lex-llm-bedrock 0.3.9 → 0.3.11

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: 1d0b15f1168f45e6f11211963cc8c0939085562bd92e720bcd74426367824318
-  data.tar.gz: e68d8865e321e62f0b7c54cf16a573286174626531229b325e53b0aecbc8c3ea
+  metadata.gz: 52fa9adf70213041c0b2e6e96b44ff50b64220ae09f9cdab5619808f725df914
+  data.tar.gz: 12b5b4b31b712be0f0831e5e37dd045e6ca89442296bc645100c05cb41f4f8e0
 SHA512:
-  metadata.gz: ac5ed2ff6e4d586891edc3d05a8935cd32b25b561862a882c04a17d2c42324f8cfde527338c8d464cc8373d06dfa4b758fc3a7ab4c1ac70072d22d9998799207
-  data.tar.gz: 1e6f4bd752aa5fe9eeb033426405718826135763e2f6974674a7238ffbf74134024ad8f164962d02dc908aa8276417451087c7092e623beda5e0df58977efbf4
+  metadata.gz: 3519d77bffac0e5ea90805fe64cc095f831e6069bc21bf93ffa3e0441977560ebd556b8d2c82b5fad163301c3a78c54e5efdf6bc8060f0ad4935cdce07ab2cfd
+  data.tar.gz: 96f1a2d3d394e17b0a517a3523445d027b9a11819785c889b2aae483bfd97753d8b7926aeadaab827330831525a3c86e3b4a3e2d64462b6d35113d359431e7fa

data/CHANGELOG.md

@@ -1,5 +1,26 @@
 # Changelog
 
+## 0.3.11 - 2026-05-31
+
+### Security
+- **BEDROCK-CRED-01**: Static AWS credentials now emit a deprecation warning. New setting `security.block_static_aws_credentials=true` rejects them entirely, forcing IAM role-based authentication.
+
+### Fixed
+- **TRANSLATION-BUG-07**: Bedrock streaming now preserves thinking (chain-of-thought) blocks in the final `Message`. Previously CoT was accumulated by the wire handler but silently dropped from the returned response.
+
+### Added
+- **PROMPT-CACHE-01**: System blocks, tool definitions, and early conversation messages (first 4, never the last) now include `cache_control: { type: "cache_control" }` markers for Anthropic prompt caching via Bedrock Converse.
+- **PROMPT-CACHE-02**: Response parser extracts `cached_input_tokens` (`cache_read_input_tokens`) and `cache_creation_tokens` (`cache_creation_input_tokens`) from Bedrock usage metadata into `Message#cached_tokens` and `Message#cache_creation_tokens`.
+
+## 0.3.10 - 2026-05-21
+
+- Add `default_transport`/`default_tier` class declarations, remove `configured_transport`/`configured_tier`
+- Add `model_allowed?` filtering in `discover_offerings` (handles ModelOffering objects)
+- Move `DEFAULT_REGION` to settings[:region]
+- Default tier corrected from :frontier to :cloud
+- Identity headers included via base provider
+
+
 ## 0.3.9 - 2026-05-18
 
 - Fix streaming tool call parsing: `stream_converse` now handles content_block_start/delta/stop events for tool_use blocks, capturing tool ids, names, and accumulated input JSON. Previously only text deltas were captured and tool calls were silently dropped.

data/lib/legion/extensions/llm/bedrock/provider.rb

@@ -14,8 +14,6 @@ module Legion
         class Provider < Legion::Extensions::Llm::Provider # rubocop:disable Metrics/ClassLength
           include Legion::Logging::Helper
 
-          DEFAULT_REGION = 'us-east-1'
-
           STATIC_MODELS = [
             { model: 'anthropic.claude-3-haiku-20240307-v1:0', alias: 'claude-3-haiku' },
             { model: 'amazon.titan-text-express-v1', alias: 'titan-text-express' },
@@ -50,6 +48,8 @@ module Legion
 
           class << self
             def slug = 'bedrock'
+            def default_transport = :aws_sdk
+            def default_tier = :cloud
 
             def configuration_options
               %i[
@@ -77,11 +77,24 @@ module Legion
 
             INFERENCE_PROFILE_PREFIXES = %w[anthropic. meta. mistral. cohere. ai21.].freeze
 
-            def inference_profile_id(model)
+            def inference_profile_id(model, region: nil)
               return model if model.start_with?('us.', 'eu.', 'ap.', 'arn:')
               return model unless INFERENCE_PROFILE_PREFIXES.any? { |p| model.start_with?(p) }
 
