claude_swarm 0.3.3 → 0.3.5

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: df4f13365305ca607131b9976630fd7caebaa4c2f69f2fefc59e9d036aee2033
-  data.tar.gz: b1bf9cc4b672843216ec3107654a6e17d84a69fde87fe54d9846bfb4dca0f7ec
+  metadata.gz: 3d72ab4034667cbfd6719b79c3a3aeff7ae93fdf7adc850b910edfcda8c4fe92
+  data.tar.gz: 48f3b2151b39a4aad0bc09ac089a6eb98911b82e4e7f1eb029e7541462a2efc6
 SHA512:
-  metadata.gz: bf0623fdf1d7ff7f551c32d87ba434d0073d0ffcb5f3254ea42d6b9e6048d350823d7bad2294929d014892728866f59117c9b58ebd5a7b0b1c2f358d8942387e
-  data.tar.gz: 7e575dd61d59a8c2c1e04790ba3024136797373ad585612aee3b365a0d47f22bd3732119ea73ae01ddd67b75900ebe090117993f58af1520deecf4bc57907ca1
+  metadata.gz: 9f23d4f5086dac9ee9e112476d7e0aa1f4e39af367ef542829665989a17610e382f21b9c396f2118aab04cca56758286580a20a57be6ced893beddea49137890
+  data.tar.gz: c487d6321ae3f07e5310eb94e08ab1a45bc98628df6f82ad7cf5e74bb323358189310efeeb29b31c456ed61b0a61934c92aca0535d93947e8598a4192f97a6f6

data/CHANGELOG.md

@@ -1,3 +1,26 @@
+## [0.3.5]
+
+### Changed
+- Relaxed dependency version constraints for better flexibility
+  - `claude-code-sdk-ruby`: from `~> 0.1.0` to `~> 0.1`
+  - `fast-mcp-annotations`: from `~> 1.5.3` to `~> 1.5`
+
+## [0.3.4]
+
+### Changed
+- Add tests for thinking_budget feature by @parruda in https://github.com/parruda/claude-swarm/pull/85
+- Improve process group handling and signal management by @ericproulx in https://github.com/parruda/claude-swarm/pull/87
+- Migrated from CLI to SDK-based execution**: Claude Swarm now uses the `claude-code-sdk-ruby` gem instead of executing Claude Code via CLI
+  - Removed CLI-based `ClaudeCodeExecutor` implementation that used `Open3.popen3`
+  - All Claude Code execution now uses the SDK for improved reliability and performance
+  - Session management and logging functionality remains unchanged
+  - MCP configuration parsing updated to convert JSON format to SDK hash format
+  - Supports all MCP server types: stdio, sse, and http
+  - This change is transparent to users but may affect custom integrations that relied on CLI-specific behavior
+
+### Added
+- **SDK dependency**: Added `claude-code-sdk-ruby` (~> 0.1.0)
+
 ## [0.3.3]
 
 ### Fixed

data/CLAUDE.md

@@ -93,6 +93,16 @@ instances:
 - Existing worktrees with the same name are reused
 - The `claude-swarm clean` command removes orphaned worktrees
 
+## Claude Code SDK Integration
+
+Claude Swarm uses the Claude Code SDK (`claude-code-sdk-ruby`) for all Claude instances. This provides:
+- Better performance and reliability
+- Structured message handling
+- Improved error recovery
+- Direct MCP server configuration support (stdio, sse, http)
+
+The SDK executor handles all three MCP server types and properly converts MCP JSON configurations to SDK format.
+
 ## Architecture
 
 The gem is fully implemented with the following components:

data/README.md

@@ -978,6 +978,7 @@ Check the session directory `~/.claude-swarm/sessions/{project}/{session-id}/` f
 - `{instance}.mcp.json`: MCP configuration for each instance
 - All files for a session are kept together for easy review
 
