language-operator 0.1.58 → 0.1.59

data/lib/language_operator/cli/commands/agent.rb

@@ -491,6 +491,188 @@ module LanguageOperator
           end
         end
 
+        desc 'optimize NAME', 'Optimize neural tasks to symbolic based on learned patterns'
+        long_desc <<-DESC
+          Analyze agent execution patterns and propose optimizations to convert
+          neural (LLM-based) tasks into symbolic (code-based) implementations.
+
+          This command queries OpenTelemetry traces to detect deterministic patterns
+          in task execution, then generates optimized symbolic code that runs faster
+          and costs less while maintaining the same behavior.
+
+          Requirements:
+            • OpenTelemetry backend configured (SigNoz, Jaeger, or Tempo)
+            • Neural task has executed at least 10 times
+            • Execution pattern consistency >= 85%
+
+          Examples:
+            aictl agent optimize my-agent                    # Analyze and propose optimizations
+            aictl agent optimize my-agent --dry-run          # Show what would be optimized
+            aictl agent optimize my-agent --status-only      # Show learning status only
+            aictl agent optimize my-agent --auto-accept      # Auto-accept high-confidence optimizations
+            aictl agent optimize my-agent --tasks task1,task2  # Optimize specific tasks only
+        DESC
+        option :cluster, type: :string, desc: 'Override current cluster context'
+        option :dry_run, type: :boolean, default: false, desc: 'Show what would be optimized without applying'
+        option :status_only, type: :boolean, default: false, desc: 'Show learning status without optimizing'
+        option :auto_accept, type: :boolean, default: false, desc: 'Auto-accept optimizations above min-confidence'
+        option :min_confidence, type: :numeric, default: 0.90, desc: 'Minimum consistency for auto-accept (0.0-1.0)'
+        option :tasks, type: :array, desc: 'Only optimize specific tasks'
+        option :since, type: :string, desc: 'Only analyze traces since (e.g., "2h", "1d", "7d")'
+        option :use_synthesis, type: :boolean, default: false, desc: 'Use LLM synthesis instead of pattern detection'
+        option :synthesis_model, type: :string, desc: 'Model to use for synthesis (default: cluster default)'
+        def optimize(name)
+          handle_command_error('optimize agent') do
+            require_relative '../../learning/trace_analyzer'
+            require_relative '../../learning/pattern_detector'
+            require_relative '../../learning/optimizer'
+            require_relative '../../learning/task_synthesizer'
+            require_relative '../../agent/safety/ast_validator'
+            require_relative '../formatters/optimization_formatter'
+
+            ctx = Helpers::ClusterContext.from_options(options)
+
+            # Get agent to verify it exists
+            get_resource_or_exit('LanguageAgent', name)
+
+            # Get agent code/definition
+            agent_definition = load_agent_definition(ctx, name)
+            unless agent_definition
+              Formatters::ProgressFormatter.error("Could not load agent definition for '#{name}'")
+              exit 1
+            end
+
+            # Check for OpenTelemetry configuration
+            unless ENV['OTEL_QUERY_ENDPOINT']
+              Formatters::ProgressFormatter.warn('OpenTelemetry endpoint not configured')
+              puts
+              puts 'Set OTEL_QUERY_ENDPOINT to enable learning:'
+              puts '  export OTEL_QUERY_ENDPOINT=https://your-signoz-instance.com'
+              puts '  export OTEL_QUERY_API_KEY=your-api-key          # For SigNoz authentication'
+              puts '  export OTEL_QUERY_BACKEND=signoz                # Optional: signoz, jaeger, or tempo'
+              puts
+              puts 'Auto-detection tries backends in order: SigNoz → Jaeger → Tempo'
+              puts 'Set OTEL_QUERY_BACKEND to skip auto-detection and use a specific backend.'
+              puts
+              exit 1
+            end
+
+            # Initialize learning components
+            trace_analyzer = Learning::TraceAnalyzer.new(
+              endpoint: ENV.fetch('OTEL_QUERY_ENDPOINT', nil),
+              api_key: ENV.fetch('OTEL_QUERY_API_KEY', nil),
+              backend: ENV.fetch('OTEL_QUERY_BACKEND', nil)
+            )
+
+            unless trace_analyzer.available?
+              Formatters::ProgressFormatter.error('OpenTelemetry backend not available')
+              puts
+              puts 'Check your OTEL_QUERY_ENDPOINT configuration and backend status.'
