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-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": "models.dev",
+      "provider_id": "amazon-bedrock",
+      "open_weights": false,
+      "attachment": true,
+      "temperature": false,
+      "last_updated": "2026-06-01",
+      "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": "models.dev",
+      "provider_id": "amazon-bedrock",
+      "open_weights": false,
+      "attachment": true,
+      "temperature": false,
+      "last_updated": "2026-06-01",
+      "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-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": "models.dev",
+      "provider_id": "amazon-bedrock",
+      "open_weights": false,
+      "attachment": true,
+      "temperature": false,
+      "last_updated": "2026-06-01",
+      "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-oss-120b",
     "name": "gpt-oss-120b",

lib/ruby_llm/protocols/mantle_responses.rb

@@ -0,0 +1,46 @@
+# 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.
+      FRONTIER_GPT5_PATTERN = /\Aopenai\.gpt-5/
+
+      def initialize(provider, model = nil)
+        super
+        @connection = provider.mantle_connection
+      end
+
+      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
+        )
+      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)

lib/ruby_llm/providers/bedrock.rb

@@ -9,8 +9,20 @@ class Bedrock < Provider
 
       protocol :converse, Protocols::Converse, batches: Protocols::Converse::Batches
       protocol :bedrock_invoke_model, Protocols::BedrockInvokeModel
+      protocol :mantle_responses, Protocols::MantleResponses
       files Bedrock::Files
 
+      # SigV4 requests to bedrock-mantle sign against this service namespace, not "bedrock" —
+      # AWS models bedrock-mantle as a separate service (see the AmazonBedrockMantleFullAccess
+      # managed policy and the bedrock-mantle:CreateInference IAM action). Named here so it is a
+      # one-line change if this assumption turns out to be wrong.
+      MANTLE_SIGNING_SERVICE = 'bedrock-mantle'
+
+      # openai.gpt-5.x ids are only reachable on bedrock-mantle (no Converse, no InvokeModel).
+      # Deliberately narrower than /\Aopenai\./ — openai.gpt-oss-* models ARE served by
+      # bedrock-runtime and must keep routing to Converse.
+      MANTLE_ONLY_MODEL_PATTERN = /\Aopenai\.gpt-5/
+
       def api_base
         @config.bedrock_api_base || "https://bedrock-runtime.#{bedrock_region}.amazonaws.com"
       end
@@ -19,15 +31,26 @@ def control_api_base
         @config.bedrock_api_base || "https://bedrock.#{bedrock_region}.amazonaws.com"
       end
 
+      def mantle_api_base
+        @config.bedrock_mantle_api_base || "https://bedrock-mantle.#{mantle_region}.api.aws"
+      end
+
+      def mantle_connection
+        @mantle_connection ||= Connection.new(self, @config, base_url: mantle_api_base)
+      end
+
       def headers
         {}
       end
 
       def complete(messages, model:, params: {}, **rest, &)
-        super(messages, model:, params: normalize_params(params, model:), **rest, &)
+        params = normalize_params(params, model:) unless mantle_only_model?(model)
+        super
       end
 
       def protocol_for(model, **)
+        return fetch_protocol(:mantle_responses) if mantle_only_model?(model)
+
         invoke_model?(model) ? fetch_protocol(:bedrock_invoke_model) : fetch_protocol(:converse)
       end
 
@@ -37,7 +60,7 @@ def parse_error(response)
         body = try_parse_json(response.body)
         return body if body.is_a?(String)
 
-        body['message'] || body['Message'] || body['error'] || body['__type'] || super
+        extract_error_message(body) || super
       end
 
       def list_models
@@ -54,6 +77,8 @@ def configuration_options
             bedrock_session_token
             bedrock_credential_provider
             bedrock_api_base
+            bedrock_mantle_api_base
+            bedrock_mantle_region
             bedrock_batch_s3_uri
             bedrock_batch_role_arn
             bedrock_use_invoke_model
@@ -96,10 +121,26 @@ def ensure_configured!
 
       private
 
+      # Bedrock errors are shaped like {"message" => "..."} or {"__type" => "..."};
+      # mantle (OpenAI Responses) errors are shaped like {"error" => {"message" => "..."}}.
+      def extract_error_message(body)
+        body.dig('error', 'message') || body['message'] || body['Message'] || body['error'] || body['__type']
+      end
+
       def bedrock_region
         @config.bedrock_region
       end
 
+      def mantle_region
+        @config.bedrock_mantle_region || bedrock_region
+      end
+
+      # openai.gpt-5.x ids cannot serve Converse or InvokeModel; routing here is automatic and
+      # needs no config knob (unlike bedrock_use_invoke_model, which is an optimization choice).
+      def mantle_only_model?(model)
+        MANTLE_ONLY_MODEL_PATTERN.match?(model.id.to_s)
+      end
+
       def bedrock_credentials_requirement
         if @config.bedrock_credential_provider
           'bedrock_credential_provider responding to #credentials'