ruby_llm-agents 3.13.0 → 3.14.0

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: 5fc64ee9b7db541e40c144d3b8007c0967f7ad63a7b5c404079d7c4c3398a646
-  data.tar.gz: 1c5b0d4c8c55390cfb5bee3ad394b47f4cae44dbd0452c5135c6ad7e7ac66a0b
+  metadata.gz: d974a549e2d99bbcd16c8345547d16c16aff7ae602da503974d9db1f32be8ce6
+  data.tar.gz: 9947da35f39521706e7c7bef349dcf9f6a63f5f732409ff1774ce0607bd6429e
 SHA512:
-  metadata.gz: 9c187e97882d3b4cf91dc98db12004d809db61a04bf95463aa32f1fa64b53a54d7f0800dfe105e0e6c5bf541b79bcdbe8f3d7537767b9c0d9814095f8fe6889d
-  data.tar.gz: 63b31bfcfcd27acaded706b78cfa96db5825e98ebd6f4a7fe4b2786ed0f15d048c706fdd157c4364618388816ab739a8c0dac34c773bcfaf0f67d9dc90397766
+  metadata.gz: 8aa785d541d2e00d237a71d5a652643e7e2da83750f8264d754c00fce6c2e33fc7d6d47caa75c1ac94659fd05703a9c488a229064cd1d01157b6845e4ef3970e
+  data.tar.gz: 89bf72286b0a6e70418268457beea87df58bb16514a170730518adeadae84eac6d6ff6553e9726e481b00c2358ad9d909c9f20170bcf41224eb7efdaf3d7f1ac

data/README.md

@@ -317,7 +317,7 @@ mount RubyLLM::Agents::Engine => "/agents"
 
 - **Ruby** >= 3.1.0
 - **Rails** >= 7.0
-- **RubyLLM** >= 1.12.0
+- **RubyLLM** >= 1.16.0
 
 ## Contributing
 

data/app/controllers/ruby_llm/agents/analytics_controller.rb

@@ -35,6 +35,7 @@ module RubyLLM
         @days = range_to_days(@selected_range)
         parse_custom_dates if @selected_range == "custom"
 
+        set_active_filters
         current_scope = apply_filters(time_scoped(tenant_scoped_executions))
         prior_scope = apply_filters(prior_period_scope(tenant_scoped_executions))
 
@@ -114,6 +115,13 @@ module RubyLLM
           []
         end
 
+        set_active_filters
+      end
+
+      # Resolves the active agent/model/tenant filters from request params.
+      # Shared by index (which also loads dropdown options) and chart_data
+      # (JSON only) so apply_filters behaves consistently on both endpoints.
+      def set_active_filters
         @filter_agent = params[:agent].presence
         @filter_model = params[:model].presence
         @filter_tenant = params[:filter_tenant].presence

data/app/controllers/ruby_llm/agents/executions_controller.rb

@@ -214,8 +214,9 @@ module RubyLLM
         # Only show root executions - children are nested under parents
         scope = scope.where(parent_execution_id: nil)
 
-        # Eager load children for grouping
-        scope.includes(:child_executions)
+        # Eager load children for grouping and detail for error_message, which
+        # the list renders per row (otherwise an N+1 on error rows).
+        scope.includes(:child_executions, :detail)
       end
 
       # Checks whether turbo-rails is available in the host application

data/app/controllers/ruby_llm/agents/tenants_controller.rb

@@ -19,7 +19,9 @@ module RubyLLM
       # @return [void]
       def index
         @sort_params = parse_tenant_sort_params
-        scope = TenantBudget.all
+        # Eager-load tenant_record so display_name's live name resolution does
+        # not issue a query per row.
+        scope = TenantBudget.all.includes(:tenant_record)
 
         if params[:q].present?
           @search_query = params[:q].to_s.strip
@@ -67,7 +69,11 @@ module RubyLLM
       # @return [void]
       def update
         @tenant = TenantBudget.find(params[:id])
-        if @tenant.update(tenant_params)
+        attrs = tenant_params
+        # Linked tenants derive their name live from the host record, so ignore
+        # any submitted name — it would be overwritten on the next record sync.
+        attrs = attrs.except(:name) if @tenant.linked?
+        if @tenant.update(attrs)
           redirect_to tenant_path(@tenant), notice: "Tenant updated successfully"
         else
           render :edit, status: :unprocessable_entity