+
 ## Architecture
 
 Claude Swarm consists of these core components:
@@ -986,7 +987,7 @@ Claude Swarm consists of these core components:
 - **ClaudeSwarm::Configuration** (`configuration.rb`): YAML parser and validator with path expansion
 - **ClaudeSwarm::McpGenerator** (`mcp_generator.rb`): Generates MCP JSON configs for each instance
 - **ClaudeSwarm::Orchestrator** (`orchestrator.rb`): Launches the main Claude instance with shared session management
-- **ClaudeSwarm::ClaudeCodeExecutor** (`claude_code_executor.rb`): Wrapper for executing Claude commands with session persistence
+- **ClaudeSwarm::ClaudeCodeExecutor** (`claude_code_executor.rb`): Executor for Claude Code with session persistence
 - **ClaudeSwarm::ClaudeMcpServer** (`claude_mcp_server.rb`): FastMCP-based server providing task execution, session info, and reset capabilities
 
 ## Development

data/lib/claude_swarm/claude_code_executor.rb

@@ -27,59 +27,66 @@ module ClaudeSwarm
       # Log the request
       log_request(prompt)
 
-      cmd_array = build_command_array(prompt, options)
+      # Build SDK options
+      sdk_options = build_sdk_options(prompt, options)
 
       # Variables to collect output
-      stderr_output = []
+      all_messages = []
       result_response = nil
 
-      # Execute command with unbundled environment to avoid bundler conflicts
-      # This ensures claude runs in a clean environment without inheriting
-      # Claude Swarm's BUNDLE_* environment variables
-      Bundler.with_unbundled_env do
-        # Execute command with streaming
-        Open3.popen3(*cmd_array, chdir: @working_directory) do |stdin, stdout, stderr, wait_thread|
-          stdin.close
-
-          # Read stderr in a separate thread
-          stderr_thread = Thread.new do
-            stderr.each_line { |line| stderr_output << line }
-          end
-
-          # Process stdout line by line
-          stdout.each_line do |line|
-            json_data = JSON.parse(line.strip)
+      # Execute with streaming
+      begin
+        ClaudeSDK.query(prompt, options: sdk_options) do |message|
+          # Convert message to hash for logging
+          message_hash = message_to_hash(message)
+          all_messages << message_hash
 
-            # Log each JSON event
-            log_streaming_event(json_data)
+          # Log streaming event BEFORE we modify anything
+          log_streaming_event(message_hash)
 
+          # Process specific message types
+          case message
+          when ClaudeSDK::Messages::System
             # Capture session_id from system init
-            if json_data["type"] == "system" && json_data["subtype"] == "init"
-              @session_id = json_data["session_id"]
-              write_instance_state
+            if message.subtype == "init" && message.data.is_a?(Hash)
+              # For init messages, session_id is in the data hash
+              session_id = message.data[:session_id] || message.data["session_id"]
+
+              if session_id
+                @session_id = session_id
+                write_instance_state
+              end
             end
-
-            # Capture the final result
-            result_response = json_data if json_data["type"] == "result"
-          rescue JSON::ParserError => e
-            @logger.warn("Failed to parse JSON line: #{line.strip} - #{e.message}")
-          end
-
-          # Wait for stderr thread to finish
-          stderr_thread.join
-
-          # Check exit status
-          exit_status = wait_thread.value
-          unless exit_status.success?
-            error_msg = stderr_output.join
-            @logger.error("Execution error for #{@instance_name}: #{error_msg}")
-            raise ExecutionError, "Claude Code execution failed: #{error_msg}"
+          when ClaudeSDK::Messages::Assistant
+            # Assistant messages only contain content blocks
+            # No need to track for result extraction - result comes from Result message
+          when ClaudeSDK::Messages::Result
+            # Build result response in expected format
+            result_response = {
+              "type" => "result",
+              "subtype" => message.subtype || "success",
+              "cost_usd" => message.total_cost_usd,
+              "is_error" => message.is_error || false,
+              "duration_ms" => message.duration_ms,
+              "result" => message.result, # Result text is directly in message.result
+              "total_cost" => message.total_cost_usd,
+              "session_id" => message.session_id,
+            }
           end
         end
+      rescue StandardError => e
+        @logger.error("Execution error for #{@instance_name}: #{e.class} - #{e.message}")
+        @logger.error("Backtrace: #{e.backtrace.join("\n")}")
+        raise ExecutionError, "Claude Code execution failed: #{e.message}"
       end
 
