llm_cost_tracker 0.14.1 → 0.14.2

data/lib/llm_cost_tracker/charges/line_item.rb

@@ -58,7 +58,11 @@ module LlmCostTracker
       def self.from_token_usage(token_usage)
         return [] unless token_usage
 
-        token_usage.priced_quantities.filter_map do |key, quantity|
+        from_quantities(token_usage.priced_quantities)
+      end
+
+      def self.from_quantities(quantities)
+        quantities.filter_map do |key, quantity|
           next unless quantity.positive?
 
           dimension = Usage::Catalog.fetch(key)

data/lib/llm_cost_tracker/configuration/budgets.rb

@@ -35,7 +35,7 @@ module LlmCostTracker
       end
 
       def finalize!
-        @per_tag = deep_freeze(@per_tag || {})
+        @per_tag = deep_freeze(@per_tag)
       end
 
       private

data/lib/llm_cost_tracker/configuration/pricing.rb

@@ -29,7 +29,7 @@ module LlmCostTracker
       end
 
       def finalize!
-        @overrides = deep_freeze(@overrides || {})
+        @overrides = deep_freeze(@overrides)
       end
     end
   end

data/lib/llm_cost_tracker/configuration.rb

@@ -18,8 +18,6 @@ module LlmCostTracker
       budgets: Budgets, capture: Capture, ingestion: Ingestion, pricing: Pricing, tags: Tags
     }.freeze
 
-    SCALAR_ATTRIBUTES = %i[enabled].freeze
-
     DEPRECATED_OPTIONS = {
       monthly_budget: { to: %i[budgets monthly] },
       daily_budget: { to: %i[budgets daily] },
@@ -46,7 +44,7 @@ module LlmCostTracker
       log_level: { to: nil, note: "LlmCostTracker logs through Rails.logger, which owns the level" }
     }.freeze
 
-    attr_reader(*SCALAR_ATTRIBUTES, *SECTIONS.keys, :instrumented_integrations)
+    attr_reader(:enabled, *SECTIONS.keys, :instrumented_integrations)
 
     def initialize
       SECTIONS.each { |name, klass| instance_variable_set(:"@#{name}", klass.new(self)) }
@@ -67,11 +65,9 @@ module LlmCostTracker
       @instrumented_integrations.include?(name)
     end
 
-    SCALAR_ATTRIBUTES.each do |name|
-      define_method(:"#{name}=") do |value|
-        ensure_mutable!
-        instance_variable_set(:"@#{name}", value)
-      end
+    def enabled=(value)
+      ensure_mutable!
+      @enabled = value
     end
 
     DEPRECATED_OPTIONS.each do |old_name, spec|
@@ -101,7 +97,7 @@ module LlmCostTracker
 
     def finalize!
       SECTIONS.each_key { |name| public_send(name).finalize! }
-      @instrumented_integrations = deep_freeze(@instrumented_integrations || Set.new)
+      @instrumented_integrations = deep_freeze(@instrumented_integrations)
       @finalized = true
     end
 

data/lib/llm_cost_tracker/doctor/price_check.rb

@@ -20,8 +20,8 @@ module LlmCostTracker
         updated_at = metadata["updated_at"] || metadata[:updated_at]
         return configured_check(:warn, path, count, "metadata.updated_at missing; #{REFRESH_COMMAND}") unless updated_at
 
-        age_days = (Date.today - Date.iso8601(updated_at.to_s)).to_i
-        if age_days > STALE_AFTER_DAYS
+        file_date = Date.iso8601(updated_at.to_s)
+        if (Date.today - file_date).to_i > STALE_AFTER_DAYS
           return configured_check(
             :warn,
             path,
@@ -29,6 +29,15 @@ module LlmCostTracker
             "updated_at=#{updated_at} is older than #{STALE_AFTER_DAYS} days; #{REFRESH_COMMAND}"
           )
         end
+        bundled_at = LlmCostTracker::Pricing::Registry.metadata["updated_at"]
+        if bundled_at && file_date < Date.iso8601(bundled_at.to_s)
+          return configured_check(
+            :warn,
+            path,
+            count,
+            "updated_at=#{updated_at} is older than the bundled prices (#{bundled_at}); #{REFRESH_COMMAND}"
+          )
+        end
 