data/app/models/ruby_llm/agents/execution.rb

@@ -107,7 +107,13 @@ module RubyLLM
       validates :finish_reason, inclusion: {in: FINISH_REASONS}, allow_nil: true
 
       before_save :calculate_total_tokens, if: -> { input_tokens_changed? || output_tokens_changed? }
-      before_save :calculate_total_cost, if: -> { input_cost_changed? || output_cost_changed? }
+      # Derive total_cost from its components only when the caller did not set an
+      # explicit total in the same save. The pipeline records a
+      # cache/reasoning-aware total alongside input_cost/output_cost, and that
+      # richer value must not be overwritten with the text-only input+output sum.
+      # (Deriving from metadata is unsafe — metadata merges user-supplied agent
+      # data and a colliding key would corrupt the total.)
+      before_save :calculate_total_cost, if: -> { (input_cost_changed? || output_cost_changed?) && !total_cost_changed? }
 
       # Aggregates costs from all attempts using each attempt's model pricing
       #
@@ -474,7 +480,11 @@ module RubyLLM
         self.total_tokens = (input_tokens || 0) + (output_tokens || 0)
       end
 
-      # Calculates and sets total_cost from input and output costs
+      # Calculates and sets total_cost from input and output costs.
+      #
+      # Only runs when the caller did not provide an explicit total_cost (see
+      # the before_save guard), so a cache/reasoning-aware total supplied by the
+      # pipeline is preserved rather than collapsed to the text-only sum.
       #
       # @return [BigDecimal] The calculated total
       def calculate_total_cost

data/app/models/ruby_llm/agents/tenant.rb

@@ -145,11 +145,17 @@ module RubyLLM
         alias_method :for_tenant!, :for!
       end
 
-      # Display name (name or tenant_id fallback)
+      # Display name.
+      #
+      # For tenants linked to a host model (Account, Organization, ...) the name
+      # is resolved live from that record, so a renamed record is reflected
+      # immediately instead of showing the snapshot taken when the tenant was
+      # created. Unlinked (string-id) tenants fall back to the stored name
+      # column, and tenant_id is the final fallback so this is never blank.
       #
       # @return [String]
       def display_name
-        name.presence || tenant_id
+        linked_record_name.presence || name.presence || tenant_id
       end
 
       # Check if tenant is linked to a user model
@@ -179,6 +185,28 @@ module RubyLLM
       def activate!
         update!(active: true)
       end
+
+      private
+
+      # Live display name from the linked host record (Account/Organization),
+      # or nil when this tenant is unlinked or the record is unavailable.
+      # Prefers the model's llm_tenant_name (which honours the configured name
+      # method), falling back to a plain #name. Never raises — name resolution
+      # must not break rendering.
+      #
+      # @return [String, nil]
+      def linked_record_name
+        record = tenant_record
+        return nil unless record
+
+        if record.respond_to?(:llm_tenant_name)
+          record.llm_tenant_name
+        elsif record.respond_to?(:name)
+          record.name
+        end
+      rescue
+        nil
+      end
     end
   end
 end

