feat: add GPT-6 mantle support, fix item/cost loss in Responses

lib/ruby_llm/active_record/chat_methods.rb

@@ -417,7 +417,7 @@ def prepare_content_for_storage(content)
         case content
         when RubyLLM::Content::Raw
           content_raw = content.value
-          content_text = nil
+          content_text = content.respond_to?(:text) ? content.text.presence : nil
         when RubyLLM::Content
           attachments = content.attachments if content.attachments.any?
           content_text = content.text

lib/ruby_llm/aliases.json

@@ -447,6 +447,15 @@
   "gpt-5.6-terra": {
     "bedrock": "openai.gpt-5.6-terra"
   },
+  "gpt-6-luna": {
+    "bedrock": "openai.gpt-6-luna"
+  },
+  "gpt-6-sol": {
+    "bedrock": "openai.gpt-6-sol"
+  },
+  "gpt-6.1-sol": {
+    "bedrock": "openai.gpt-6.1-sol"
+  },
   "gpt-audio": {
     "openai": "gpt-audio",
     "openrouter": "openai/gpt-audio"

lib/ruby_llm/chat.rb

@@ -224,6 +224,13 @@ def add_completion(response)
       response
     end
 
+    # Compacts the conversation through the provider's standalone compaction endpoint and
+    # returns the compaction Message without adding it to #messages; the caller decides
+    # whether to add it. Raises UnsupportedFeatureError for protocols without one.
+    def compact_context
+      @provider.compact(messages, model: @model, protocol: @protocol, headers: @headers)
+    end
+
     # The request this chat would send for its next completion.
     def render
       @provider.render(
@@ -361,9 +368,14 @@ def provider_completion(&)
     end
 
     def normalize_schema_response(response)
-      return unless @schema && response.content.is_a?(String) && !response.tool_call?
+      return unless @schema && !response.tool_call?
 
-      response.content = JSON.parse(response.content)
+      case response.content
+      when String
+        response.content = JSON.parse(response.content)
+      when RubyLLM::Protocols::Responses::OutputItems
+        response.content = JSON.parse(response.content.text)
+      end
     rescue JSON::ParserError
       # If parsing fails, keep content as string.
     end

lib/ruby_llm/error.rb

@@ -27,6 +27,7 @@ class PromptNotFoundError < StandardError; end
   class InvalidRoleError < StandardError; end
   class InvalidToolChoiceError < StandardError; end
   class ModelNotFoundError < StandardError; end
+  class UnsupportedFeatureError < StandardError; end
 
   # Raised when RubyLLM cannot format an attachment for the selected provider.
   class UnsupportedAttachmentError < StandardError