         configured_check(:ok, path, count, "updated_at=#{updated_at}")
       rescue Date::Error
@@ -49,7 +58,8 @@ module LlmCostTracker
         Check.new(
           :warn,
           "prices",
-          "using bundled prices updated_at=#{updated_at}; commit a prices_file for production releases"
+          "using bundled prices updated_at=#{updated_at}; " \
+          "commit a prices_file (config.pricing.file) for production releases"
         )
       end
 

data/lib/llm_cost_tracker/doctor/schema_check.rb

@@ -22,8 +22,7 @@ module LlmCostTracker
         Check.new(
           :error,
           @name,
-          "current schema required; #{errors.join('; ')}; " \
-          "run bin/rails generate llm_cost_tracker:install && bin/rails db:migrate"
+          "current schema required; #{errors.join('; ')}; see docs/upgrading.md"
         )
       end
     end

data/lib/llm_cost_tracker/doctor.rb

@@ -167,7 +167,7 @@ module LlmCostTracker
       return live_rollups_check unless LlmCostTracker.configuration.budgets.totals_source == :cache
 
       errors = LlmCostTracker::Ledger::Schema::CallRollups.current_schema_errors
-      return Check.new(:ok, "call rollups", "llm_cost_tracker_call_rollups exists") if errors.empty?
+      return rollups_drift_check if errors.empty?
 
       Check.new(
         :error,
@@ -176,13 +176,30 @@ module LlmCostTracker
       )
     end
 
+    def rollups_drift_check
+      month_start = LlmCostTracker::Ledger::Period.range_start(:month, Time.now)
+      today = LlmCostTracker::Ledger::Period.range_start(:day, Time.now)
+      rolled = LlmCostTracker::CallRollup.where(period: "day", period_start: month_start.to_date...today.to_date)
+                                         .sum(:total_cost)
+      recorded = LlmCostTracker::Call.where(tracked_at: month_start...today).sum(:total_cost)
+      return Check.new(:ok, "call rollups", "llm_cost_tracker_call_rollups exists") if rolled == recorded
+
+      Check.new(
+        :warn,
+        "call rollups",
+        "this month's day rollups before today total #{rolled.to_d.to_s('F')} but the calls total " \
+        "#{recorded.to_d.to_s('F')}, and monthly budgets read the rollups; " \
+        "run bin/rails llm_cost_tracker:rebuild_rollups"
+      )
+    end
+
     def live_rollups_check
       if Probe.table_exists?("llm_cost_tracker_call_rollups")
         Check.new(
           :warn,
           "call rollups",
           "budgets.totals_source=:ledger but llm_cost_tracker_call_rollups exists. " \
-          "Set config.budgets.totals_source = :cache to keep budget reads on the rollups fast path or drop the table."
+          "Set config.budgets.totals_source = :cache to use it or drop the table."
         )
       else
         Check.new(

data/lib/llm_cost_tracker/engine.rb

@@ -3,7 +3,6 @@
 require "rails"
 require_relative "../llm_cost_tracker"
 require_relative "assets"
-require "rack/files"
 
 module LlmCostTracker
   class Engine < ::Rails::Engine

data/lib/llm_cost_tracker/event.rb

@@ -1,5 +1,7 @@
 # frozen_string_literal: true
 
+require "active_support/core_ext/digest/uuid"
+
 require_relative "pricing/mode"
 
 module LlmCostTracker
@@ -53,6 +55,12 @@ module LlmCostTracker
       Pricing::Mode.tokenize(pricing_mode.to_s).include?("batch")
     end
 
+    def keyed_by_response_id
+      return self unless provider_response_id
+
+      with(event_id: Digest::UUID.uuid_v5(Digest::UUID::OID_NAMESPACE, "#{provider}/#{provider_response_id}"))
+    end
+
     def self.resolve_line_items(service_items)
       Array(service_items).map do |item|
         item.is_a?(Charges::LineItem) ? item : Charges::LineItem.build(item)

data/lib/llm_cost_tracker/generators/llm_cost_tracker/async_ingestion_generator.rb

