active_harness 0.2.27 → 0.2.28

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: 5b406bc5feeebdb4cc26e215a65d25dd931605684c6f05ee7cdd71de5910a1d6
-  data.tar.gz: a34d2e22840ae906d47b0e27fe271d97b42d97a0fce64a19a58e15c296f00bf7
+  metadata.gz: ce3118f24bd7e48d024a4b5c14d0e01531e453be2f56479918759dc425bc932d
+  data.tar.gz: 561d4d3f46b98a224d53ed964c319fbb1a73bffc5c6a3dd189d01e24d1af2427
 SHA512:
-  metadata.gz: f53ee25213ac7404840349cee732aa59b1dc25168b4e362028e7542fd51a0c51e9f2a2af4b8cbad9a552e74a9c4769d147b7d642e3b051df9c293af58bc87f34
-  data.tar.gz: d56efdd52b38afd9153903611fd860e316a59eb8f5d5b5c3ed0510b22647b8008a3b22f4e0c8430322548bb82a50c69cc4d6c086e8d965653c22d795fd43cae7
+  metadata.gz: 29f2a581e10290eb176a80e142a120ba5e9c9f0bc9e5decfacb7237fd11f35621bd35a64ebca2999399edd5f5865aec343edb8a11c8040e539329857b51c6d23
+  data.tar.gz: 86ef4129befa8b92af78f5666a4f6346f9d9eba46f39e880cf9968d03b7339f6de9620a0b558d67571bd4650425da0ea3455b8251c356e366717c87233dc3516

data/lib/active_harness/agent/prompt.rb

@@ -49,11 +49,15 @@ module ActiveHarness
       prompt
     end
 
+    # Injects agent state into a prompt class instance before #call.
+    # Available in prompt classes: @input, @context, @params, @memory, @context_window, @config
     def inject_agent_state(obj)
-      obj.instance_variable_set(:@input,   @input)
-      obj.instance_variable_set(:@context, @context)
-      obj.instance_variable_set(:@config,  @config)
-      obj.instance_variable_set(:@memory,  @memory)
+      obj.instance_variable_set(:@input,          @input)
+      obj.instance_variable_set(:@context,        @context)
+      obj.instance_variable_set(:@params,         @params)
+      obj.instance_variable_set(:@config,         @config)
+      obj.instance_variable_set(:@memory,         @memory)
+      obj.instance_variable_set(:@context_window, @context_window)
     end
   end
 end

data/lib/active_harness/agent.rb

@@ -54,6 +54,7 @@ module ActiveHarness
                   :params,
                   :memory
     attr_reader   :result,
+                  :context_window,
                   :token_stream,
                   :event_stream
 
@@ -76,9 +77,10 @@ module ActiveHarness
       @context         = context
       @params          = params
       @memory          = memory
-      @models_override = Array(models) if models
-      @token_stream    = streams[:token]
-      @event_stream    = streams[:agent]
+      @models_override  = Array(models) if models
+      @context_window   = lookup_context_window(self.models.to_a.first)
+      @token_stream     = streams[:token]
+      @event_stream     = streams[:agent]
       fire(:setup)
     end
 
@@ -144,9 +146,10 @@ module ActiveHarness
     end
 
     def build_result(response, entry, attempts, elapsed)
-      raw    = response[:content]
+      raw       = response[:content]
       processed = parse_output(raw)
-      usage  = response[:usage]
+      usage     = response[:usage]
+      cw        = lookup_context_window(entry)
 
       Result.new(
         input:          @input,
@@ -160,10 +163,18 @@ module ActiveHarness
         attempts:       attempts,
         execution_time: elapsed,
         usage:          usage,
-        cost:           calculate_cost(entry[:model], usage)
+        cost:           calculate_cost(entry[:model], usage),
+        context_window: cw
       )
     end
 
