lex-llm 0.7.6 → 0.8.4

data/lib/legion/extensions/llm/inventory/weight_schema.rb

@@ -2,6 +2,7 @@
 
 require 'legion/extensions/llm/settings_cascade'
 require 'legion/extensions/llm/inventory/identity'
+require 'legion/extensions/llm/taxonomies'
 
 module Legion
   module Extensions
@@ -11,15 +12,17 @@ module Legion
         # reads the stored scalar; it is computed ONLY here, at write events, from a
         # live settings read. Each multiplicative axis reads ITS OWN scope — never a
         # fall-through cascade (a fall-through would double-count the other axes).
+        # The offering-scope settings key is the lane's 5 tuple — operator-readable,
+        # composed via Identity.compose_lane_id.
         module WeightSchema
           module_function
 
           IDENTITY = 100
 
-          # Exact 4-key hash. offering_id derives from instance_key + the REAL
-          # provider native key (NOT the model — azure's native key is
-          # deployment_name, which can differ from model_id).
-          def weight_inputs(settings:, instance_key:, provider_native_key:, model:, tier:)
+          # Exact 4-key hash. The offering-scope settings key is the 5 tuple of
+          # the lanes the draft's operation evidence supports (one per distinct
+          # lane type) — finally operator-usable, no digest.
+          def weight_inputs(settings:, instance_key:, model:, tier:, operation_evidence:)
             llm_conf = scope_hash(settings.dig(:extensions, :llm), path: 'extensions.llm')
             provider_conf = scope_hash(
               SettingsCascade.lookup(llm_conf, instance_key.provider_family),
@@ -35,10 +38,10 @@ module Legion
             tier_weights = scope_hash(
               settings.dig(:llm, :routing, :tier_weights), path: 'llm.routing.tier_weights'
             )
-            offering_id = Identity.offering_id(instance_key: instance_key, provider_native_key: provider_native_key)
             offerings = scope_hash(SettingsCascade.lookup(provider_conf, :offerings), path: 'provider.offerings')
-            offering_conf = scope_hash(
-              SettingsCascade.lookup(offerings, offering_id), path: "provider.offerings.#{offering_id}"
+            offering_w = offering_scope_weight(
+              offerings: offerings, instance_key: instance_key, model: model, tier: tier,
+              operation_evidence: operation_evidence
             )
             provider_models = scope_hash(
               SettingsCascade.lookup(provider_conf, :models), path: 'provider.models'
@@ -59,7 +62,6 @@ module Legion
               path: "merged model scope #{model}"
             )
 
-            offering_w = SettingsCascade.lookup(offering_conf, :weight)
             model_w = SettingsCascade.lookup(model_scopes, :weight)
             {
               tier: component(SettingsCascade.lookup(tier_weights, tier), IDENTITY),
@@ -72,6 +74,43 @@ module Legion
             }
           end
 
+          # The configured offering-scope weight for the draft: the operator keys
+          # `extensions.llm.<provider>.offerings` by the lane's 5 tuple. The lanes
+          # a draft supports are one 5 tuple per distinct lane type; the draft
+          # carries a SINGLE weight pair, so configured weights across its lanes
+          # must agree — a disagreement is an ambiguous operator input and RAISES
+          # instead of silently picking one.
+          def offering_scope_weight(offerings:, instance_key:, model:, tier:, operation_evidence:)
+            lane_ids = supported_lane_ids(
+              instance_key: instance_key, model: model, tier: tier, operation_evidence: operation_evidence
+            )
+            weights = lane_ids.filter_map do |lane_id|
+              scope = scope_hash(SettingsCascade.lookup(offerings, lane_id), path: "provider.offerings.#{lane_id}")
+              SettingsCascade.lookup(scope, :weight)
+            end
+            return nil if weights.empty?
+
+            unless weights.uniq.length == 1
+              raise ArgumentError, "offering weight scope is ambiguous: the draft's lanes " \
+                                   "#{lane_ids.join(', ')} carry differing configured weights"
+            end
+
+            weights.first
+          end
+
+          def supported_lane_ids(instance_key:, model:, tier:, operation_evidence:)
+            operation_evidence.each_value.with_object([]) do |evidence, lane_ids|
+              next unless evidence.supported?
+
+              lane_id = Identity.compose_lane_id(
+                tier: tier, provider_family: instance_key.provider_family,
+                instance_id: instance_key.instance_id,
+                type: Taxonomies.lane_type_for(operation: evidence.operation), model: model
+              )
+              lane_ids << lane_id unless lane_ids.include?(lane_id)
+            end
+          end
+
           # nil → default; an explicit 0 passes through (0 = operator disable —
           # 0 is TRUTHY in Ruby and must not be defaulted); any other non-Integer
           # RAISES instead of silently applying a different weight.