data/app/views/ruby_llm/agents/tenants/_form.html.erb

@@ -17,13 +17,22 @@
 
     <div class="grid grid-cols-1 md:grid-cols-2 gap-6">
       <div>
-        <%= f.label :name, "display name", class: "block text-xs font-mono text-gray-500 dark:text-gray-400 mb-1" %>
-        <%= f.text_field :name,
-            class: "w-full px-3 py-2 bg-transparent border border-gray-200 dark:border-gray-800 rounded text-sm font-mono text-gray-900 dark:text-gray-100 focus:ring-1 focus:ring-gray-400 dark:focus:ring-gray-600 focus:border-gray-400 dark:focus:border-gray-600 placeholder-gray-400 dark:placeholder-gray-600",
-            placeholder: "e.g., Acme Corporation" %>
-        <p class="mt-1 text-[10px] font-mono text-gray-400 dark:text-gray-600">
-          Falls back to tenant ID if not set.
-        </p>
+        <label class="block text-xs font-mono text-gray-500 dark:text-gray-400 mb-1">display name</label>
+        <% if tenant.linked? %>
+          <p class="w-full px-3 py-2 bg-gray-50 dark:bg-gray-900 border border-gray-200 dark:border-gray-800 rounded text-sm font-mono text-gray-500 dark:text-gray-400">
+            <%= tenant.display_name %>
+          </p>
+          <p class="mt-1 text-[10px] font-mono text-gray-400 dark:text-gray-600">
+            Managed by the linked <%= tenant.tenant_record_type %>. Renaming it updates this automatically.
+          </p>
+        <% else %>
+          <%= f.text_field :name,
+              class: "w-full px-3 py-2 bg-transparent border border-gray-200 dark:border-gray-800 rounded text-sm font-mono text-gray-900 dark:text-gray-100 focus:ring-1 focus:ring-gray-400 dark:focus:ring-gray-600 focus:border-gray-400 dark:focus:border-gray-600 placeholder-gray-400 dark:placeholder-gray-600",
+              placeholder: "e.g., Acme Corporation" %>
+          <p class="mt-1 text-[10px] font-mono text-gray-400 dark:text-gray-600">
+            Falls back to tenant ID if not set.
+          </p>
+        <% end %>
       </div>
 
       <div>

data/lib/ruby_llm/agents/base_agent.rb

@@ -262,6 +262,24 @@ module RubyLLM
           @tools || (superclass.respond_to?(:tools) ? superclass.tools : [])
         end
 
+        # Sets or returns how this agent runs multiple tool calls returned in
+        # a single LLM response.
+        #
+        # Mirrors RubyLLM's tool_concurrency: +false+ runs them sequentially,
+        # +true+ or +:threads+ runs them in Ruby threads, and +:fibers+ runs
+        # them in fibers (requires the async gem). When unset, the agent
+        # inherits its superclass value and ultimately the global
+        # RubyLLM tool_concurrency configuration.
+        #
+        # @param value [Boolean, Symbol] Concurrency mode (omit to read)
+        # @return [Boolean, Symbol, nil] Configured mode, or nil when unset
+        def tool_concurrency(*value)
+          @tool_concurrency = value.first unless value.empty?
+          return @tool_concurrency if instance_variable_defined?(:@tool_concurrency)
+
+          superclass.respond_to?(:tool_concurrency) ? superclass.tool_concurrency : nil
+        end
+
         # @!endgroup
 
         # @!group Temperature DSL
@@ -738,6 +756,7 @@ module RubyLLM
       def execute(context)
         @context = context
         client = build_client(context)
+        @client = client
 
         # Make context available to Tool instances during tool execution
         previous_context = Thread.current[:ruby_llm_agents_caller_context]
@@ -788,7 +807,16 @@ module RubyLLM
         end
 
         client = client.with_schema(schema) if schema
-        client = client.with_tools(*resolved_tools) if resolved_tools.any?
+        if resolved_tools.any?
+          # Only pass concurrency when the agent overrides it; otherwise let
+          # RubyLLM apply its globally configured tool_concurrency default.
+          concurrency = self.class.tool_concurrency
+          client = if concurrency.nil?
+            client.with_tools(*resolved_tools)
+          else
+            client.with_tools(*resolved_tools, concurrency: concurrency)
+          end
+        end
         apply_tool_prompt_caching(client) if use_prompt_caching && resolved_tools.any?
         client = setup_tool_tracking(client) if resolved_tools.any?
         client = apply_messages(client, resolved_messages) if resolved_messages.any?
@@ -891,35 +919,80 @@ module RubyLLM
 
       # Captures response metadata to the context
       #
-      # @param response [RubyLLM::Message] The response
+      # When a tool returns RubyLLM::Tool::Halt, the response is a Halt
+      # instance with no token metadata. In that case we pull metadata from
+      # the last assistant message in the client's history.
+      #
+      # @param response [RubyLLM::Message, RubyLLM::Tool::Halt] The response
       # @param context [Pipeline::Context] The context
       def capture_response(response, context)
-        context.input_tokens = response.input_tokens
-        context.output_tokens = response.output_tokens
-        context.model_used = response.model_id || model
-        # finish_reason may not be available on all RubyLLM::Message versions
-        context.finish_reason = response.respond_to?(:finish_reason) ? response.finish_reason : nil
+        is_halt = response.is_a?(RubyLLM::Tool::Halt)
+        metadata = is_halt ? last_assistant_message_from_client : response
 