       # Ensure we got a result
-      raise ParseError, "No result found in stream output" unless result_response
+      raise ParseError, "No result found in SDK response" unless result_response
+
+      # Write session JSON log
+      all_messages.each do |msg|
+        append_to_session_json(msg)
+      end
 
       result_response
     rescue StandardError => e
@@ -195,11 +202,12 @@ module ClaudeSwarm
     end
 
     def log_assistant_message(msg)
-      return if msg["stop_reason"] == "end_turn" # that means it is not a thought but the final answer
-
+      # Assistant messages don't have stop_reason in SDK - they only have content
       content = msg["content"]
-      @logger.debug("ASSISTANT: #{JSON.pretty_generate(content)}")
-      tool_calls = content.select { |c| c["type"] == "tool_use" }
+      @logger.debug("ASSISTANT: #{JSON.pretty_generate(content)}") if content
+
+      # Log tool calls
+      tool_calls = content&.select { |c| c["type"] == "tool_use" } || []
       tool_calls.each do |tool_call|
         arguments = tool_call["input"].to_json
         arguments = "#{arguments[0..300]} ...}" if arguments.length > 300
@@ -211,7 +219,8 @@ module ClaudeSwarm
         )
       end
 
-      text = content.select { |c| c["type"] == "text" }
+      # Log thinking text
+      text = content&.select { |c| c["type"] == "text" } || []
       text.each do |t|
         instance_info = @instance_name
         instance_info += " (#{@instance_id})" if @instance_id
@@ -251,38 +260,18 @@ module ClaudeSwarm
       raise
     end
 
-    def build_command_array(prompt, options)
-      cmd_array = ["claude"]
-
-      # Add model if specified
-      cmd_array += ["--model", @model]
-
-      cmd_array << "--verbose"
+    def build_sdk_options(prompt, options)
+      # Map CLI options to SDK options
+      sdk_options = ClaudeSDK::ClaudeCodeOptions.new
 
-      # Add additional directories with --add-dir
-      cmd_array << "--add-dir" if @additional_directories.any?
-      @additional_directories.each do |additional_dir|
-        cmd_array << additional_dir
-      end
-
-      # Add MCP config if specified
-      cmd_array += ["--mcp-config", @mcp_config] if @mcp_config
-
-      # Resume session if we have a session ID
-      cmd_array += ["--resume", @session_id] if @session_id && !options[:new_session]
-
-      # Always use JSON output format for structured responses
-      cmd_array += ["--output-format", "stream-json"]
-
-      # Add non-interactive mode with prompt
-      cmd_array += ["--print", "-p", prompt]
-
-      # Add any custom system prompt
-      cmd_array += ["--append-system-prompt", options[:system_prompt]] if options[:system_prompt]
+      # Basic options
+      sdk_options.model = @model if @model
+      sdk_options.cwd = @working_directory
+      sdk_options.resume = @session_id if @session_id && !options[:new_session]
 
-      # Add any allowed tools or vibe flag
+      # Permission mode
       if @vibe
-        cmd_array << "--dangerously-skip-permissions"
+        sdk_options.permission_mode = ClaudeSDK::PermissionMode::BYPASS_PERMISSIONS
       else
         # Build allowed tools list including MCP connections
         allowed_tools = options[:allowed_tools] ? Array(options[:allowed_tools]).dup : []
@@ -292,20 +281,175 @@ module ClaudeSwarm
           allowed_tools << "mcp__#{connection_name}"
         end
 