-              "us.#{model}"
+              prefix = region ? region_prefix(region) : 'us'
+              "#{prefix}.#{model}"
+            end
+
+            # Region-based inference profile prefix mapping.
+            # Bare model IDs (e.g. anthropic.claude-sonnet-4) get the region prefix.
+            REGION_PREFIX = {
+              'us-east-1' => 'us', 'us-east-2' => 'us', 'us-west-1' => 'us', 'us-west-2' => 'us',
+              'eu-central-1' => 'eu', 'eu-west-1' => 'eu', 'eu-west-2' => 'eu', 'eu-west-3' => 'eu',
+              'ap-south-1' => 'ap', 'ap-southeast-1' => 'ap', 'ap-southeast-2' => 'ap', 'ap-northeast-1' => 'ap'
+            }.freeze
+
+            def region_prefix(region)
+              REGION_PREFIX.fetch(region.to_s, 'us')
             end
           end
 
@@ -113,7 +126,7 @@ module Legion
           def count_tokens_url = 'CountTokens'
 
           def region
-            config.bedrock_region || DEFAULT_REGION
+            config.bedrock_region || settings[:region] || 'us-east-1'
           end
 
           def discover_offerings(live: false, **filters)
@@ -126,8 +139,12 @@ module Legion
 
             log.info { "bedrock.provider.discover_offerings: listing foundation models (region=#{region})" }
             response = bedrock_client.list_foundation_models(**filters)
-            @cached_offerings = Array(value(response, :model_summaries)).map do |summary|
-              offering_from_summary(summary)
+            @cached_offerings = Array(value(response, :model_summaries)).filter_map do |summary|
+              offering = offering_from_summary(summary)
+              model_id = offering.respond_to?(:model) ? offering.model : (offering[:model] || offering[:id])
+              next unless model_allowed?(model_id.to_s)
+
+              offering
             end
             log.info { "bedrock.provider.discover_offerings: found #{@cached_offerings.size} models" }
             @cached_offerings
@@ -230,7 +247,7 @@ module Legion
             log.debug { "bedrock.provider.count_tokens: model=#{model_id(model)}" }
             request = Utils.deep_merge(
               {
-                model_id: self.class.inference_profile_id(model_id(model)),
+                model_id: self.class.inference_profile_id(model_id(model), region: region),
                 input: { converse: { messages: format_messages(messages), system: system_blocks(system) }.compact }
               },
               params
@@ -323,8 +340,8 @@ module Legion
             Legion::Extensions::Llm::Routing::ModelOffering.new(
               provider_family: :bedrock,
               instance_id: instance_id,
-              transport: configured_transport(:aws_sdk),
-              tier: configured_tier(:frontier),
+              transport: offering_transport,
+              tier: offering_tier,
               model: model,
               usage_type: usage_type,
               capabilities: capabilities || default_capabilities(model),
@@ -346,30 +363,48 @@ module Legion
             ctx ? { context_window: ctx } : nil
           end
 
-          def configured_transport(default)
-            config.respond_to?(:transport) ? config.transport : default
-          end
-
-          def configured_tier(default)
-            config.respond_to?(:tier) ? config.tier : default
-          end
-
-          def converse_request(messages, model:, temperature:, max_tokens:, tools:, tool_prefs:)
+          def converse_request(messages, model:, temperature:, max_tokens:, tools:, tool_prefs:, guardrail_config: nil)
             {
-              model_id: self.class.inference_profile_id(model_id(model)),
+              model_id: self.class.inference_profile_id(model_id(model), region: region),
               messages: format_messages(messages.reject { |message| message.role == :system }),
               system: format_system(messages),
               inference_config: { temperature: temperature, max_tokens: max_tokens || model_max_tokens(model) }.compact,
-              tool_config: format_tool_config(tools, tool_prefs)
+              tool_config: format_tool_config(tools, tool_prefs),
+              guardrail_config: guardrail_config
             }.compact
           end
 
           def format_messages(messages)
-            messages.filter_map do |message|
-              blocks = content_blocks(message.content)
+            total = messages.size
+            messages.filter_map.with_index do |message, idx|
+              blocks = message.role == :tool ? tool_result_blocks(message) : content_blocks(message.content)
               next if blocks.empty?
 
