DUX-9411: Add Bedrock bedrock-mantle endpoint support for GPT-5.6 models

lib/ruby_llm/aliases.json

@@ -422,6 +422,15 @@
     "openai": "gpt-5.5-pro",
     "openrouter": "openai/gpt-5.5-pro"
   },
+  "gpt-5.6-luna": {
+    "bedrock": "openai.gpt-5.6-luna"
+  },
+  "gpt-5.6-sol": {
+    "bedrock": "openai.gpt-5.6-sol"
+  },
+  "gpt-5.6-terra": {
+    "bedrock": "openai.gpt-5.6-terra"
+  },
   "gpt-audio": {
     "openai": "gpt-audio",
     "openrouter": "openai/gpt-audio"

lib/ruby_llm/connection.rb

@@ -24,11 +24,11 @@ def self.basic(&)
       end
     end
 
-    def initialize(provider, config)
+    def initialize(provider, config, base_url: provider.api_base)
       @provider = provider
       @config = config
 
-      @connection = Faraday.new(provider.api_base) do |faraday|
+      @connection = Faraday.new(base_url) do |faraday|
         setup_timeout(faraday)
         setup_logging(faraday)
         setup_retry(faraday)

lib/ruby_llm/models.json

@@ -16784,6 +16784,204 @@
       "knowledge": "2025-12-01"
     }
   },
+  {
+    "id": "openai.gpt-5.6-luna",
+    "name": "GPT-5.6 Luna",
+    "provider": "bedrock",
+    "family": "gpt",
+    "created_at": "2026-07-13 00:00:00 UTC",
+    "context_window": 272000,
+    "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": 1.0,
+          "output_per_million": 6.0,
+          "cache_read_input_per_million": 0.1
+        }
+      }
+    },
+    "metadata": {
+      "source": "manual",
+      "provider_id": "amazon-bedrock",
+      "open_weights": false,
+      "attachment": true,
+      "temperature": false,
+      "last_updated": "2026-07-13",
+      "reasoning_options": [
+        {
+          "type": "effort",
+          "values": [
+            "none",
+            "low",
+            "medium",
+            "high",
+            "xhigh",
+            "max"
+          ]
+        }
+      ],
+      "cost": {
+        "input": 1.0,
+        "output": 6.0,
+        "cache_read": 0.1
+      },
+      "limit": {
+        "context": 272000,
+        "output": 128000
+      },
+      "knowledge": null
+    }
+  },
+  {
+    "id": "openai.gpt-5.6-sol",
+    "name": "GPT-5.6 Sol",
+    "provider": "bedrock",
+    "family": "gpt",
+    "created_at": "2026-07-13 00:00:00 UTC",
+    "context_window": 272000,
+    "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": 5.0,
+          "output_per_million": 30.0,
+          "cache_read_input_per_million": 0.5
+        }
+      }
+    },
+    "metadata": {
+      "source": "manual",
+      "provider_id": "amazon-bedrock",
+      "open_weights": false,
+      "attachment": true,
+      "temperature": false,
+      "last_updated": "2026-07-13",
+      "reasoning_options": [
+        {
+          "type": "effort",
+          "values": [
+            "none",
+            "low",
+            "medium",
+            "high",
+            "xhigh",
+            "max"
+          ]
+        }
+      ],
+      "cost": {
+        "input": 5.0,
+        "output": 30.0,
+        "cache_read": 0.5
+      },
+      "limit": {
+        "context": 272000,
+        "output": 128000
+      },
+      "knowledge": null
+    }
+  },
+  {
+    "id": "openai.gpt-5.6-terra",
+    "name": "GPT-5.6 Terra",
+    "provider": "bedrock",
+    "family": "gpt",
+    "created_at": "2026-07-13 00:00:00 UTC",
+    "context_window": 272000,
+    "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": 2.5,
+          "output_per_million": 15.0,
+          "cache_read_input_per_million": 0.25
+        }
+      }
+    },
+    "metadata": {
+      "source": "manual",
+      "provider_id": "amazon-bedrock",
+      "open_weights": false,
+      "attachment": true,
+      "temperature": false,
+      "last_updated": "2026-07-13",
+      "reasoning_options": [
+        {
+          "type": "effort",
+          "values": [
+            "none",
+            "low",
+            "medium",
+            "high",
+            "xhigh",
+            "max"
+          ]
+        }
+      ],
+      "cost": {
+        "input": 2.5,
+        "output": 15.0,
+        "cache_read": 0.25
+      },
+      "limit": {
+        "context": 272000,
+        "output": 128000
+      },
+      "knowledge": null
+    }
+  },
   {
     "id": "openai.gpt-oss-120b",
     "name": "gpt-oss-120b",

lib/ruby_llm/protocols/mantle_responses.rb

@@ -0,0 +1,55 @@
+# frozen_string_literal: true
+
+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 or InvokeModel (the GPT-5.x frontier
+    # family). 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.
+      #
+      # Kept in sync with Providers::Bedrock::MANTLE_ONLY_MODEL_PATTERN — that pattern picks
+      # mantle vs Converse/InvokeModel, 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/
+
+      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.
+      # 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'
+      end
+
+      private
+
+      def sync_response(payload, additional_headers = {})
+        super(payload, additional_headers.merge(mantle_signature_headers(payload)))
+      end
+
+      def stream_response(payload, additional_headers = {}, &)
+        super(payload, additional_headers.merge(mantle_signature_headers(payload)), &)
+      end
+
+      def mantle_signature_headers(payload)
+        body = JSON.generate(payload)
+        @provider.sign_headers(
+          'POST', completion_url, body,
+          base_url: @provider.mantle_api_base,
+          service: Providers::Bedrock::MANTLE_SIGNING_SERVICE,
+          region: @provider.mantle_region
+        )
+      end
+    end
+  end
+end

lib/ruby_llm/protocols/responses/chat.rb

@@ -82,7 +82,7 @@ def parse_output_citations(output, content)
         end
 
         def reasoning_model?(model_id)
-          model_id.match?(/^o\d|^gpt-5/)
+          model_id.match?(/\A(?:openai\.)?(?:o\d|gpt-5)/)
         end
 
         def parse_usage(usage)