debug_all_memories,
+        )
+      end
+
+      # Build system reminder for discovered skills
+      #
+      # @param skills [Array<Hash>] Skill search results
+      # @return [String] Formatted system reminder
+      def build_skill_discovery_reminder(skills)
+        reminder = "<system-reminder>\n"
+        reminder += "🎯 Found #{skills.size} skill(s) in memory that may be relevant:\n\n"
+
+        skills.each do |skill|
+          match_pct = (skill[:similarity] * 100).round
+          reminder += "**#{skill[:title]}** (#{match_pct}% match)\n"
+          reminder += "Path: `#{skill[:path]}`\n"
+          reminder += "To use: `LoadSkill(file_path: \"#{skill[:path]}\")`\n\n"
+        end
+
+        reminder += "**If a skill matches your task:** Load it to get step-by-step instructions and adapted tools.\n"
+        reminder += "**If none match (false positive):** Ignore and proceed normally.\n"
+        reminder += "</system-reminder>"
+
+        reminder
+      end
+
+      # Build system reminder for discovered memories
+      #
+      # @param memories [Array<Hash>] Memory search results (concepts/facts/experiences)
+      # @return [String] Formatted system reminder
+      def build_memory_discovery_reminder(memories)
+        reminder = "<system-reminder>\n"
+        reminder += "📚 Found #{memories.size} memory entr#{memories.size == 1 ? "y" : "ies"} that may provide context:\n\n"
+
+        memories.each do |memory|
+          match_pct = (memory[:similarity] * 100).round
+          type = memory.dig(:metadata, "type")
+          type_emoji = case type
+          when "concept" then "💡"
+          when "fact" then "📋"
+          when "experience" then "🔍"
+          else "📄"
+          end
+
+          reminder += "#{type_emoji} **#{memory[:title]}** (#{type}, #{match_pct}% match)\n"
+          reminder += "Path: `#{memory[:path]}`\n"
+          reminder += "Read with: `MemoryRead(file_path: \"#{memory[:path]}\")`\n\n"
+        end
+
+        reminder += "**These entries may contain relevant knowledge for your task.**\n"
+        reminder += "Read them to inform your approach, or ignore if not helpful.\n"
+        reminder += "</system-reminder>"
+
+        reminder
+      end
+    end
+  end
+end