+              exit 1
+            end
+
+            validator = LanguageOperator::Agent::Safety::ASTValidator.new
+            pattern_detector = LanguageOperator::Learning::PatternDetector.new(
+              trace_analyzer: trace_analyzer,
+              validator: validator
+            )
+
+            # Create task synthesizer for fallback (or forced via --use-synthesis)
+            # Synthesis is used when pattern detection fails OR --use-synthesis is set
+            task_synthesizer = nil
+            llm_client = create_synthesis_llm_client(ctx, options[:synthesis_model])
+            if llm_client
+              task_synthesizer = LanguageOperator::Learning::TaskSynthesizer.new(
+                llm_client: llm_client,
+                validator: validator
+              )
+              Formatters::ProgressFormatter.info('LLM synthesis mode (forced)') if options[:use_synthesis]
+            elsif options[:use_synthesis]
+              Formatters::ProgressFormatter.warn('Could not create LLM client for synthesis')
+            end
+
+            optimizer = LanguageOperator::Learning::Optimizer.new(
+              agent_name: name,
+              agent_definition: agent_definition,
+              trace_analyzer: trace_analyzer,
+              pattern_detector: pattern_detector,
+              task_synthesizer: task_synthesizer
+            )
+
+            formatter = Formatters::OptimizationFormatter.new
+
+            # Parse --since option into time range
+            time_range = parse_since_option(options[:since])
+
+            # Analyze for opportunities
+            opportunities = optimizer.analyze(time_range: time_range)
+
+            # Display analysis only in status-only mode
+            if options[:status_only]
+              puts formatter.format_analysis(agent_name: name, opportunities: opportunities)
+              return
+            end
+
+            # Exit if no opportunities
+            return if opportunities.empty?
+
+            # Filter opportunities:
+            # - If synthesis available: try any task with enough executions
+            # - Otherwise: only tasks ready for pattern detection
+            candidates = if task_synthesizer
+                           # With synthesis, try any task that has min executions
+                           opportunities.select { |opp| opp[:execution_count] >= 10 }
+                         else
+                           opportunities.select { |opp| opp[:ready_for_learning] }
+                         end
+            return if candidates.empty?
+
+            # Process each opportunity
+            candidates.each do |opp|
+              task_name = opp[:task_name]
+
+              # Skip if not in requested tasks list
+              next if options[:tasks] && !options[:tasks].include?(task_name)
+
+              # Generate proposal
+              begin
+                proposal = optimizer.propose(task_name: task_name, use_synthesis: options[:use_synthesis])
+              rescue ArgumentError => e
+                Formatters::ProgressFormatter.warn("Cannot optimize '#{task_name}': #{e.message}")
+                next
+              end
+
+              # Display proposal
+              puts formatter.format_proposal(proposal: proposal)
+
+              # Get user confirmation or auto-accept
+              accepted = if options[:auto_accept] && proposal[:consistency_score] >= options[:min_confidence]
+                           consistency_pct = (proposal[:consistency_score] * 100).round(1)
+                           threshold_pct = (options[:min_confidence] * 100).round(1)
+                           puts pastel.green("✓ Auto-accepting (consistency: #{consistency_pct}% >= #{threshold_pct}%)")
+                           true
+                         elsif options[:dry_run]
+                           puts pastel.yellow('[DRY RUN] Would prompt for acceptance')
+                           false
+                         else
+                           prompt_for_optimization_acceptance(proposal)
+                         end
+
+              # Apply if accepted
+              if accepted && !options[:dry_run]
+                result = apply_optimization(ctx, name, proposal)
+                puts formatter.format_success(result: result)
+              elsif accepted
+                puts pastel.yellow('[DRY RUN] Would apply optimization')
+              else
+                puts pastel.yellow("Skipped optimization for '#{task_name}'")
+              end
+
+              puts
+            end
+          end
+        end
+
         desc 'workspace NAME', 'Browse agent workspace files'
         long_desc <<-DESC
           Browse and manage the workspace files for an agent.