+    def lookup_context_window(entry)
+      return nil unless entry
+      Costs.find(entry[:model])&.context_window
+    rescue StandardError
+      nil
+    end
+
     def normalize_input!
       return if @config.fetch(:normalize_input, true) == false
       @input = @input&.strip&.gsub(/\s+/, " ")

data/lib/active_harness/costs.rb

@@ -62,12 +62,15 @@ module ActiveHarness
       :output_per_million,
       :cache_read_input_per_million,
       :cache_write_input_per_million,
+      :context_window,
+      :max_output_tokens,
       keyword_init: true
     ) do
       def inspect
         parts = ["id=#{id.inspect}", "provider=#{provider.inspect}"]
         parts << "input=$#{input_per_million}/M"  if input_per_million
         parts << "output=$#{output_per_million}/M" if output_per_million
+        parts << "ctx=#{context_window}"           if context_window
         "#<ModelCost #{parts.join(' ')}>"
       end
     end
@@ -230,10 +233,12 @@ module ActiveHarness
             }.compact
 
             {
-              id:       m[:id],
-              name:     m[:name] || m[:id],
-              provider: ah_provider,
-              pricing:  standard.any? ? { text_tokens: { standard: standard } } : {}
+              id:                m[:id],
+              name:              m[:name] || m[:id],
+              provider:          ah_provider,
+              context_window:    m[:context_window] || m.dig(:limit, :context),
+              max_output_tokens: m[:max_output_tokens] || m.dig(:limit, :output),
+              pricing:           standard.any? ? { text_tokens: { standard: standard } } : {}
             }
           end
         end
@@ -248,7 +253,9 @@ module ActiveHarness
           input_per_million:             standard[:input_per_million],
           output_per_million:            standard[:output_per_million],
           cache_read_input_per_million:  standard[:cache_read_input_per_million],
-          cache_write_input_per_million: standard[:cache_write_input_per_million]
+          cache_write_input_per_million: standard[:cache_write_input_per_million],
+          context_window:                raw[:context_window],
+          max_output_tokens:             raw[:max_output_tokens]
         )
       end
 

data/lib/active_harness/memory.rb

@@ -138,13 +138,15 @@ module ActiveHarness
 
     # Returns messages array for LLM consumption, respecting depth.
     # Optional filters:
-    #   filter: ->(turn) { turn[:agent] == "SupportAgent" }
-    #   since:  Time.now - 3600
-    def to_messages(filter: nil, since: nil)
+    #   filter:       ->(turn) { turn[:agent] == "SupportAgent" }
+    #   since:        Time.now - 3600
+    #   token_budget: 4000  # rough limit (chars / 4 estimate); trims oldest turns first
+    def to_messages(filter: nil, since: nil, token_budget: nil)
       turns = @turns.dup
       turns.select! { |t| filter.call(t) }   if filter
       turns.select! { |t| after?(t, since) } if since
       turns = turns.last(@depth)              if @depth
+      turns = trim_to_token_budget(turns, token_budget) if token_budget
 
       turns.flat_map do |t|
         [
@@ -210,5 +212,16 @@ module ActiveHarness
       turn_time = Time.parse(turn[:at].to_s) rescue nil
       turn_time ? turn_time >= time : true
     end
+
+    def trim_to_token_budget(turns, budget)
+      return turns if turns.empty?
+
+      total = 0
+      turns.reverse.take_while do |t|
+        tokens = ((t[:request].to_s.length + t[:response].to_s.length) / 4.0).ceil
+        total += tokens
+        total <= budget
+      end.reverse
+    end
   end
 end

data/lib/active_harness/result.rb

@@ -11,7 +11,7 @@ module ActiveHarness
     :input,
     :output,
     :processed,
-    :system_prompt, 
+    :system_prompt,
     :provider, :model,
     :temperature,
     :model_list,
@@ -19,5 +19,6 @@ module ActiveHarness
     :execution_time,
     :usage,
     :cost,
+    :context_window,
     keyword_init: true)
 end

data/lib/active_harness.rb

@@ -30,7 +30,7 @@ require_relative "active_harness/pipeline"
 require_relative "active_harness/railtie" if defined?(Rails::Railtie)
 
 module ActiveHarness
-  VERSION = "0.2.27"
+  VERSION = "0.2.28"
 
   class << self
     # Configure ActiveHarness.

metadata

@@ -1,7 +1,7 @@
 --- !ruby/object:Gem::Specification
 name: active_harness
 version: !ruby/object:Gem::Version
-  version: 0.2.27
+  version: 0.2.28
 platform: ruby
 authors:
 - the-teacher