@@ -32,12 +32,6 @@ module LlmCostTracker
           inline. The doctor check warns about unused async ingestion tables.
         MSG
       end
-
-      private
-
-      def migration_version
-        "[#{ActiveRecord::VERSION::MAJOR}.#{ActiveRecord::VERSION::MINOR}]"
-      end
     end
   end
 end

data/lib/llm_cost_tracker/generators/llm_cost_tracker/call_rollups_generator.rb

@@ -10,7 +10,7 @@ module LlmCostTracker
 
       source_root File.expand_path("templates", __dir__)
 
-      desc "Creates the optional llm_cost_tracker_call_rollups table for fast budget reads. " \
+      desc "Creates the optional llm_cost_tracker_call_rollups table of running budget totals. " \
            "Required when config.budgets.totals_source = :cache."
 
       def create_migration_file
@@ -28,16 +28,13 @@ module LlmCostTracker
               config.budgets.totals_source = :cache
             end
 
-          Without it Tracker keeps reading budget totals as live SUM aggregates over
+          Once that is deployed, run bin/rails llm_cost_tracker:rebuild_rollups. Until then
+          calls recorded before the switch are missing from monthly budget totals.
+
+          Without the setting Tracker keeps reading budget totals as live SUM aggregates over
           llm_cost_tracker_calls. The doctor check warns about an unused rollups table.
         MSG
       end
-
-      private
-
-      def migration_version
-        "[#{ActiveRecord::VERSION::MAJOR}.#{ActiveRecord::VERSION::MINOR}]"
-      end
     end
   end
 end

data/lib/llm_cost_tracker/generators/llm_cost_tracker/install_generator.rb

@@ -53,12 +53,6 @@ module LlmCostTracker
           unauthenticated would expose spend, tags, and provider IDs to anyone.
         MSG
       end
-
-      private
-
-      def migration_version
-        "[#{ActiveRecord::VERSION::MAJOR}.#{ActiveRecord::VERSION::MINOR}]"
-      end
     end
   end
 end

data/lib/llm_cost_tracker/generators/llm_cost_tracker/templates/create_llm_cost_tracker_async_ingestion.rb.erb

@@ -1,4 +1,4 @@
-class CreateLlmCostTrackerAsyncIngestion < ActiveRecord::Migration<%= migration_version %>
+class CreateLlmCostTrackerAsyncIngestion < ActiveRecord::Migration[<%= ActiveRecord::Migration.current_version %>]
   def change
     create_table :llm_cost_tracker_ingestion_inbox_entries do |t|
       t.string   :event_id, null: false

data/lib/llm_cost_tracker/generators/llm_cost_tracker/templates/create_llm_cost_tracker_call_rollups.rb.erb

@@ -1,4 +1,4 @@
-class CreateLlmCostTrackerCallRollups < ActiveRecord::Migration<%= migration_version %>
+class CreateLlmCostTrackerCallRollups < ActiveRecord::Migration[<%= ActiveRecord::Migration.current_version %>]
   def change
     create_table :llm_cost_tracker_call_rollups do |t|
       t.string :period, null: false

data/lib/llm_cost_tracker/generators/llm_cost_tracker/templates/create_llm_cost_tracker_calls.rb.erb

@@ -1,7 +1,7 @@
 require "llm_cost_tracker/charges/cost_status"
 require "llm_cost_tracker/ledger/schema/adapter"
 
-class CreateLlmCostTrackerCalls < ActiveRecord::Migration<%= migration_version %>
+class CreateLlmCostTrackerCalls < ActiveRecord::Migration[<%= ActiveRecord::Migration.current_version %>]
   def change
     create_table :llm_cost_tracker_calls do |t|
       t.string  :event_id,      null: false

data/lib/llm_cost_tracker/generators/llm_cost_tracker/templates/initializer.rb.erb