@@ -538,6 +720,183 @@ module LanguageOperator
 
         private
 
+        # Parse --since option into seconds (time range)
+        #
+        # @param since [String, nil] Duration string (e.g., "2h", "1d", "7d")
+        # @return [Integer, nil] Seconds or nil if not specified
+        def parse_since_option(since)
+          return nil unless since
+
+          match = since.match(/^(\d+)([hHdDwW])$/)
+          unless match
+            Formatters::ProgressFormatter.warn("Invalid --since format '#{since}', using default (24h)")
+            Formatters::ProgressFormatter.info('Valid formats: 2h (hours), 1d (days), 1w (weeks)')
+            return nil
+          end
+
+          value = match[1].to_i
+          unit = match[2].downcase
+
+          case unit
+          when 'h' then value * 3600
+          when 'd' then value * 86_400
+          when 'w' then value * 604_800
+          end
+        end
+
+        # Create LLM client for task synthesis using cluster model
+        #
+        # @param ctx [ClusterContext] Cluster context
+        # @param model_name [String, nil] Specific model to use (defaults to first available)
+        # @return [Object, nil] LLM client or nil if unavailable
+        def create_synthesis_llm_client(ctx, model_name = nil)
+          # Get model from cluster
+          selected_model = model_name || select_synthesis_model(ctx)
+          return nil unless selected_model
+
+          # Get model resource to extract model ID
+          # Always use port-forwarding to deployment (LiteLLM proxy for cost controls)
+          begin
+            model = ctx.client.get_resource('LanguageModel', selected_model, ctx.namespace)
+            model_id = model.dig('spec', 'modelName')
+            return nil unless model_id
+
+            ClusterLLMClient.new(
+              ctx: ctx,
+              model_name: selected_model,
+              model_id: model_id,
+              agent_command: self
+            )
+          rescue StandardError => e
+            @logger&.warn("Failed to create cluster LLM client: #{e.message}")
+            nil
+          end
+        end
+
+        # Select model for synthesis (first available if not specified)
+        #
+        # @param ctx [ClusterContext] Cluster context
+        # @return [String, nil] Model name or nil
+        def select_synthesis_model(ctx)
+          models = ctx.client.list_resources('LanguageModel', namespace: ctx.namespace)
+          return nil if models.empty?
+
+          models.first.dig('metadata', 'name')
+        rescue StandardError
+          nil
+        end
+
+        # LLM client that uses port-forwarding to cluster model deployments (LiteLLM proxy)
+        class ClusterLLMClient
+          def initialize(ctx:, model_name:, model_id:, agent_command:)
+            @ctx = ctx
+            @model_name = model_name
+            @model_id = model_id
+            @agent_command = agent_command
+          end
+
+          def chat(prompt)
+            require 'faraday'
+            require 'json'
+
+            pod = get_model_pod
+            pod_name = pod.dig('metadata', 'name')
+
+            local_port = find_available_port
+            port_forward_pid = nil
+
+            begin
+              port_forward_pid = start_port_forward(pod_name, local_port, 4000)
+              wait_for_port(local_port)
+
+              conn = Faraday.new(url: "http://localhost:#{local_port}") do |f|
+                f.request :json
+                f.response :json
+                f.adapter Faraday.default_adapter
+                f.options.timeout = 120
+                f.options.open_timeout = 10
+              end
+
+              payload = {
+                model: @model_id,
+                messages: [{ role: 'user', content: prompt }],
+                max_tokens: 4000,
+                temperature: 0.3
+              }
+
+              response = conn.post('/v1/chat/completions', payload)
+              result = response.body
+
+              raise "LLM error: #{result['error']['message'] || result['error']}" if result['error']
+
+              result.dig('choices', 0, 'message', 'content')
+            ensure
+              cleanup_port_forward(port_forward_pid) if port_forward_pid
+            end
+          end
+
+          private
+
+          def get_model_pod
+            # Get the deployment for the model
+            deployment = @ctx.client.get_resource('Deployment', @model_name, @ctx.namespace)
+            raise "Deployment '#{@model_name}' not found in namespace '#{@ctx.namespace}'" if deployment.nil?
+
+            labels = deployment.dig('spec', 'selector', 'matchLabels')
+            raise "Deployment '#{@model_name}' has no selector labels" if labels.nil?
+
+            # Convert to hash if needed (K8s API may return K8s::Resource)
+            labels_hash = labels.respond_to?(:to_h) ? labels.to_h : labels
+            raise "Deployment '#{@model_name}' has empty selector labels" if labels_hash.empty?
+
+            label_selector = labels_hash.map { |k, v| "#{k}=#{v}" }.join(',')
+
+            # Find a running pod
+            pods = @ctx.client.list_resources('Pod', namespace: @ctx.namespace, label_selector: label_selector)
+            raise "No pods found for model '#{@model_name}'" if pods.empty?
+
+            running_pods = pods.select { |p| p.dig('status', 'phase') == 'Running' }
+            raise "No running pods found for model '#{@model_name}'" if running_pods.empty?