-        # Store tracked tool calls in context for instrumentation
-        context[:tool_calls] = @tracked_tool_calls if @tracked_tool_calls.any?
+        if metadata
+          context.input_tokens = metadata.input_tokens if metadata.respond_to?(:input_tokens)
+          context.output_tokens = metadata.output_tokens if metadata.respond_to?(:output_tokens)
+          context.model_used = (metadata.respond_to?(:model_id) && metadata.model_id) || model
 
-        # Capture Anthropic prompt caching metrics
-        if response.respond_to?(:cached_tokens) && response.cached_tokens&.positive?
-          context[:cached_tokens] = response.cached_tokens
+          # Capture Anthropic prompt caching metrics
+          if metadata.respond_to?(:cached_tokens) && metadata.cached_tokens&.positive?
+            context[:cached_tokens] = metadata.cached_tokens
+          end
+          if metadata.respond_to?(:cache_creation_tokens) && metadata.cache_creation_tokens&.positive?
+            context[:cache_creation_tokens] = metadata.cache_creation_tokens
+          end
+        else
+          context.model_used = model
         end
-        if response.respond_to?(:cache_creation_tokens) && response.cache_creation_tokens&.positive?
-          context[:cache_creation_tokens] = response.cache_creation_tokens
+
+        context.finish_reason = if is_halt
+          "halt"
+        elsif response.respond_to?(:finish_reason)
+          response.finish_reason
         end
 
-        calculate_costs(response, context) if context.input_tokens
+        # Store tracked tool calls in context for instrumentation
+        context[:tool_calls] = @tracked_tool_calls if @tracked_tool_calls.any?
+
+        calculate_costs(metadata, context) if metadata && context.input_tokens
       end
 
-      # Calculates costs for the response
+      # Finds the most recent assistant message with usage metadata in
+      # the active client's history. Used to recover token/model metadata
+      # when the LLM call short-circuits via Tool::Halt.
+      #
+      # @return [RubyLLM::Message, nil]
+      def last_assistant_message_from_client
+        messages = @client&.messages
+        return nil unless messages
+
+        messages.reverse_each.find do |m|
+          m.respond_to?(:role) && m.role == :assistant &&
+            m.respond_to?(:input_tokens) && m.input_tokens
+        end
+      end
+
+      # Calculates costs for the response.
+      #
+      # Providers often return dated model variants (e.g.
+      # "anthropic/claude-4.6-sonnet-20260217") that aren't in the
+      # RubyLLM::Models registry, while the agent is configured with a
+      # stable alias (e.g. "anthropic/claude-sonnet-4.6") that is. When the
+      # response's model_id misses, fall back to the agent's configured
+      # model so cost calculation still finds pricing.
+      #
+      # Text input/output are priced from the context's token counts. These
+      # reflect the final attempt's usage (a retry/fallback overwrites them per
+      # attempt); failed attempts that erred at the provider are typically not
+      # billed, so the final attempt is the charged one. On top of the text
+      # cost, cache reads/writes and reasoning tokens — which exist on the
+      # response and are billed at their own rates — are priced via RubyLLM's
+      # first-class cost helper (RubyLLM::Cost) and added in.
       #
       # @param response [RubyLLM::Message] The response
       # @param context [Pipeline::Context] The context
       def calculate_costs(response, context)
-        model_info = find_model_info(response.model_id || model)
+        model_info = find_model_info(response.model_id) || find_model_info(model)
         return unless model_info
 
         input_tokens = context.input_tokens || 0
@@ -929,16 +1002,111 @@ module RubyLLM
         output_price = model_info.pricing&.text_tokens&.output || 0
 