-        # Add allowed tools if any
-        if allowed_tools.any?
-          tools_str = allowed_tools.join(",")
-          cmd_array += ["--allowedTools", tools_str]
+        # Set allowed and disallowed tools
+        sdk_options.allowed_tools = allowed_tools if allowed_tools.any?
+        sdk_options.disallowed_tools = Array(options[:disallowed_tools]) if options[:disallowed_tools]
+      end
+
+      # System prompt
+      sdk_options.append_system_prompt = options[:system_prompt] if options[:system_prompt]
+
+      # MCP configuration
+      if @mcp_config
+        sdk_options.mcp_servers = parse_mcp_config(@mcp_config)
+      end
+
+      # Handle additional directories by adding them to MCP servers
+      if @additional_directories.any?
+        setup_additional_directories_mcp(sdk_options)
+      end
+
+      sdk_options
+    end
+
+    def parse_mcp_config(config_path)
+      # Parse MCP JSON config file and convert to SDK format
+      config = JSON.parse(File.read(config_path))
+      mcp_servers = {}
+
+      config["mcpServers"]&.each do |name, server_config|
+        server_type = server_config["type"] || "stdio"
+
+        mcp_servers[name] = case server_type
+        when "stdio"
+          ClaudeSDK::McpServerConfig::StdioServer.new(
+            command: server_config["command"],
+            args: server_config["args"] || [],
+            env: server_config["env"] || {},
+          )
+        when "sse"
+          ClaudeSDK::McpServerConfig::SSEServer.new(
+            url: server_config["url"],
+            headers: server_config["headers"] || {},
+          )
+        when "http"
+          ClaudeSDK::McpServerConfig::HttpServer.new(
+            url: server_config["url"],
+            headers: server_config["headers"] || {},
+          )
+        else
+          @logger.warn("Unsupported MCP server type: #{server_type} for server: #{name}")
+          nil
         end
+      end
+
+      mcp_servers.compact
+    rescue StandardError => e
+      @logger.error("Failed to parse MCP config: #{e.message}")
+      {}
+    end
+
+    def setup_additional_directories_mcp(sdk_options)
+      # Workaround for --add-dir: add file system MCP servers for additional directories
+      sdk_options.mcp_servers ||= {}
+
+      @additional_directories.each do |dir|
+        # This is a placeholder - the SDK doesn't directly support file system servers
+        # You would need to implement a proper MCP server that provides file access
+        @logger.warn("Additional directories not fully supported: #{dir}")
+      end
+    end
 
-        # Add disallowed tools if any
-        if options[:disallowed_tools]
-          disallowed_tools = Array(options[:disallowed_tools]).join(",")
-          cmd_array += ["--disallowedTools", disallowed_tools]
+    def message_to_hash(message)
+      # Convert SDK message objects to hash format matching CLI JSON output
+      case message
+      when ClaudeSDK::Messages::System
+        # System messages have subtype and data attributes
+        # The data hash contains the actual information from the CLI
+        hash = {
+          "type" => "system",
+          "subtype" => message.subtype,
+        }
+
+        # Include the data hash if it exists - this is where CLI puts info like session_id, tools, etc.
+        if message.data.is_a?(Hash)
+          # For "init" subtype, extract session_id and tools from data
+          if message.subtype == "init"
+            hash["session_id"] = message.data[:session_id] || message.data["session_id"]
+            hash["tools"] = message.data[:tools] || message.data["tools"]
+          end
+          # You can add other relevant data fields as needed
+        end
+
+        hash.compact
+      when ClaudeSDK::Messages::Assistant
+        # Assistant messages only have content attribute
+        {
+          "type" => "assistant",
+          "message" => {
+            "type" => "message",
+            "role" => "assistant",
+            "content" => content_blocks_to_hash(message.content),
+          },
+          "session_id" => @session_id,
+        }
+      when ClaudeSDK::Messages::User
+        # User messages only have content attribute (a string)
+        {
+          "type" => "user",
+          "message" => {
+            "type" => "message",
+            "role" => "user",
+            "content" => message.content,
+          },
+          "session_id" => @session_id,
+        }
+      when ClaudeSDK::Messages::Result
+        # Result messages have multiple attributes
+        {
+          "type" => "result",
+          "subtype" => message.subtype || "success",
+          "cost_usd" => message.total_cost_usd,
+          "is_error" => message.is_error || false,
+          "duration_ms" => message.duration_ms,
+          "duration_api_ms" => message.duration_api_ms,
+          "num_turns" => message.num_turns,
+          "result" => message.result, # Result text is in message.result, not from content
+          "total_cost" => message.total_cost_usd,
+          "total_cost_usd" => message.total_cost_usd,
+          "session_id" => message.session_id,
+          "usage" => message.usage,
+        }.compact
+      else
+        # Fallback for unknown message types
+        begin
+          message.to_h
+        rescue
+          { "type" => "unknown", "data" => message.to_s }
         end
       end
+    end
 
