ruby_llm 1.9.2 → 1.10.0

data/lib/ruby_llm/providers/bedrock/streaming/payload_processing.rb

@@ -57,10 +57,15 @@ module RubyLLM
               role: :assistant,
               model_id: extract_model_id(data),
               content: extract_streaming_content(data),
+              thinking: Thinking.build(
+                text: extract_thinking_delta(data),
+                signature: extract_signature_delta(data)
+              ),
               input_tokens: extract_input_tokens(data),
               output_tokens: extract_output_tokens(data),
               cached_tokens: extract_cached_tokens(data),
               cache_creation_tokens: extract_cache_creation_tokens(data),
+              thinking_tokens: extract_thinking_tokens(data),
               tool_calls: extract_tool_calls(data)
             }
           end

data/lib/ruby_llm/providers/gemini/chat.rb

@@ -14,7 +14,7 @@ module RubyLLM
           "models/#{@model}:generateContent"
         end
 
-        def render_payload(messages, tools:, temperature:, model:, stream: false, schema: nil) # rubocop:disable Metrics/ParameterLists,Lint/UnusedMethodArgument
+        def render_payload(messages, tools:, temperature:, model:, stream: false, schema: nil, thinking: nil) # rubocop:disable Metrics/ParameterLists,Lint/UnusedMethodArgument
           @model = model.id
           payload = {
             contents: format_messages(messages),
@@ -24,11 +24,30 @@ module RubyLLM
           payload[:generationConfig][:temperature] = temperature unless temperature.nil?
 
           payload[:generationConfig].merge!(structured_output_config(schema, model)) if schema
+          payload[:generationConfig][:thinkingConfig] = build_thinking_config(model, thinking) if thinking&.enabled?
 
           payload[:tools] = format_tools(tools) if tools.any?
           payload
         end
 
+        def build_thinking_config(_model, thinking)
+          config = { includeThoughts: true }
+
+          config[:thinkingLevel] = resolve_effort_level(thinking) if thinking&.effort
+          config[:thinkingBudget] = resolve_budget(thinking) if thinking&.budget
+
+          config
+        end
+
+        def resolve_effort_level(thinking)
+          thinking.respond_to?(:effort) ? thinking.effort : thinking
+        end
+
+        def resolve_budget(thinking)
+          budget = thinking.respond_to?(:budget) ? thinking.budget : thinking
+          budget.is_a?(Integer) ? budget : nil
+        end
+
         private
 
         def format_messages(messages)
@@ -56,20 +75,43 @@ module RubyLLM
           elsif msg.tool_result?
             format_tool_result(msg)
           else
-            Media.format_content(msg.content)
+            format_message_parts(msg)
           end
         end
 
+        def format_message_parts(msg)
+          parts = []
+
+          parts << build_thought_part(msg.thinking) if msg.role == :assistant && msg.thinking
+
+          content_parts = Media.format_content(msg.content)
+          parts.concat(content_parts.is_a?(Array) ? content_parts : [content_parts])
+          parts
+        end
+
+        def build_thought_part(thinking)
+          part = { thought: true }
+          part[:text] = thinking.text if thinking.text
+          part[:thoughtSignature] = thinking.signature if thinking.signature
+          part
+        end
+
         def parse_completion_response(response)
           data = response.body
+          parts = data.dig('candidates', 0, 'content', 'parts') || []
           tool_calls = extract_tool_calls(data)
 
           Message.new(
             role: :assistant,
-            content: parse_content(data),
+            content: extract_text_parts(parts) || parse_content(data),
+            thinking: Thinking.build(
+              text: extract_thought_parts(parts),
+              signature: extract_thought_signature(parts)
+            ),
             tool_calls: tool_calls,
             input_tokens: data.dig('usageMetadata', 'promptTokenCount'),
             output_tokens: calculate_output_tokens(data),
+            thinking_tokens: data.dig('usageMetadata', 'thoughtsTokenCount'),
             model_id: data['modelVersion'] || response.env.url.path.split('/')[3].split(':')[0],
             raw: response
           )
@@ -93,6 +135,30 @@ module RubyLLM
           build_response_content(parts)
         end
 