@@ -50,8 +50,9 @@ LlmCostTracker.configure do |config|
 
   # One budget applied to every distinct value of a tag: each tenant gets its own
   # 1000 a month, not 1000 shared between them. Declare as many tags as you need; a rule
-  # can set its own behavior and callback instead of following the global ones. Needs the
-  # cost columns from `bin/rails generate llm_cost_tracker:upgrade_per_tag_budgets`.
+  # can set its own behavior and callback instead of following the global ones. Installs created
+  # before v0.14 add the cost columns with `bin/rails generate llm_cost_tracker:upgrade_per_tag_budgets`
+  # and `bin/rails db:migrate`, then run `bin/rails llm_cost_tracker:backfill_tag_costs` so earlier calls count.
   # Budget high-cardinality tags only — a tag with a handful of values covers most of the
   # ledger, so its check cannot use an index and every call pays for a near-full scan.
   # config.budgets.per_tag = {
@@ -63,9 +64,9 @@ LlmCostTracker.configure do |config|
   # }
 
   # Where budget checks read period spend from. :ledger sums llm_cost_tracker_calls on
-  # every check; :cache also keeps running totals in llm_cost_tracker_call_rollups and
-  # takes the greater of the two, so it guards against drift but does not make checks
-  # faster. Generate the table first with `bin/rails generate llm_cost_tracker:call_rollups`.
+  # every check; :cache reads earlier days from llm_cost_tracker_call_rollups and sums
+  # only today's calls. Run `bin/rails generate llm_cost_tracker:call_rollups`, deploy
+  # this setting, then run `bin/rails llm_cost_tracker:rebuild_rollups`.
   # config.budgets.totals_source = :cache
 
   # Ingestion path — :inline (default) writes events synchronously from the request
@@ -74,13 +75,14 @@ LlmCostTracker.configure do |config|
   # inbox/leases tables created by `bin/rails generate llm_cost_tracker:async_ingestion`.
   # Synchronous inbox writes use a dedicated ActiveRecord pool (defaults to 2 connections)
   # so they don't compete with request threads for the default pool when a tracked call
-  # happens inside an open caller transaction. Bump ingestion.pool_size if your Puma
-  # worker count outgrows that.
+  # happens inside an open caller transaction. The pool is per process: bump ingestion.pool_size
+  # if concurrent tracked calls in one process (e.g. Puma threads) outgrow that.
   # config.ingestion.mode = :async
   # config.ingestion.pool_size = 5
 
-  # Register OpenAI-compatible gateway hosts so their calls are attributed to a
-  # provider name instead of the generic openai_compatible bucket.
+  # Register OpenAI-compatible gateway hosts beyond the built-in OpenRouter, DeepSeek and Groq;
+  # the Faraday middleware captures only listed hosts. Its calls, and openai gem calls whose
+  # base_url is a listed host, are recorded under the provider name given here.
   # config.capture.openai_compatible_providers["llm.my-company.com"] = "internal_gateway"
 
   # Streaming calls only report token usage when the request asks for it. Leave this

data/lib/llm_cost_tracker/generators/llm_cost_tracker/templates/upgrade_call_rollups_provider.rb.erb

@@ -1,6 +1,6 @@
 require "llm_cost_tracker/ledger/schema/adapter"
 
-class UpgradeLlmCostTrackerCallRollupsProvider < ActiveRecord::Migration<%= migration_version %>
+class UpgradeLlmCostTrackerCallRollupsProvider < ActiveRecord::Migration[<%= ActiveRecord::Migration.current_version %>]
   disable_ddl_transaction!
 
   TABLE = :llm_cost_tracker_call_rollups

data/lib/llm_cost_tracker/generators/llm_cost_tracker/templates/upgrade_call_tags_key_value_index.rb.erb

@@ -1,6 +1,6 @@
 require "llm_cost_tracker/ledger/schema/adapter"
 
-class UpgradeLlmCostTrackerCallTagsKeyValueIndex < ActiveRecord::Migration<%= migration_version %>
+class UpgradeLlmCostTrackerCallTagsKeyValueIndex < ActiveRecord::Migration[<%= ActiveRecord::Migration.current_version %>]
   disable_ddl_transaction!
 