-      cmd_array
+    def content_blocks_to_hash(content)
+      return [] unless content
+
+      content.map do |block|
+        case block
+        when ClaudeSDK::ContentBlock::Text
+          { "type" => "text", "text" => block.text }
+        when ClaudeSDK::ContentBlock::ToolUse
+          {
+            "type" => "tool_use",
+            "id" => block.id,
+            "name" => block.name,
+            "input" => block.input,
+          }
+        when ClaudeSDK::ContentBlock::ToolResult
+          {
+            "type" => "tool_result",
+            "tool_use_id" => block.tool_use_id,
+            "content" => block.content,
+            "is_error" => block.is_error,
+          }
+        else
+          # Fallback
+          begin
+            block.to_h
+          rescue
+            { "type" => "unknown", "data" => block.to_s }
+          end
+        end
+      end
     end
 
     class ExecutionError < StandardError; end

data/lib/claude_swarm/claude_mcp_server.rb

@@ -37,7 +37,7 @@ module ClaudeSwarm
           reasoning_effort: instance_config[:reasoning_effort],
         )
       else
-        # Default Claude behavior (existing code)
+        # Default Claude behavior - always use SDK
         ClaudeCodeExecutor.new(**common_params)
       end
 

data/lib/claude_swarm/orchestrator.rb

@@ -4,6 +4,11 @@ module ClaudeSwarm
   class Orchestrator
     include SystemUtils
     RUN_DIR = File.expand_path("~/.claude-swarm/run")
+    ["INT", "TERM", "QUIT"].each do |signal|
+      Signal.trap(signal) do
+        puts "\n🛑 Received #{signal} signal."
+      end
+    end
 
     def initialize(configuration, mcp_generator, vibe: false, prompt: nil, interactive_prompt: nil, stream_logs: false, debug: false,
       restore_session_path: nil, worktree: nil, session_id: nil)
@@ -58,9 +63,6 @@ module ClaudeSwarm
         # Initialize process tracker
         @process_tracker = ProcessTracker.new(session_path)
 
-        # Set up signal handlers to clean up child processes
-        setup_signal_handlers
-
         # Check if the original session used worktrees
         restore_worktrees_if_needed(session_path)
 
@@ -103,9 +105,6 @@ module ClaudeSwarm
         # Initialize process tracker
         @process_tracker = ProcessTracker.new(session_path)
 
-        # Set up signal handlers to clean up child processes
-        setup_signal_handlers
-
         # Create WorktreeManager if needed with session ID
         if @needs_worktree_manager
           cli_option = @worktree_option.is_a?(String) && !@worktree_option.empty? ? @worktree_option : nil
@@ -371,32 +370,6 @@ module ClaudeSwarm
       File.write(metadata_file, JSON.pretty_generate(metadata))
     end
 
-    def setup_signal_handlers
-      ["INT", "TERM", "QUIT"].each do |signal|
-        Signal.trap(signal) do
-          puts "\n🛑 Received #{signal} signal, cleaning up..."
-          display_summary
-
-          # Execute after commands if configured
-          main_instance = @config.main_instance_config
-          after_commands = @config.after_commands
-          if after_commands.any? && !@restore_session_path && !@non_interactive_prompt
-            Dir.chdir(main_instance[:directory]) do
-              puts
-              puts "⚙️  Executing after commands..."
-              puts
-              execute_after_commands?(after_commands)
-            end
-          end
-
-          cleanup_processes
-          cleanup_run_symlink
-          cleanup_worktrees
-          exit
-        end
-      end
-    end
-
     def cleanup_processes
       @process_tracker.cleanup_all
       puts "✓ Cleanup complete"