-              { role: bedrock_role(message.role), content: blocks }
+              cache_blocks = should_cache_message?(idx, total) ? add_cache_control_to_blocks(blocks) : blocks
+              { role: bedrock_role(message.role), content: cache_blocks }
+            end
+          end
+
+          def tool_result_blocks(message)
+            return [] unless message.tool_result?
+
+            [{
+              type: 'tool_result',
+              tool_use: { tool_use_id: message.tool_call_id },
+              content: [{ type: 'text', text: message.tool_results.to_s }]
+            }]
+          end
+
+          def should_cache_message?(index, total)
+            # Cache first 4 messages, never the last message
+            return false if index == total - 1
+
+            index < 4
+          end
+
+          def add_cache_control_to_blocks(blocks)
+            blocks.map do |block|
+              block.dup.merge(cache_control: { type: 'cache_control' })
             end
           end
 
@@ -382,7 +417,7 @@ module Legion
           def system_blocks(system)
             return nil if system.to_s.empty?
 
-            [{ text: system }]
+            [{ text: system, cache_control: { type: 'cache_control' } }]
           end
 
           def bedrock_role(role)
@@ -393,12 +428,45 @@ module Legion
             raw = raw_content(content)
             return raw if raw
 
+            return image_blocks(content) if content.respond_to?(:attachments) && !content.attachments.empty?
+
             text = content_text(content)
             return [] if text.strip.empty?
 
             [{ text: text }]
           end
 
+          def image_blocks(content)
+            blocks = []
+            text = content_text(content)
+            blocks << { text: text } if text.strip.present?
+
+            content.attachments.each do |attachment|
+              if attachment.is_a?(Legion::Extensions::Llm::Content::ImageAttachment)
+                blocks << format_image_attachment(attachment)
+              end
+            end
+            blocks
+          end
+
+          def format_image_attachment(attachment)
+            {
+              image: {
+                format: image_format(attachment.format),
+                source: { bytes: attachment.data }
+              }
+            }
+          end
+
+          def image_format(fmt)
+            case fmt.to_s.downcase
+            when 'jpeg', 'jpg' then 'jpeg'
+            when 'png' then 'png'
+            when 'gif' then 'gif'
+            when 'webp' then 'webp'
+            end || 'jpeg'
+          end
+
           def raw_content(content)
             return nil unless content.is_a?(Legion::Extensions::Llm::Content::Raw)
 
@@ -418,7 +486,14 @@ module Legion
               "bedrock.provider.tools: formatting tools=#{tools.keys.map(&:to_s).sort.join(',')} " \
                 "tool_choice=#{tool_choice_label(tool_prefs)}"
             end
-            { tools: tools.values.map { |tool| tool_definition(tool) }, tool_choice: tool_choice(tool_prefs) }.compact
+            {
+              tools: tools.values.map { |tool| tool_definition_with_cache(tool) },
+              tool_choice: tool_choice(tool_prefs)
+            }.compact
+          end
+
+          def tool_definition_with_cache(tool)
+            tool_definition(tool).merge(cache_control: { type: 'cache_control' })
           end
 
           def tool_definition(tool)
@@ -469,26 +544,33 @@ module Legion
               tool_calls: parse_tool_calls(value(message, :content)),
               input_tokens: value(usage, :input_tokens),
               output_tokens: value(usage, :output_tokens),
+              cached_tokens: cache_read_tokens(usage),
+              cache_creation_tokens: cache_write_tokens(usage),
               raw: normalize_response(response)
             )
           end
 
           def stream_converse(request, fallback_model)
-            state = { accumulated: +'', final_usage: nil, stop_reason: nil, tool_use_blocks: [], current_tool_use: nil }
+            state = { accumulated: +'', thinking: +'', final_usage: nil, stop_reason: nil,
+                      tool_use_blocks: [], current_tool_use: nil, in_thinking: false }
 
             runtime_client.converse_stream(**request) do |stream|
               wire_stream_handlers(stream, state, fallback_model) { |chunk| yield chunk if block_given? }
             end
 
-            Legion::Extensions::Llm::Message.new(
+            msg_attrs = {
               role: :assistant,
               content: state[:accumulated],
               model_id: fallback_model,
               tool_calls: build_stream_tool_calls(state[:tool_use_blocks]),
               input_tokens: value(state[:final_usage], :input_tokens),
               output_tokens: value(state[:final_usage], :output_tokens),
+              cached_tokens: cache_read_tokens(state[:final_usage]),
+              cache_creation_tokens: cache_write_tokens(state[:final_usage]),
               stop_reason: state[:stop_reason]
-            )
+            }
+            msg_attrs[:thinking] = state[:thinking] unless state[:thinking].empty?
+            Legion::Extensions::Llm::Message.new(**msg_attrs)
           end
 
           def wire_stream_handlers(stream, state, fallback_model, &)
@@ -504,6 +586,13 @@ module Legion
 
             stream.on_content_block_start_event do |event|
               start = value(event, :start)
+
+              if value(start, :thinking)
+                state[:in_thinking] = true
+                next
+              end
+
+              state[:in_thinking] = false
               tool_start = value(start, :tool_use) if start
               next unless tool_start
 