         context.input_cost = (input_tokens / 1_000_000.0) * input_price
-        context.output_cost = (output_tokens / 1_000_000.0) * output_price
-        context.total_cost = (context.input_cost + context.output_cost).round(6)
+
+        # Price cache/reasoning extras first so we know whether reasoning was
+        # actually billed at the reasoning rate. Only then exclude those tokens
+        # from the output charge — never subtract tokens that weren't charged
+        # elsewhere, or a degraded cost helper would make reasoning vanish.
+        extra = extra_token_costs(response, model_info, context)
+        billable_output = output_tokens - reasoning_tokens_charged(response, context)
+        context.output_cost = ([billable_output, 0].max / 1_000_000.0) * output_price
+
+        context.total_cost = (context.input_cost + context.output_cost + extra).round(6)
+      end
+
+      # Number of reasoning (thinking) tokens that were actually charged at the
+      # reasoning rate, recorded in the cost breakdown by +extra_token_costs+.
+      #
+      # Reasoning providers fold reasoning tokens into the reported
+      # output_tokens, so when they are billed separately they must be removed
+      # from the output-rate charge to avoid double billing. Returns 0 when no
+      # reasoning was charged (non-reasoning model, or a degraded cost helper),
+      # so reasoning tokens are never silently dropped from the output charge.
+      #
+      # @param response [Object] The response (RubyLLM::Message in production)
+      # @param context [Pipeline::Context] The context
+      # @return [Integer] Reasoning tokens to exclude from the output charge
+      def reasoning_tokens_charged(response, context)
+        breakdown = context[:cost_breakdown]
+        return 0 unless breakdown.is_a?(Hash) && breakdown.key?(:thinking)
+        return 0 unless response.respond_to?(:reasoning_tokens)
+
+        response.reasoning_tokens.to_i
+      end
+
+      # Prices the non-text token components (cache reads/writes, reasoning)
+      # that RubyLLM::Cost exposes on a response, records them in metadata for
+      # visibility, and returns their sum to add on top of text input/output.
+      #
+      # Returns 0.0 for responses that don't expose cost (plain structs/mocks)
+      # or when the registry lacks the relevant prices, so cache/reasoning
+      # accuracy is additive and never regresses text pricing.
+      #
+      # @param response [Object] The response (RubyLLM::Message in production)
+      # @param model_info [RubyLLM::Model::Info] Resolved pricing source
+      # @param context [Pipeline::Context] The context
+      # @return [Float] Combined cache + reasoning cost, or 0.0
+      def extra_token_costs(response, model_info, context)
+        cost = response_cost(response, model_info)
+        return 0.0 unless cost
+
+        components = {
+          cache_read: cost.cache_read,
+          cache_write: cost.cache_write,
+          thinking: cost.thinking
+        }.compact.reject { |_, value| value.zero? }
+        return 0.0 if components.empty?
+
+        # Round per component and sum the rounded values so the stored
+        # breakdown reconciles exactly with the amount added to total_cost.
+        breakdown = components.transform_values { |value| value.round(6) }
+        context[:cost_breakdown] = breakdown
+        breakdown.values.sum
+      rescue => e
+        # Non-standard pricing shapes can't price these components; degrade to
+        # text-only rather than failing the cost calculation.
+        log_cost_warning("extra_token_costs", e)
+        0.0
+      end
+
+      # Returns a RubyLLM::Cost for the response, priced against the resolved
+      # model_info (which may differ from the response's own dated model
+      # variant). Returns nil for responses that don't expose cost — e.g.
+      # simple structs/mocks in tests — so callers skip the extra components.
+      #
+      # @param response [Object] The response (RubyLLM::Message in production)
+      # @param model_info [RubyLLM::Model::Info] Resolved pricing source
+      # @return [RubyLLM::Cost, nil]
+      def response_cost(response, model_info)
+        return nil unless response.respond_to?(:cost)
+
+        response.cost(model: model_info)
+      rescue => e
+        log_cost_warning("response_cost", e)
+        nil
+      end
+
+      # Leaves a debug breadcrumb for a swallowed cost-calculation error.
+      # Cost components are best-effort, so we degrade gracefully rather than
+      # raise, but record why instead of failing silently. Logging itself must
+      # never break cost handling.
+      #
+      # @param source [String] The method that degraded
+      # @param error [Exception] The swallowed error
+      def log_cost_warning(source, error)
+        return unless defined?(Rails) && Rails.respond_to?(:logger) && Rails.logger
+
+        Rails.logger.debug("[RubyLLM::Agents] #{source} skipped: #{error.class}: #{error.message}")
+      rescue
+        nil
       end
 