+
+            running_pods.first
+          end
+
+          def find_available_port
+            server = TCPServer.new('127.0.0.1', 0)
+            port = server.addr[1]
+            server.close
+            port
+          end
+
+          def start_port_forward(pod_name, local_port, remote_port)
+            pid = spawn(
+              'kubectl', 'port-forward',
+              '-n', @ctx.namespace,
+              "pod/#{pod_name}",
+              "#{local_port}:#{remote_port}",
+              %i[out err] => '/dev/null'
+            )
+            Process.detach(pid)
+            pid
+          end
+
+          def wait_for_port(port, max_attempts: 30)
+            max_attempts.times do
+              TCPSocket.new('127.0.0.1', port).close
+              return true
+            rescue Errno::ECONNREFUSED
+              sleep 0.1
+            end
+            raise "Port #{port} not available after #{max_attempts} attempts"
+          end
+
+          def cleanup_port_forward(pid)
+            Process.kill('TERM', pid)
+          rescue Errno::ESRCH
+            # Process already gone
+          end
+        end
+
         def handle_agent_not_found(name, ctx)
           # Get available agents for fuzzy matching
           agents = ctx.client.list_resources('LanguageAgent', namespace: ctx.namespace)
@@ -1131,6 +1490,222 @@ module LanguageOperator
         def format_timestamp(time)
           Formatters::ValueFormatter.timestamp(time)
         end
+
+        # Load agent definition from ConfigMap
+        def load_agent_definition(ctx, agent_name)
+          # Try to get the agent code ConfigMap
+          configmap_name = "#{agent_name}-code"
+          begin
+            configmap = ctx.client.get_resource('ConfigMap', configmap_name, ctx.namespace)
+            code_content = configmap.dig('data', 'agent.rb')
+
+            return nil unless code_content
+
+            # Parse the code to extract agent definition
+            # For now, we'll create a mock definition with the task structure
+            # In a full implementation, this would eval the code safely
+            parse_agent_code(code_content)
+          rescue K8s::Error::NotFound
+            nil
+          rescue StandardError => e
+            @logger&.error("Failed to load agent definition: #{e.message}")
+            nil
+          end
+        end
+
+        # Parse agent code to extract definition
+        def parse_agent_code(code)
+          require_relative '../../dsl/agent_definition'
+
+          # Create a minimal agent definition structure
+          agent_def = Struct.new(:tasks, :name, :mcp_servers) do
+            def initialize
+              super({}, 'agent', {})
+            end
+          end
+
+          agent = agent_def.new
+
+          # Parse tasks from code - extract full task definitions
+          code.scan(/task\s+:(\w+),?\s*(.*?)(?=\n\s*(?:task\s+:|main\s+do|end\s*$))/m) do |match|
+            task_name = match[0].to_sym
+            task_block = match[1]
+
+            # Check if neural (has instructions but no do block) or symbolic
+            is_neural = task_block.include?('instructions:') && !task_block.match?(/\bdo\s*\|/)
+
+            # Extract instructions
+            instructions = extract_string_value(task_block, 'instructions')
+
+            # Extract inputs hash
+            inputs = extract_hash_value(task_block, 'inputs')
+
+            # Extract outputs hash
+            outputs = extract_hash_value(task_block, 'outputs')
+
+            task = Struct.new(:name, :neural?, :instructions, :inputs, :outputs).new(
+              task_name, is_neural, instructions, inputs, outputs
+            )
+
+            agent.tasks[task_name] = task
+          end
+
+          agent
+        end
+
+        # Extract a string value from DSL code (e.g., instructions: "...")
+        def extract_string_value(code, key)
+          # Match both single and double quoted strings, including multi-line
+          match = code.match(/#{key}:\s*(['"])(.*?)\1/m) ||
+                  code.match(/#{key}:\s*(['"])(.+?)\1/m)
+          match ? match[2] : ''
+        end
+
+        # Extract a hash value from DSL code (e.g., inputs: { foo: 'bar' })
+        def extract_hash_value(code, key)
+          match = code.match(/#{key}:\s*\{([^}]*)\}/)
+          return {} unless match
+
+          hash_content = match[1].strip
+          return {} if hash_content.empty?