+        def extract_text_parts(parts)
+          text_parts = parts.reject { |p| p['thought'] }
+          content = text_parts.filter_map { |p| p['text'] }.join
+          content.empty? ? nil : content
+        end
+
+        def extract_thought_parts(parts)
+          thought_parts = parts.select { |p| p['thought'] }
+          thoughts = thought_parts.filter_map { |p| p['text'] }.join
+          thoughts.empty? ? nil : thoughts
+        end
+
+        def extract_thought_signature(parts)
+          parts.each do |part|
+            signature = part['thoughtSignature'] ||
+                        part['thought_signature'] ||
+                        part.dig('functionCall', 'thoughtSignature') ||
+                        part.dig('functionCall', 'thought_signature')
+            return signature if signature
+          end
+
+          nil
+        end
+
         def function_call?(candidate)
           parts = candidate.dig('content', 'parts')
           parts&.any? { |p| p['functionCall'] }

data/lib/ruby_llm/providers/gemini/streaming.rb

@@ -10,12 +10,19 @@ module RubyLLM
         end
 
         def build_chunk(data)
+          parts = data.dig('candidates', 0, 'content', 'parts') || []
+
           Chunk.new(
             role: :assistant,
             model_id: extract_model_id(data),
-            content: extract_content(data),
+            content: extract_text_content(parts),
+            thinking: Thinking.build(
+              text: extract_thought_content(parts),
+              signature: extract_thought_signature(parts)
+            ),
             input_tokens: extract_input_tokens(data),
             output_tokens: extract_output_tokens(data),
+            thinking_tokens: data.dig('usageMetadata', 'thoughtsTokenCount'),
             tool_calls: extract_tool_calls(data)
           )
         end
@@ -26,6 +33,30 @@ module RubyLLM
           data['modelVersion']
         end
 
+        def extract_text_content(parts)
+          text_parts = parts.reject { |p| p['thought'] }
+          text = text_parts.filter_map { |p| p['text'] }.join
+          text.empty? ? nil : text
+        end
+
+        def extract_thought_content(parts)
+          thought_parts = parts.select { |p| p['thought'] }
+          thoughts = thought_parts.filter_map { |p| p['text'] }.join
+          thoughts.empty? ? nil : thoughts
+        end
+
+        def extract_thought_signature(parts)
+          parts.each do |part|
+            signature = part['thoughtSignature'] ||
+                        part['thought_signature'] ||
+                        part.dig('functionCall', 'thoughtSignature') ||
+                        part.dig('functionCall', 'thought_signature')
+            return signature if signature
+          end
+
+          nil
+        end
+
         def extract_content(data)
           return nil unless data['candidates']&.any?
 

data/lib/ruby_llm/providers/gemini/tools.rb

@@ -13,7 +13,7 @@ module RubyLLM
           }]
         end
 
-        def format_tool_call(msg)
+        def format_tool_call(msg) # rubocop:disable Metrics/PerceivedComplexity
           parts = []
 
           if msg.content && !(msg.content.respond_to?(:empty?) && msg.content.empty?)
@@ -21,13 +21,24 @@ module RubyLLM
             parts.concat(formatted_content.is_a?(Array) ? formatted_content : [formatted_content])
           end
 