-      # Finds model pricing info
+      # Finds model pricing info.
       #
       # @param model_id [String] The model ID
       # @return [Hash, nil] Model info with pricing
       def find_model_info(model_id)
-        return nil unless defined?(RubyLLM::Models)
+        return nil unless defined?(RubyLLM::Models) && model_id
 
         RubyLLM::Models.find(model_id)
       rescue

data/lib/ruby_llm/agents/core/configuration.rb

@@ -373,10 +373,18 @@ module RubyLLM
         gemini_api_base
         gpustack_api_base
         ollama_api_base
+        bedrock_api_base
+        mistral_api_base
+        perplexity_api_base
+        vertexai_api_base
         vertexai_project_id
         vertexai_location
+        xai_api_base
         request_timeout
         max_retries
+        faraday_adapter
+        deprecation_behavior
+        tool_concurrency
       ].freeze
 
       FORWARDED_RUBY_LLM_ATTRIBUTES.each do |attr|

data/lib/ruby_llm/agents/core/llm_tenant.rb

@@ -113,6 +113,11 @@ module RubyLLM
 
           # Auto-create tenant record callback
           after_create :create_default_llm_tenant if llm_tenant_options[:budget]
+
+          # Keep the denormalized Tenant#name column fresh so the dashboard's
+          # SQL search/sort by name keeps working for linked tenants. Display
+          # already resolves the name live, so this only powers SQL.
+          after_update :sync_llm_tenant_name
         end
 
         private
@@ -143,6 +148,17 @@ module RubyLLM
         send(id_method).to_s
       end
 
+      # Returns this model's tenant display name, resolved live from the
+      # configured name method (`llm_tenant name: :company_name`). Resolving on
+      # read means the tenant always reflects the current value instead of the
+      # snapshot taken when its Tenant record was first created.
+      #
+      # @return [String] The current display name
+      def llm_tenant_name
+        name_method = self.class.llm_tenant_options[:name] || :to_s
+        send(name_method).to_s
+      end
+
       # Returns API keys resolved from the DSL configuration
       #
       # Maps provider names (e.g., :openai, :anthropic) to their resolved values
@@ -354,6 +370,30 @@ module RubyLLM
         tenant.tenant_record = self
         tenant.save!
       end
+
+      # Pushes the current name into the linked Tenant row when the source
+      # column changed, keeping the denormalized copy fresh for the dashboard's
+      # SQL search/sort. Display already resolves live, so this is best-effort
+      # and never raises. Only runs when the name is backed by a column we can
+      # detect a change on (method-based names are skipped — display stays
+      # correct via live resolution, only SQL search/sort may lag for those).
+      #
+      # @return [void]
+      def sync_llm_tenant_name
+        name_method = self.class.llm_tenant_options[:name]
+        return unless name_method
+
+        change_predicate = "saved_change_to_#{name_method}?"
+        return unless respond_to?(change_predicate) && public_send(change_predicate)
+
+        record = llm_tenant_record
+        return unless record&.persisted?
+        return if record.read_attribute(:name) == llm_tenant_name
+
+        record.update_column(:name, llm_tenant_name)
+      rescue
+        nil
+      end
     end
   end
 end

data/lib/ruby_llm/agents/core/version.rb

@@ -4,6 +4,6 @@ module RubyLLM
   module Agents
     # Current version of the RubyLLM::Agents gem
     # @return [String] Semantic version string
-    VERSION = "3.13.0"
+    VERSION = "3.14.0"
   end
 end

data/lib/ruby_llm/agents/infrastructure/execution_logger_job.rb

@@ -37,8 +37,10 @@ module RubyLLM
           execution.create_detail!(detail_data)
         end
 