+
+          # Parse simple key: 'value' or key: "value" pairs
+          result = {}
+          hash_content.scan(/(\w+):\s*(['"])([^'"]*)\2/) do |k, _quote, v|
+            result[k.to_sym] = v
+          end
+          result
+        end
+
+        # Prompt user for optimization acceptance
+        def prompt_for_optimization_acceptance(proposal)
+          require 'tty-prompt'
+          prompt = TTY::Prompt.new
+
+          choices = [
+            { name: 'Yes - apply this optimization', value: :yes },
+            { name: 'No - skip this task', value: :no },
+            { name: 'View full code diff', value: :diff },
+            { name: 'Skip all remaining', value: :skip_all }
+          ]
+
+          loop do
+            choice = prompt.select(
+              "Accept optimization for '#{proposal[:task_name]}'?",
+              choices,
+              per_page: 10
+            )
+
+            case choice
+            when :yes
+              return true
+            when :no
+              return false
+            when :diff
+              show_code_diff(proposal)
+              # Loop to ask again
+            when :skip_all
+              throw :skip_all
+            end
+          end
+        end
+
+        # Show full code diff
+        def show_code_diff(proposal)
+          puts
+          puts pastel.bold('Full Generated Code:')
+          puts pastel.dim('=' * 70)
+          puts proposal[:full_generated_code]
+          puts pastel.dim('=' * 70)
+          puts
+        end
+
+        # Apply optimization by updating ConfigMap and restarting pod
+        def apply_optimization(ctx, agent_name, proposal)
+          configmap_name = "#{agent_name}-code"
+          task_name = proposal[:task_name]
+
+          # Get current ConfigMap
+          configmap = ctx.client.get_resource('ConfigMap', configmap_name, ctx.namespace)
+          current_code = configmap.dig('data', 'agent.rb')
+
+          raise "ConfigMap '#{configmap_name}' does not contain agent.rb" unless current_code
+
+          # Replace the neural task with the symbolic implementation
+          updated_code = replace_task_in_code(current_code, task_name, proposal[:proposed_code])
+
+          # Build updated ConfigMap resource
+          # Add annotation to prevent controller from overwriting optimized code
+          updated_configmap = {
+            'apiVersion' => 'v1',
+            'kind' => 'ConfigMap',
+            'metadata' => {
+              'name' => configmap_name,
+              'namespace' => ctx.namespace,
+              'resourceVersion' => configmap.metadata.resourceVersion,
+              'annotations' => {
+                'langop.io/optimized' => 'true',
+                'langop.io/optimized-at' => Time.now.iso8601,
+                'langop.io/optimized-task' => task_name
+              }
+            },
+            'data' => {
+              'agent.rb' => updated_code
+            }
+          }
+
+          # Update ConfigMap
+          ctx.client.update_resource('ConfigMap', configmap_name, ctx.namespace, updated_configmap, 'v1')
+
+          # Restart the agent pod to pick up changes
+          restart_agent_pod(ctx, agent_name)
+
+          {
+            success: true,
+            task_name: task_name,
+            updated_code: proposal[:proposed_code],
+            action: 'applied',
+            message: "Optimization for '#{task_name}' applied successfully"
+          }
+        rescue StandardError => e
+          {
+            success: false,
+            task_name: task_name,
+            error: e.message,
+            action: 'failed',
+            message: "Failed to apply optimization: #{e.message}"
+          }
+        end
+
+        # Replace a task definition in agent code
+        def replace_task_in_code(code, task_name, new_task_code)
+          # Match the task definition including any trailing do block
+          # Pattern matches: task :name, ... (neural) or task :name, ... do |inputs| ... end (symbolic)
+          task_pattern = /task\s+:#{Regexp.escape(task_name.to_s)},?\s*.*?(?=\n\s*(?:task\s+:|main\s+do|end\s*$))/m
+
+          raise "Could not find task ':#{task_name}' in agent code" unless code.match?(task_pattern)
+
+          # Ensure new_task_code has proper trailing newline
+          new_code = "#{new_task_code.strip}\n\n"
+
+          code.gsub(task_pattern, new_code.strip)
+        end
+
+        # Restart agent pod by deleting it (Deployment will recreate)
+        def restart_agent_pod(ctx, agent_name)
+          # Find pods for this agent
+          pods = ctx.client.list_resources('Pod', namespace: ctx.namespace, label_selector: "app=#{agent_name}")
+
+          pods.each do |pod|
+            pod_name = pod.dig('metadata', 'name')
+            begin
+              ctx.client.delete_resource('Pod', pod_name, ctx.namespace)
+              Formatters::ProgressFormatter.info("Restarting pod '#{pod_name}'")
+            rescue StandardError => e
+              Formatters::ProgressFormatter.warn("Could not delete pod '#{pod_name}': #{e.message}")
+            end
+          end
+        end
       end
     end
   end