lib/ruby_llm/models.json

@@ -17978,6 +17978,209 @@
       "knowledge": null
     }
   },
+  {
+    "id": "openai.gpt-6-luna",
+    "name": "GPT-6 Luna",
+    "provider": "bedrock",
+    "family": "gpt",
+    "created_at": "2026-10-05 00:00:00 UTC",
+    "context_window": 1050000,
+    "max_output_tokens": 128000,
+    "knowledge_cutoff": null,
+    "modalities": {
+      "input": [
+        "text",
+        "image",
+        "pdf"
+      ],
+      "output": [
+        "text"
+      ]
+    },
+    "capabilities": [
+      "function_calling",
+      "structured_output",
+      "reasoning",
+      "vision"
+    ],
+    "pricing": {
+      "text_tokens": {
+        "standard": {
+          "input_per_million": 0.11,
+          "output_per_million": 0.55,
+          "cache_read_input_per_million": 0.011,
+          "cache_write_input_per_million": 0.1375
+        }
+      }
+    },
+    "metadata": {
+      "source": "manual",
+      "provider_id": "amazon-bedrock",
+      "open_weights": false,
+      "attachment": true,
+      "temperature": false,
+      "last_updated": "2026-10-05",
+      "reasoning_options": [
+        {
+          "type": "effort",
+          "values": [
+            "none",
+            "low",
+            "medium",
+            "high",
+            "xhigh",
+            "max"
+          ]
+        }
+      ],
+      "cost": {
+        "input": 0.11,
+        "output": 0.55,
+        "cache_read": 0.011,
+        "cache_write": 0.1375
+      },
+      "limit": {
+        "context": 1050000,
+        "output": 128000
+      },
+      "knowledge": null
+    }
+  },
+  {
+    "id": "openai.gpt-6-sol",
+    "name": "GPT-6 Sol",
+    "provider": "bedrock",
+    "family": "gpt",
+    "created_at": "2026-10-05 00:00:00 UTC",
+    "context_window": 1050000,
+    "max_output_tokens": 128000,
+    "knowledge_cutoff": null,
+    "modalities": {
+      "input": [
+        "text",
+        "image",
+        "pdf"
+      ],
+      "output": [
+        "text"
+      ]
+    },
+    "capabilities": [
+      "function_calling",
+      "structured_output",
+      "reasoning",
+      "vision"
+    ],
+    "pricing": {
+      "text_tokens": {
+        "standard": {
+          "input_per_million": 2.2,
+          "output_per_million": 11.0,
+          "cache_read_input_per_million": 0.22,
+          "cache_write_input_per_million": 2.75
+        }
+      }
+    },
+    "metadata": {
+      "source": "manual",
+      "provider_id": "amazon-bedrock",
+      "open_weights": false,
+      "attachment": true,
+      "temperature": false,
+      "last_updated": "2026-10-05",
+      "reasoning_options": [
+        {
+          "type": "effort",
+          "values": [
+            "none",
+            "low",
+            "medium",
+            "high",
+            "xhigh",
+            "max"
+          ]
+        }
+      ],
+      "cost": {
+        "input": 2.2,
+        "output": 11.0,
+        "cache_read": 0.22,
+        "cache_write": 2.75
+      },
+      "limit": {
+        "context": 1050000,
+        "output": 128000
+      },
+      "knowledge": null
+    }
+  },
+  {
+    "id": "openai.gpt-6.1-sol",
+    "name": "GPT-6.1 Sol",
+    "provider": "bedrock",
+    "family": "gpt",
+    "created_at": "2026-10-05 00:00:00 UTC",
+    "context_window": 1050000,
+    "max_output_tokens": 128000,
+    "knowledge_cutoff": null,
+    "modalities": {
+      "input": [
+        "text",
+        "image",
+        "pdf"
+      ],
+      "output": [
+        "text"
+      ]
+    },
+    "capabilities": [
+      "function_calling",
+      "structured_output",
+      "reasoning",
+      "vision"
+    ],
+    "pricing": {
+      "text_tokens": {
+        "standard": {
+          "input_per_million": 2.2,
+          "output_per_million": 11.0,
+          "cache_read_input_per_million": 0.11,
+          "cache_write_input_per_million": 2.75
+        }
+      }
+    },
+    "metadata": {
+      "source": "manual",
+      "provider_id": "amazon-bedrock",
+      "open_weights": false,
+      "attachment": true,
+      "temperature": false,
+      "last_updated": "2026-10-05",
+      "reasoning_options": [
+        {
+          "type": "effort",
+          "values": [
+            "low",
+            "medium",
+            "high",
+            "xhigh",
+            "max"
+          ]
+        }
+      ],
+      "cost": {
+        "input": 2.2,
+        "output": 11.0,
+        "cache_read": 0.11,
+        "cache_write": 2.75
+      },
+      "limit": {
+        "context": 1050000,
+        "output": 128000
+      },
+      "knowledge": null
+    }
+  },
   {
     "id": "openai.gpt-oss-120b",
     "name": "gpt-oss-120b",
@@ -67385,4 +67588,4 @@
       "owned_by": "xai"
     }
   }
-]
+]