   TABLE = :llm_cost_tracker_call_tags

data/lib/llm_cost_tracker/generators/llm_cost_tracker/templates/upgrade_image_tokens.rb.erb

@@ -1,4 +1,4 @@
-class UpgradeLlmCostTrackerImageTokens < ActiveRecord::Migration<%= migration_version %>
+class UpgradeLlmCostTrackerImageTokens < ActiveRecord::Migration[<%= ActiveRecord::Migration.current_version %>]
   TABLE = :llm_cost_tracker_calls
   COLUMNS = %i[image_input_tokens image_output_tokens].freeze
 

data/lib/llm_cost_tracker/generators/llm_cost_tracker/templates/upgrade_indexes.rb.erb

@@ -1,6 +1,6 @@
 require "llm_cost_tracker/ledger/schema/adapter"
 
-class UpgradeLlmCostTrackerIndexes < ActiveRecord::Migration<%= migration_version %>
+class UpgradeLlmCostTrackerIndexes < ActiveRecord::Migration[<%= ActiveRecord::Migration.current_version %>]
   disable_ddl_transaction!
 
   CALLS = :llm_cost_tracker_calls

data/lib/llm_cost_tracker/generators/llm_cost_tracker/templates/upgrade_per_tag_budgets.rb.erb

@@ -1,6 +1,6 @@
 require "llm_cost_tracker/ledger/schema/adapter"
 
-class UpgradeLlmCostTrackerPerTagBudgets < ActiveRecord::Migration<%= migration_version %>
+class UpgradeLlmCostTrackerPerTagBudgets < ActiveRecord::Migration[<%= ActiveRecord::Migration.current_version %>]
   disable_ddl_transaction!
 
   TABLE = :llm_cost_tracker_call_tags

data/lib/llm_cost_tracker/generators/llm_cost_tracker/upgrade_call_rollups_provider_generator.rb

@@ -18,12 +18,6 @@ module LlmCostTracker
           "db/migrate/upgrade_llm_cost_tracker_call_rollups_provider.rb"
         )
       end
-
-      private
-
-      def migration_version
-        "[#{ActiveRecord::VERSION::MAJOR}.#{ActiveRecord::VERSION::MINOR}]"
-      end
     end
   end
 end

data/lib/llm_cost_tracker/generators/llm_cost_tracker/upgrade_call_tags_key_value_index_generator.rb

@@ -19,12 +19,6 @@ module LlmCostTracker
           "db/migrate/upgrade_llm_cost_tracker_call_tags_key_value_index.rb"
         )
       end
-
-      private
-
-      def migration_version
-        "[#{ActiveRecord::VERSION::MAJOR}.#{ActiveRecord::VERSION::MINOR}]"
-      end
     end
   end
 end

data/lib/llm_cost_tracker/generators/llm_cost_tracker/upgrade_image_tokens_generator.rb

@@ -18,12 +18,6 @@ module LlmCostTracker
           "db/migrate/upgrade_llm_cost_tracker_image_tokens.rb"
         )
       end
-
-      private
-
-      def migration_version
-        "[#{ActiveRecord::VERSION::MAJOR}.#{ActiveRecord::VERSION::MINOR}]"
-      end
     end
   end
 end

data/lib/llm_cost_tracker/generators/llm_cost_tracker/upgrade_indexes_generator.rb

@@ -10,8 +10,9 @@ module LlmCostTracker
 
       source_root File.expand_path("templates", __dir__)
 
-      desc "Adds a partial index over unpriced calls and drops the unused ingestion inbox lock index" \
-           "scanning the whole ledger."
+      desc "Adds the unpriced-calls index (partial on PostgreSQL, so backfill_unknown_pricing stops scanning " \
+           "the whole ledger; a plain id index on MySQL, which has no partial indexes) and drops the unused " \
+           "ingestion inbox lock index."
 