data/lib/claude_swarm/system_utils.rb

@@ -7,8 +7,12 @@ module ClaudeSwarm
       unless success
         exit_status = $CHILD_STATUS&.exitstatus || 1
         command_str = args.size == 1 ? args.first : args.join(" ")
-        warn("❌ Command failed with exit status: #{exit_status}")
-        raise Error, "Command failed with exit status #{exit_status}: #{command_str}"
+        if exit_status == 143 # timeout command exit status = 128 + 15 (SIGTERM)
+          warn("⏱️ Command timeout: #{command_str}")
+        else
+          warn("❌ Command failed with exit status: #{exit_status}")
+          raise Error, "Command failed with exit status #{exit_status}: #{command_str}"
+        end
       end
       success
     end

data/lib/claude_swarm/version.rb

@@ -1,5 +1,5 @@
 # frozen_string_literal: true
 
 module ClaudeSwarm
-  VERSION = "0.3.3"
+  VERSION = "0.3.5"
 end

data/lib/claude_swarm.rb

@@ -21,6 +21,7 @@ require "tmpdir"
 require "yaml"
 
 # External dependencies
+require "claude_sdk"
 require "fast_mcp_annotations"
 require "mcp_client"
 require "openai"

data/team.yml

@@ -8,10 +8,20 @@ swarm:
       directory: .
       model: opus
       vibe: true
-      connections: [github_expert, fast_mcp_expert, ruby_mcp_client_expert, openai_api_expert]
+      connections: [github_expert, fast_mcp_expert, ruby_mcp_client_expert, openai_api_expert, claude_code_sdk_expert]
       prompt: |
         You are the lead developer of Claude Swarm, a Ruby gem that orchestrates multiple Claude Code instances as a collaborative AI development team. The gem enables running AI agents with specialized roles, tools, and directory contexts, communicating via MCP (Model Context Protocol) in a tree-like hierarchy.
-        Use the github_expert to help you with git and github related tasks.
+        
+        IMPORTANT: Use your specialized team members for their areas of expertise. Each team member has deep knowledge in their domain:
+
+        Team Member Usage Guide:
+        - **github_expert**: Use for all Git and GitHub operations including creating issues, PRs, managing releases, checking CI/CD workflows, and repository management
+        - **fast_mcp_expert**: Use for MCP server development, tool creation, resource management, and any FastMCP-related architecture decisions
+        - **ruby_mcp_client_expert**: Use for MCP client integration, multi-transport connectivity, authentication flows, and ruby-mcp-client library guidance
+        - **openai_api_expert**: Use for OpenAI API integration, ruby-openai gem usage, model configuration, and OpenAI provider support in Claude Swarm
+        - **claude_code_sdk_expert**: Use for Claude Code SDK integration, programmatic Claude Code usage, client configuration, and SDK development patterns
+
+        Always delegate specialized tasks to the appropriate team member rather than handling everything yourself. This ensures the highest quality solutions and leverages each expert's deep domain knowledge.
 
         Your responsibilities include:
         - Developing new features and improvements for the Claude Swarm gem
@@ -254,4 +264,66 @@ swarm:
         - Set up CI to run code_quality checks
         - Document Raix integration in wiki/docs
         