lib/ruby_llm/protocols/mantle_responses.rb

@@ -4,47 +4,51 @@ module RubyLLM
   module Protocols
     # AWS Bedrock's bedrock-mantle endpoint, speaking the OpenAI Responses API.
     # Reachable only via bedrock-mantle (not bedrock-runtime), and only for models
-    # that physically cannot serve Converse (the GPT-5.x frontier
-    # family). Talks to the provider's mantle connection instead of the default
+    # that physically cannot serve Converse (the GPT-5.x and GPT-6.x frontier
+    # families). Talks to the provider's mantle connection instead of the default
     # bedrock-runtime connection, and SigV4-signs every request against the
     # "bedrock-mantle" service namespace instead of "bedrock".
     class MantleResponses < Responses
-      # Frontier openai.gpt-5.x models are served at /openai/v1/responses; every other
-      # mantle model (e.g. openai.gpt-oss-*) is served at /v1/responses. See the
-      # AWS Bedrock model cards for GPT-5.6.
+      # Frontier openai.gpt-5.x and openai.gpt-6.x models are served at /openai/v1/responses;
+      # every other mantle model (e.g. openai.gpt-oss-*) is served at /v1/responses. See the
+      # AWS Bedrock model cards for GPT-5.6 and GPT-6.
       #
       # Kept in sync with Providers::Bedrock::MANTLE_ONLY_MODEL_PATTERN — that pattern picks
       # mantle vs Converse, this one picks the /openai/v1 vs /v1 mantle path. They
       # coincide today but are distinct concepts; update both when a new frontier family lands.
-      FRONTIER_GPT5_PATTERN = /\Aopenai\.gpt-5/
+      FRONTIER_GPT_PATTERN = /\Aopenai\.gpt-[56]/
 
       def initialize(provider, model = nil)
         super
         @connection = provider.mantle_connection
       end
 
       # The '/v1/responses' branch is forward-looking: today Providers::Bedrock#protocol_for only
-      # ever routes /\Aopenai\.gpt-5/ ids here, so this branch is unreachable in production (e.g.
+      # ever routes /\Aopenai\.gpt-[56]/ ids here, so this branch is unreachable in production (e.g.
       # openai.gpt-oss-* still routes to Converse). It exists so this protocol is ready if a
       # future mantle-only, non-frontier model needs it, without another round of plumbing.
       def completion_url
-        FRONTIER_GPT5_PATTERN.match?(@model.id) ? '/openai/v1/responses' : '/v1/responses'
+        FRONTIER_GPT_PATTERN.match?(@model.id) ? '/openai/v1/responses' : '/v1/responses'
       end
 
       private
 
       def sync_response(payload, additional_headers = {})
-        super(payload, additional_headers.merge(mantle_signature_headers(payload)))
+        super(payload, additional_headers.merge(mantle_signature_headers(completion_url, payload)))
       end
 
       def stream_response(payload, additional_headers = {}, &)
-        super(payload, additional_headers.merge(mantle_signature_headers(payload)), &)
+        super(payload, additional_headers.merge(mantle_signature_headers(completion_url, payload)), &)
       end
 
-      def mantle_signature_headers(payload)
+      def compaction_response(payload, additional_headers = {})
+        super(payload, additional_headers.merge(mantle_signature_headers(compaction_url, payload)))
+      end
+
+      def mantle_signature_headers(path, payload)
         body = JSON.generate(payload)
         @provider.sign_headers(
-          'POST', completion_url, body,
+          'POST', path, body,
           base_url: @provider.mantle_api_base,
           service: Providers::Bedrock::MANTLE_SIGNING_SERVICE,
           region: @provider.mantle_region

lib/ruby_llm/protocols/responses.rb

@@ -12,6 +12,16 @@ class Responses < ChatCompletions
       include Responses::Streaming
       include Responses::Tools
 
+      # Compacts the conversation through the standalone /compact endpoint and returns the
+      # response.compaction object as an assistant Message holding OutputItems. GPT-6 only.
+      def compact(messages, headers: {})
+        unless gpt6_model?(model.id)
+          raise UnsupportedFeatureError, "Standalone compaction is only supported for GPT-6 models, not #{model.id}"
+        end
+
+        compaction_response(render_compaction_payload(messages), headers)
+      end
+
       # Reasoning models reject the temperature parameter on this API.
       def maybe_normalize_temperature(temperature, model)
         return super unless reasoning_model?(model.id)
@@ -21,6 +31,15 @@ def maybe_normalize_temperature(temperature, model)
         end
         nil
       end
+
+      private
+
+      def compaction_response(payload, additional_headers = {})
+        response = @connection.post compaction_url, payload do |req|
+          req.headers = additional_headers.merge(req.headers) unless additional_headers.empty?
+        end
+        parse_compaction_response response
+      end
     end
   end
 end