       def create_migration_file
         migration_template(
@@ -19,12 +20,6 @@ module LlmCostTracker
           "db/migrate/upgrade_llm_cost_tracker_indexes.rb"
         )
       end
-
-      private
-
-      def migration_version
-        "[#{ActiveRecord::VERSION::MAJOR}.#{ActiveRecord::VERSION::MINOR}]"
-      end
     end
   end
 end

data/lib/llm_cost_tracker/generators/llm_cost_tracker/upgrade_per_tag_budgets_generator.rb

@@ -10,8 +10,9 @@ module LlmCostTracker
 
       source_root File.expand_path("templates", __dir__)
 
-      desc "Carries each call's cost and time onto its tag rows so per-tag budgets read " \
-           "llm_cost_tracker_call_tags without a join. Required when config.budgets.per_tag is set."
+      desc "Adds cost and time columns to llm_cost_tracker_call_tags so per-tag budgets read it " \
+           "without a join; run llm_cost_tracker:backfill_tag_costs afterwards so earlier calls count. " \
+           "Required for config.budgets.per_tag on installs created before v0.14."
 
       def create_migration_file
         migration_template(
@@ -19,12 +20,6 @@ module LlmCostTracker
           "db/migrate/upgrade_llm_cost_tracker_per_tag_budgets.rb"
         )
       end
-
-      private
-
-      def migration_version
-        "[#{ActiveRecord::VERSION::MAJOR}.#{ActiveRecord::VERSION::MINOR}]"
-      end
     end
   end
 end

data/lib/llm_cost_tracker/ingestion/inbox.rb

@@ -18,7 +18,7 @@ module LlmCostTracker
         end
 
         def event_from_row(row)
-          payload = JSON.parse(row.payload, symbolize_names: true)
+          payload = JSON.parse(row.payload, symbolize_names: true, allow_duplicate_key: true)
           schema_version = payload[:schema_version]
           unless schema_version == PAYLOAD_SCHEMA_VERSION
             raise LlmCostTracker::Error, "unsupported ledger inbox payload schema version #{schema_version.inspect}"
@@ -69,13 +69,7 @@ module LlmCostTracker
         end
 
         def payload_for(event)
-          event.to_h.merge(
-            schema_version: PAYLOAD_SCHEMA_VERSION,
-            event_id: event.event_id,
-            provider: event.provider,
-            model: event.model,
-            tracked_at: event.tracked_at.iso8601(6)
-          )
+          event.to_h.merge(schema_version: PAYLOAD_SCHEMA_VERSION, tracked_at: event.tracked_at.iso8601(6))
         end
 
         def insert_row(row)

data/lib/llm_cost_tracker/ingestion.rb

@@ -28,6 +28,8 @@ module LlmCostTracker
         end
 
         guards_for_current_config.each do |schema_module, table_name|
+          next if schema_module == Ledger::Schema::CallRollups && !Ledger::Rollups.cache_active?
+
           errors = schema_module.current_schema_errors
           next if errors.empty?
 
@@ -124,7 +126,7 @@ module LlmCostTracker
 
       def cleanup_verification_call(response_id)
         relation = LlmCostTracker::Call.where(provider_response_id: response_id)
-        records = relation.select(:id, :tracked_at, :total_cost, :pricing_snapshot, :provider).to_a
+        records = relation.select(:id, *LlmCostTracker::Ledger::Rollups::SOURCE_COLUMNS).to_a
         return if records.empty?
 
         relation.delete_all

data/lib/llm_cost_tracker/integrations/anthropic.rb

@@ -1,7 +1,6 @@
 # frozen_string_literal: true
 
 require_relative "base"
-require_relative "../providers/anthropic/usage_extractor"
 require_relative "../providers/anthropic/response_parser"
 
 module LlmCostTracker
@@ -17,7 +16,8 @@ module LlmCostTracker
             patch_target("Anthropic::Resources::Messages", with: MessagesPatch),
             patch_target("Anthropic::Resources::Beta::Messages", with: MessagesPatch, optional: true),
             patch_target("Anthropic::Resources::Messages::Batches", with: BatchesPatch, optional: true),
-            patch_target("Anthropic::Resources::Beta::Messages::Batches", with: BatchesPatch, optional: true)
+            patch_target("Anthropic::Resources::Beta::Messages::Batches", with: BatchesPatch, optional: true),
+            patch_target("Anthropic::BetaRefusalFallbackMiddleware", with: FallbackMiddlewarePatch, optional: true)
           ]
         end
 