-        # Calculate costs if token data is available
-        if execution.input_tokens && execution.output_tokens
+        # Calculate costs if token data is available. Skip when the pipeline
+        # already supplied an accurate total (RubyLLM::Cost, which prices cache
+        # and reasoning tokens) so we don't downgrade it to text-only pricing.
+        if execution.input_tokens && execution.output_tokens && !execution.total_cost&.positive?
           execution.calculate_costs!
           execution.save!
         end

data/lib/ruby_llm/agents/pipeline/middleware/instrumentation.rb

@@ -48,7 +48,7 @@ module RubyLLM
               raised_exception = nil
 
               begin
-                @app.call(context)
+                capture_llm_requests(context) { @app.call(context) }
                 context.completed_at = Time.current
 
                 begin
@@ -84,6 +84,55 @@ module RubyLLM
 
           private
 
+          # Fiber-local stack of in-flight request accumulators, innermost last.
+          REQUEST_CAPTURE_STACK = :ruby_llm_agents_request_capture
+
+          # Captures real HTTP-level provider latency for the LLM call(s) made
+          # while running the rest of the pipeline.
+          #
+          # ruby_llm 1.16 emits a "request.ruby_llm" event per HTTP request and
+          # its Railtie wires ActiveSupport::Notifications as the instrumenter
+          # in Rails, so we subscribe for the duration of the downstream call
+          # and accumulate provider time and request count (retries/fallbacks
+          # add up). This is distinct from the total pipeline duration, which
+          # also includes middleware and tool execution. The values are stored
+          # in context metadata and persisted with the execution.
+          #
+          # AS::Notifications subscriptions are process-global, so a naive
+          # subscriber would also see events from other executions running
+          # concurrently (other threads) or nested inside this one (agent-as-
+          # tool). To attribute each request to exactly one execution, we keep
+          # a fiber-local stack of accumulators and only credit the innermost
+          # one on the thread that actually emitted the event — the callback
+          # runs synchronously on the emitting thread, so its top-of-stack is
+          # the execution whose LLM call fired.
+          #
+          # @param context [Context] The execution context
+          # @return [Object] The downstream call's return value
+          def capture_llm_requests(context)
+            return yield unless defined?(ActiveSupport::Notifications)
+
+            accumulator = {ms: 0.0, count: 0}
+            stack = (Thread.current[REQUEST_CAPTURE_STACK] ||= [])
+            stack.push(accumulator)
+
+            callback = lambda do |_name, started, finished, _id, _payload|
+              top = Thread.current[REQUEST_CAPTURE_STACK]&.last
+              next unless top.equal?(accumulator)
+
+              accumulator[:ms] += (finished - started) * 1000.0
+              accumulator[:count] += 1
+            end
+
+            ActiveSupport::Notifications.subscribed(callback, "request.ruby_llm") { yield }
+          ensure
+            stack&.pop
+            if accumulator && accumulator[:count].positive?
+              context[:llm_request_ms] = accumulator[:ms].round
+              context[:llm_request_count] = accumulator[:count]
+            end
+          end
+
           # Creates initial execution record with 'running' status
           #
           # Creates the record synchronously so it appears on the dashboard immediately.
@@ -339,6 +388,8 @@ module RubyLLM
               cache_hit: context.cached?,
               input_tokens: context.input_tokens || 0,
               output_tokens: context.output_tokens || 0,
+              input_cost: context.input_cost,
+              output_cost: context.output_cost,
               total_cost: context.total_cost || 0,
               attempts_count: context.attempts_made,
               chosen_model_id: context.model_used,

metadata

@@ -1,7 +1,7 @@
 --- !ruby/object:Gem::Specification
 name: ruby_llm-agents
 version: !ruby/object:Gem::Version
-  version: 3.13.0
+  version: 3.14.0
 platform: ruby
 authors:
 - adham90
@@ -29,14 +29,14 @@ dependencies:
     requirements:
     - - ">="
       - !ruby/object:Gem::Version
-        version: 1.14.1
+        version: 1.16.0
   type: :runtime
   prerelease: false
   version_requirements: !ruby/object:Gem::Requirement
     requirements:
     - - ">="
       - !ruby/object:Gem::Version
-        version: 1.14.1
+        version: 1.16.0
 - !ruby/object:Gem::Dependency
   name: csv
   requirement: !ruby/object:Gem::Requirement