+          fallback_signature = msg.thinking&.signature
+          used_fallback = false
+
           msg.tool_calls.each_value do |tool_call|
-            parts << {
+            part = {
               functionCall: {
                 name: tool_call.name,
                 args: tool_call.arguments
               }
             }
+
+            signature = tool_call.thought_signature
+            if signature.nil? && fallback_signature && !used_fallback
+              signature = fallback_signature
+              used_fallback = true
+            end
+            part[:thoughtSignature] = signature if signature
+            parts << part
           end
 
           parts
@@ -61,11 +72,13 @@ module RubyLLM
             next unless function_data
 
             id = SecureRandom.uuid
+            thought_signature = part['thoughtSignature'] || part['thought_signature']
 
             result[id] = ToolCall.new(
               id:,
               name: function_data['name'],
-              arguments: function_data['args'] || {}
+              arguments: function_data['args'] || {},
+              thought_signature: thought_signature
             )
           end
 

data/lib/ruby_llm/providers/gpustack/chat.rb

@@ -14,7 +14,7 @@ module RubyLLM
               content: GPUStack::Media.format_content(msg.content),
               tool_calls: format_tool_calls(msg.tool_calls),
               tool_call_id: msg.tool_call_id
-            }.compact
+            }.compact.merge(OpenAI::Chat.format_thinking(msg))
           end
         end
 

data/lib/ruby_llm/providers/mistral/chat.rb

@@ -11,13 +11,70 @@ module RubyLLM
           role.to_s
         end
 
+        def format_messages(messages)
+          messages.map do |msg|
+            {
+              role: format_role(msg.role),
+              content: format_content_with_thinking(msg),
+              tool_calls: OpenAI::Tools.format_tool_calls(msg.tool_calls),
+              tool_call_id: msg.tool_call_id
+            }.compact
+          end
+        end
+
         # rubocop:disable Metrics/ParameterLists
-        def render_payload(messages, tools:, temperature:, model:, stream: false, schema: nil)
+        def render_payload(messages, tools:, temperature:, model:, stream: false, schema: nil, thinking: nil)
           payload = super
           payload.delete(:stream_options)
+          payload.delete(:reasoning_effort)
+          warn_on_unsupported_thinking(model, thinking)
           payload
         end
         # rubocop:enable Metrics/ParameterLists
+
+        def format_content_with_thinking(msg)
+          formatted_content = OpenAI::Media.format_content(msg.content)
+          return formatted_content unless msg.role == :assistant && msg.thinking
+
+          content_blocks = build_thinking_blocks(msg.thinking)
+          append_formatted_content(content_blocks, formatted_content)
+
+          content_blocks
+        end
+
+        def warn_on_unsupported_thinking(model, thinking)
+          return unless thinking&.enabled?
+          return if model.id.to_s.include?('magistral')
+
+          RubyLLM.logger.warn(
+            'Mistral thinking is only supported on Magistral models. ' \
+            "Ignoring thinking settings for #{model.id}."
+          )
+        end
+
+        def build_thinking_blocks(thinking)
+          return [] unless thinking
+
+          if thinking.text
+            [{
+              type: 'thinking',
+              thinking: [{ type: 'text', text: thinking.text }],
+              signature: thinking.signature
+            }.compact]
+          elsif thinking.signature
+            [{ type: 'thinking', signature: thinking.signature }]
+          else
+            []
+          end
+        end
+
+        def append_formatted_content(content_blocks, formatted_content)
+          if formatted_content.is_a?(Array)
+            content_blocks.concat(formatted_content)
+          elsif formatted_content
+            content_blocks << { type: 'text', text: formatted_content }
+          end
+        end
       end
     end
   end

data/lib/ruby_llm/providers/ollama/chat.rb

@@ -14,7 +14,7 @@ module RubyLLM
               content: Ollama::Media.format_content(msg.content),
               tool_calls: format_tool_calls(msg.tool_calls),
               tool_call_id: msg.tool_call_id
-            }.compact
+            }.compact.merge(OpenAI::Chat.format_thinking(msg))
           end
         end
 

data/lib/ruby_llm/providers/openai/capabilities.rb

@@ -224,8 +224,8 @@ module RubyLLM
         end
 
         def self.normalize_temperature(temperature, model_id)
-          if model_id.match?(/^(o\d|gpt-5)/)
-            RubyLLM.logger.debug "Model #{model_id} requires temperature=1.0, ignoring provided value"
+          if model_id.match?(/^(o\d|gpt-5)/) && !temperature.nil? && !temperature_close_to_one?(temperature)
+            RubyLLM.logger.debug "Model #{model_id} requires temperature=1.0, setting that instead."
             1.0
           elsif model_id.match?(/-search/)
             RubyLLM.logger.debug "Model #{model_id} does not accept temperature parameter, removing"
@@ -235,6 +235,10 @@ module RubyLLM
           end
         end
 