@@ -37,7 +37,8 @@ module LlmCostTracker
                 model: message.model || request[:model],
                 provider_response_id: message.id,
                 usage_source: Usage::Source::SDK_RESPONSE,
-                request: request
+                request: request,
+                **response_fields(message)
               ),
               latency_ms: latency_ms
             )
@@ -61,21 +62,52 @@ module LlmCostTracker
             next unless usage
             next if usage.input_tokens.nil? && usage.output_tokens.nil?
 
-            usage_hash = usage.deep_to_h
-            LlmCostTracker::Tracker.record(
-              event: Providers::Anthropic::ResponseParser.event_from_usage(
-                usage: usage_hash,
+            record_once(
+              Providers::Anthropic::ResponseParser.event_from_usage(
+                usage: usage.deep_to_h.merge(service_tier: "batch"),
                 model: message.model,
                 provider_response_id: message.id,
                 usage_source: Usage::Source::SDK_BATCH_RESULT,
-                pricing_mode: "batch"
+                **response_fields(message)
+              )
+            )
+          end
+        end
+
+        def record_refused_hop(request, response)
+          record_safely { record_message(response.parse, request: request.body, latency_ms: nil) }
+        rescue LlmCostTracker::BudgetExceededError, LlmCostTracker::UnknownPricingError
+          nil
+        end
+
+        def record_refused_stream_hop(hop)
+          return unless active?
+
+          usage = hop.dig(:refused, :usage).to_h.deep_symbolize_keys
+          output = usage[:output_tokens].to_i.positive?
+          usage = usage.slice(:server_tool_use) if output
+          return if output && Providers::Anthropic::UsageExtractor.service_line_items(usage).empty?
+
+          record_safely do
+            LlmCostTracker::Tracker.record(
+              event: Providers::Anthropic::ResponseParser.event_from_usage(
+                usage: usage,
+                model: hop[:model],
+                provider_response_id: nil,
+                usage_source: Usage::Source::STREAM_FINAL,
+                stream: true,
+                stop_reason: ("refusal" unless output),
+                refusal_category: hop.dig(:refused, :stop_details, "category")
               )
             )
           end
+        rescue LlmCostTracker::BudgetExceededError, LlmCostTracker::UnknownPricingError
+          nil
         end
 
-        def stream_pricing_mode(request)
-          Providers::Anthropic::UsageExtractor.pricing_mode(request: request || {}, usage: nil)
+        def response_fields(message)
+          { stop_reason: message.stop_reason, refusal_category: message.stop_details&.category,
+            content: message.deep_to_h[:content] }
         end
       end
 
@@ -109,6 +141,27 @@ module LlmCostTracker
         end
       end
 
+      module FallbackMiddlewarePatch
+        def call(req, nxt)
+          return super if req.streaming? || !LlmCostTracker::Integrations::Anthropic.active?
+
+          hops = []
+          response = super(req, ->(hop_req) { nxt.call(hop_req).tap { |hop| hops << [hop_req, hop] } })
+          answered = hops.select { |_hop_req, hop| hop.status < 300 }
+          answered[...-1].each { |hop| LlmCostTracker::Integrations::Anthropic.record_refused_hop(*hop) }
+          response
+        end
+
+        private
+
+        def consume_hop(*args, **kwargs)
+          refused = Thread.current[:llm_cost_tracker_refused_hop]
+          Thread.current[:llm_cost_tracker_refused_hop] = nil
+          LlmCostTracker::Integrations::Anthropic.record_refused_stream_hop(refused) if refused && kwargs[:splice]
+          super.tap { |hop| Thread.current[:llm_cost_tracker_refused_hop] = hop if hop[:refused] }
+        end
+      end
+
       module BatchesPatch
         def results_streaming(*args, **kwargs)
           raw = super