@@ -520,10 +609,14 @@ module Legion
               delta = value(event, :delta)
               text = value(delta, :text)
               if text
-                state[:accumulated] << text
-                if block_given?
-                  yield Legion::Extensions::Llm::Chunk.new(role: :assistant, content: text,
-                                                           model_id: fallback_model)
+                if state[:in_thinking]
+                  state[:thinking] << text
+                else
+                  state[:accumulated] << text
+                  if block_given?
+                    yield Legion::Extensions::Llm::Chunk.new(role: :assistant, content: text,
+                                                             model_id: fallback_model)
+                  end
                 end
               end
 
@@ -569,6 +662,18 @@ module Legion
             end
           end
 
+          def cache_read_tokens(usage)
+            return nil if usage.nil?
+
+            value(usage, :cache_read_input_tokens) || value(usage, 'cache_read_input_tokens')
+          end
+
+          def cache_write_tokens(usage)
+            return nil if usage.nil?
+
+            value(usage, :cache_creation_input_tokens) || value(usage, 'cache_creation_input_tokens')
+          end
+
           def parse_embedding_response(response, model:)
             body = parse_body(value(response, :body))
             vectors = body['embedding'] || body['embeddings'] || body.dig('data', 0, 'embedding')
@@ -595,11 +700,11 @@ module Legion
           end
 
           def bedrock_client
-            Aws::Bedrock::Client.new(client_options)
+            @bedrock_client ||= Aws::Bedrock::Client.new(client_options)
           end
 
           def runtime_client
-            Aws::BedrockRuntime::Client.new(client_options)
+            @runtime_client ||= Aws::BedrockRuntime::Client.new(client_options)
           end
 
           def client_options
@@ -626,10 +731,21 @@ module Legion
             return Aws::SharedCredentials.new(profile_name: config.bedrock_profile) if config.bedrock_profile
             return nil unless config.bedrock_access_key_id
 
+            if static_credentials_blocked?
+              raise SecurityError,
+                    'Static AWS credentials are disabled (security.block_static_aws_credentials=true); use IAM roles'
+            end
+            log.warn('[bedrock] Using static AWS credentials — prefer IAM roles for production')
             Aws::Credentials.new(config.bedrock_access_key_id, config.bedrock_secret_access_key,
                                  config.bedrock_session_token)
           end
 
+          def static_credentials_blocked?
+            return false unless defined?(::Legion::Settings)
+
+            ::Legion::Settings.dig(:extensions, :llm, :security, :block_static_aws_credentials) == true
+          end
+
           def credential_source
             return :static if config.bedrock_access_key_id
             return :profile if config.bedrock_profile

data/lib/legion/extensions/llm/bedrock/version.rb

@@ -4,7 +4,7 @@ module Legion
   module Extensions
     module Llm
       module Bedrock
-        VERSION = '0.3.9'
+        VERSION = '0.3.11'
       end
     end
   end

data/lib/legion/extensions/llm/bedrock.rb

@@ -22,8 +22,9 @@ module Legion
           ::Legion::Extensions::Llm.provider_settings(
             family: PROVIDER_FAMILY,
             instance: {
-              default_model: 'us.anthropic.claude-sonnet-4-6',
-              tier: :frontier,
+              default_model: 'anthropic.claude-sonnet-4',
+              region: 'us-east-1',
+              tier: :cloud,
               transport: :aws_sdk,
               credentials: {
                 bearer_token: nil,
@@ -73,7 +74,7 @@ module Legion
                            .transform_values do |config|
             sanitized = sanitize_instance_config(config)
             sanitized[:capabilities] ||= DEFAULT_CAPABILITIES.dup
-            sanitized[:default_model] ||= 'us.anthropic.claude-sonnet-4-6'
+            sanitized[:default_model] ||= 'anthropic.claude-sonnet-4'
             sanitized
           end
         end
@@ -197,6 +198,8 @@ module Legion
         end
 
         def self.normalize_instance_config(config)
+          return {} if config.nil?
+
           normalized = config.to_h.transform_keys { |key| key.respond_to?(:to_sym) ? key.to_sym : key }
           normalized[:bedrock_region] ||= normalized.delete(:region)
           normalized[:bedrock_endpoint] ||= normalized.delete(:endpoint)

metadata

@@ -1,7 +1,7 @@
 --- !ruby/object:Gem::Specification
 name: lex-llm-bedrock
 version: !ruby/object:Gem::Version
-  version: 0.3.9
+  version: 0.3.11
 platform: ruby
 authors:
 - LegionIO