+        def self.temperature_close_to_one?(temperature)
+          (temperature.to_f - 1.0).abs <= Float::EPSILON
+        end
+
         def modalities_for(model_id)
           modalities = {
             input: ['text'],

data/lib/ruby_llm/providers/openai/chat.rb

@@ -11,7 +11,7 @@ module RubyLLM
 
         module_function
 
-        def render_payload(messages, tools:, temperature:, model:, stream: false, schema: nil) # rubocop:disable Metrics/ParameterLists
+        def render_payload(messages, tools:, temperature:, model:, stream: false, schema: nil, thinking: nil) # rubocop:disable Metrics/ParameterLists
           payload = {
             model: model.id,
             messages: format_messages(messages),
@@ -34,6 +34,9 @@ module RubyLLM
             }
           end
 
+          effort = resolve_effort(thinking)
+          payload[:reasoning_effort] = effort if effort
+
           payload[:stream_options] = { include_usage: true } if stream
           payload
         end
@@ -49,15 +52,21 @@ module RubyLLM
 
           usage = data['usage'] || {}
           cached_tokens = usage.dig('prompt_tokens_details', 'cached_tokens')
+          thinking_tokens = usage.dig('completion_tokens_details', 'reasoning_tokens')
+          content, thinking_from_blocks = extract_content_and_thinking(message_data['content'])
+          thinking_text = thinking_from_blocks || extract_thinking_text(message_data)
+          thinking_signature = extract_thinking_signature(message_data)
 
           Message.new(
             role: :assistant,
-            content: message_data['content'],
+            content: content,
+            thinking: Thinking.build(text: thinking_text, signature: thinking_signature),
             tool_calls: parse_tool_calls(message_data['tool_calls']),
             input_tokens: usage['prompt_tokens'],
             output_tokens: usage['completion_tokens'],
             cached_tokens: cached_tokens,
             cache_creation_tokens: 0,
+            thinking_tokens: thinking_tokens,
             model_id: data['model'],
             raw: response
           )
@@ -70,7 +79,7 @@ module RubyLLM
               content: Media.format_content(msg.content),
               tool_calls: format_tool_calls(msg.tool_calls),
               tool_call_id: msg.tool_call_id
-            }.compact
+            }.compact.merge(format_thinking(msg))
           end
         end
 
@@ -82,6 +91,81 @@ module RubyLLM
             role.to_s
           end
         end
+
+        def resolve_effort(thinking)
+          return nil unless thinking
+
+          thinking.respond_to?(:effort) ? thinking.effort : thinking
+        end
+
+        def format_thinking(msg)
+          return {} unless msg.role == :assistant
+
+          thinking = msg.thinking
+          return {} unless thinking
+
+          payload = {}
+          if thinking.text
+            payload[:reasoning] = thinking.text
+            payload[:reasoning_content] = thinking.text
+          end
+          payload[:reasoning_signature] = thinking.signature if thinking.signature
+          payload
+        end
+
+        def extract_thinking_text(message_data)
+          candidate = message_data['reasoning_content'] || message_data['reasoning'] || message_data['thinking']
+          candidate.is_a?(String) ? candidate : nil
+        end
+
+        def extract_thinking_signature(message_data)
+          candidate = message_data['reasoning_signature'] || message_data['signature']
+          candidate.is_a?(String) ? candidate : nil
+        end
+
+        def extract_content_and_thinking(content)
+          return extract_think_tag_content(content) if content.is_a?(String)
+          return [content, nil] unless content.is_a?(Array)
+
+          text = extract_text_from_blocks(content)
+          thinking = extract_thinking_from_blocks(content)
+
+          [text.empty? ? nil : text, thinking.empty? ? nil : thinking]
+        end
+
+        def extract_text_from_blocks(blocks)
+          blocks.filter_map do |block|
+            block['text'] if block['type'] == 'text' && block['text'].is_a?(String)
+          end.join
+        end
+
+        def extract_thinking_from_blocks(blocks)
+          blocks.filter_map do |block|
+            next unless block['type'] == 'thinking'
+
+            extract_thinking_text_from_block(block)
+          end.join
+        end
+
+        def extract_thinking_text_from_block(block)
+          thinking_block = block['thinking']
+          return thinking_block if thinking_block.is_a?(String)
+
+          if thinking_block.is_a?(Array)
+            return thinking_block.filter_map { |item| item['text'] if item['type'] == 'text' }.join
+          end
+
+          block['text'] if block['text'].is_a?(String)
+        end
+
+        def extract_think_tag_content(text)
+          return [text, nil] unless text.include?('<think>')
+
+          thinking = text.scan(%r{<think>(.*?)</think>}m).join
+          content = text.gsub(%r{<think>.*?</think>}m, '').strip
+
+          [content.empty? ? nil : content, thinking.empty? ? nil : thinking]
+        end
       end
     end
   end