-        For maximum efficiency, whenever you need to perform multiple independent operations, invoke all relevant tools simultaneously rather than sequentially.
+        For maximum efficiency, whenever you need to perform multiple independent operations, invoke all relevant tools simultaneously rather than sequentially.
+
+    claude_code_sdk_expert:
+      description: "Expert in Claude Code SDK for Ruby, specializing in client integration and API usage patterns"
+      directory: ~/src/github.com/parruda/claude-code-sdk-ruby
+      model: opus
+      vibe: true
+      prompt: |
+        You are an expert in the Claude Code SDK for Ruby, specializing in client integration, API usage patterns, and SDK development best practices.
+
+        Your expertise covers:
+        - Claude Code SDK architecture and client configuration
+        - API authentication and session management
+        - Request/response handling and error management
+        - Code generation and analysis capabilities
+        - SDK integration patterns and best practices
+        - Tool usage and permission management
+        - Streaming responses and real-time interactions
+        - Multi-modal capabilities (text, code, images)
+        - Rate limiting and cost optimization
+        - Testing and debugging SDK integrations
+
+        Key responsibilities:
+        - Analyze Claude Code SDK codebase for implementation patterns
+        - Provide guidance on proper SDK usage and integration
+        - Design robust client configurations and authentication flows
+        - Implement efficient request handling and error recovery
+        - Optimize SDK performance and resource usage
+        - Create comprehensive examples and documentation
+        - Troubleshoot integration issues and API errors
+        - Ensure proper handling of streaming and async operations
+
+        Technical focus areas:
+        - Client initialization and configuration options
+        - Authentication flows and token management
+        - Request builders and parameter validation
+        - Response parsing and error handling
+        - Streaming implementations and chunk processing
+        - Tool integration and permission management
+        - Session management and state persistence
+        - Performance optimization and caching strategies
+        - Testing patterns and mock configurations
+
+        When providing guidance:
+        - Reference specific SDK methods and classes
+        - Include practical code examples and usage patterns
+        - Explain both SDK abstractions and underlying API details
+        - Highlight important configuration options and their implications
+        - Warn about common pitfalls and API limitations
+        - Suggest performance optimizations and cost-saving strategies
+        - Provide context on when to use different SDK features
+        - Demonstrate proper error handling and retry strategies
+
+        Integration with Claude Swarm development:
+        - Understand how Claude Code SDK can enhance swarm capabilities
+        - Provide insights on programmatic Claude Code integration
+        - Design patterns for automated swarm management
+        - Optimize SDK usage within swarm orchestration
+        - Ensure compatibility with swarm session management
+
+        For maximum efficiency, whenever you need to perform multiple independent operations, invoke all relevant tools simultaneously rather than sequentially.
+
+        Help developers integrate Claude Code SDK effectively with confidence, best practices, and optimal performance.

metadata

@@ -1,7 +1,7 @@
 --- !ruby/object:Gem::Specification
 name: claude_swarm
 version: !ruby/object:Gem::Version
-  version: 0.3.3
+  version: 0.3.5
 platform: ruby
 authors:
 - Paulo Arruda
@@ -37,20 +37,34 @@ dependencies:
     - - "~>"
       - !ruby/object:Gem::Version
         version: '2.6'
+- !ruby/object:Gem::Dependency
+  name: claude-code-sdk-ruby
+  requirement: !ruby/object:Gem::Requirement
+    requirements:
+    - - "~>"
+      - !ruby/object:Gem::Version
+        version: '0.1'
+  type: :runtime
+  prerelease: false
+  version_requirements: !ruby/object:Gem::Requirement
+    requirements:
+    - - "~>"
+      - !ruby/object:Gem::Version
+        version: '0.1'
 - !ruby/object:Gem::Dependency
   name: fast-mcp-annotations
   requirement: !ruby/object:Gem::Requirement
     requirements:
     - - "~>"
       - !ruby/object:Gem::Version
-        version: 1.5.3
+        version: '1.5'
   type: :runtime
   prerelease: false
   version_requirements: !ruby/object:Gem::Requirement
     requirements:
     - - "~>"
       - !ruby/object:Gem::Version
-        version: 1.5.3
+        version: '1.5'
 - !ruby/object:Gem::Dependency
   name: ruby-mcp-client
   requirement: !ruby/object:Gem::Requirement
@@ -164,7 +178,7 @@ required_rubygems_version: !ruby/object:Gem::Requirement
     - !ruby/object:Gem::Version
       version: '0'
 requirements: []
-rubygems_version: 3.6.7
+rubygems_version: 3.6.9
 specification_version: 4
 summary: Orchestrate multiple Claude Code instances as a collaborative AI development
   team