data/lib/ruby_llm/providers/openai/streaming.rb

@@ -14,16 +14,24 @@ module RubyLLM
         def build_chunk(data)
           usage = data['usage'] || {}
           cached_tokens = usage.dig('prompt_tokens_details', 'cached_tokens')
+          delta = data.dig('choices', 0, 'delta') || {}
+          content_source = delta['content'] || data.dig('choices', 0, 'message', 'content')
+          content, thinking_from_blocks = OpenAI::Chat.extract_content_and_thinking(content_source)
 
           Chunk.new(
             role: :assistant,
             model_id: data['model'],
-            content: data.dig('choices', 0, 'delta', 'content'),
-            tool_calls: parse_tool_calls(data.dig('choices', 0, 'delta', 'tool_calls'), parse_arguments: false),
+            content: content,
+            thinking: Thinking.build(
+              text: thinking_from_blocks || delta['reasoning_content'] || delta['reasoning'],
+              signature: delta['reasoning_signature']
+            ),
+            tool_calls: parse_tool_calls(delta['tool_calls'], parse_arguments: false),
             input_tokens: usage['prompt_tokens'],
             output_tokens: usage['completion_tokens'],
             cached_tokens: cached_tokens,
-            cache_creation_tokens: 0
+            cache_creation_tokens: 0,
+            thinking_tokens: usage.dig('completion_tokens_details', 'reasoning_tokens')
           )
         end
 

data/lib/ruby_llm/providers/openai/temperature.rb

@@ -0,0 +1,28 @@
+# frozen_string_literal: true
+
+module RubyLLM
+  module Providers
+    class OpenAI
+      # Normalizes temperature for OpenAI models with provider-specific requirements.
+      module Temperature
+        module_function
+
+        def normalize(temperature, model_id)
+          if model_id.match?(/^(o\d|gpt-5)/) && !temperature.nil? && !temperature_close_to_one?(temperature)
+            RubyLLM.logger.debug "Model #{model_id} requires temperature=1.0, setting that instead."
+            1.0
+          elsif model_id.include?('-search')
+            RubyLLM.logger.debug "Model #{model_id} does not accept temperature parameter, removing"
+            nil
+          else
+            temperature
+          end
+        end
+
+        def temperature_close_to_one?(temperature)
+          (temperature.to_f - 1.0).abs <= Float::EPSILON
+        end
+      end
+    end
+  end
+end

data/lib/ruby_llm/providers/openai.rb

@@ -27,7 +27,7 @@ module RubyLLM
       end
 
       def maybe_normalize_temperature(temperature, model)
-        OpenAI::Capabilities.normalize_temperature(temperature, model.id)
+        OpenAI::Temperature.normalize(temperature, model.id)
       end
 
       class << self