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

@@ -4,7 +4,12 @@ module Legion
   module Extensions
     module Llm
       class Provider
-        # Shared OpenAI-compatible HTTP payload and response adapter.
+        # Shared OpenAI-compatible HTTP payload and response adapter — the
+        # reference implementation for the OpenAI wire dialect (08 R3).
+        # Renders FROM Canonical (render_payload) and parses TO Canonical
+        # (parse_completion_response / build_chunk). Provider-dialect
+        # translation (usage spellings, JSON-string arguments, reasoning
+        # fields) lives here, at the edge (03 O03a).
         module OpenAICompatible
           def stream_usage_supported? = false
           def completion_url = '/v1/chat/completions'
@@ -20,68 +25,94 @@ module Legion
 
           private
 
-          def render_payload(messages, tools:, temperature:, model:, stream:, schema:, thinking:, tool_prefs:) # rubocop:disable Metrics/ParameterLists
+          def render_payload(messages, tools:, model:, stream:, schema:, thinking:, params:, tool_prefs:) # rubocop:disable Metrics/ParameterLists
             payload = {
-              model: model.id,
+              model: model,
               messages: format_openai_messages(messages),
-              temperature: temperature,
+              temperature: maybe_normalize_temperature(params),
               stream: stream,
               tools: format_openai_tools(tools),
               tool_choice: openai_tool_choice(tool_prefs),
               response_format: openai_response_format(schema),
               reasoning_effort: openai_reasoning_effort(thinking)
             }.compact
+            payload.merge!(openai_payload_params(params))
             payload[:stream_options] = { include_usage: true } if stream && stream_usage_supported?
             payload
           end
 
+          # Canonical Params → OpenAI wire keys (edge translation, O03a).
+          def openai_payload_params(params)
+            return {} unless params
+
+            {
+              max_tokens: params.max_tokens,
+              top_p: params.top_p,
+              top_k: params.top_k,
+              stop: params.stop_sequences,
+              seed: params.seed,
+              frequency_penalty: params.frequency_penalty,
+              presence_penalty: params.presence_penalty,
+              response_format: openai_response_format_value(params.response_format)
+            }.compact
+          end
+
+          # response_format travels in its provider wire shape (client
+          # translator concern): a wire Hash passes through, a mode String is
+          # wrapped.
+          def openai_response_format_value(response_format)
+            return nil if response_format.nil?
+            return response_format if response_format.is_a?(::Hash)
+
+            { type: response_format.to_s }
+          end
+
           def format_openai_messages(messages)
             messages.map do |message|
               {
                 role: message.role.to_s,
-                content: openai_content(message.content, role: message.role),
+                content: openai_content(message.content),
                 tool_call_id: message.tool_call_id,
                 tool_calls: format_openai_tool_calls(message.tool_calls)
               }.compact
             end
           end
 
-          def openai_content(content, role:)
-            return content.format if content.is_a?(Legion::Extensions::Llm::Content::Raw)
-            return sanitize_openai_text(content, role:) unless content.respond_to?(:attachments)
-            return sanitize_openai_text(content.text.to_s, role:) if content.attachments.empty?
+          # L10: the role-parameterized sanitizer is deleted — text content
+          # passes through verbatim in every role (the decision comment
+          # below documents why).
+          def openai_content(content)
+            return content.map { |block| openai_content(block) } if content.is_a?(::Array)
+            return content.to_s if content.nil? || content.is_a?(::String)
 
-            openai_content_parts(content)
-          end
+            return content.text.to_s if content.text?
 
-          def openai_content_parts(content)
-            parts = []
-            parts << { type: 'text', text: content.text.to_s } if content.text
-            content.attachments.each do |attachment|
-              parts << { type: 'image_url', image_url: { url: attachment.for_llm } } if attachment.image?
-            end
-            parts
+            {
+              type: content.type.to_s,
+              text: content.text,
+              data: content.data,
+              media_type: content.media_type,
+              source_type: content.source_type,
+              name: content.name,
+              id: content.id,
+              input: content.input,
+              tool_use_id: content.tool_use_id,
+              is_error: content.is_error
+            }.compact
           end
 
-          def sanitize_openai_text(text, role:)
-            return text unless role.to_sym == :assistant && text.is_a?(String)
-
-            # Preserve thinking tags in the content — qwen3.6 outputs thinking in
-            # <think> tags and expects to see its own reasoning on subsequent rounds.
-            # The Anthropic API layer separates thinking into distinct content blocks
-            # for client-facing responses; the OpenAI compat layer passes them through.
-            text
-          end
+          # Thinking-tag pass-through decision (L10: the vestigial sanitizer
+          # is deleted, the decision stays): qwen3.6 outputs thinking in
+          # tags and expects to see its own reasoning on subsequent rounds.
+          # The Anthropic API layer separates thinking into distinct content
+          # blocks for client-facing responses; the OpenAI compat layer passes
+          # them through untouched.
 
           def format_openai_tool_calls(tool_calls)
             return nil unless tool_calls&.any?
 
-            # Array is the canonical shape (per canonical/message.rb); Hash
-            # is the legacy lex-llm shape (id => ToolCall). Both flow through
-            # this renderer depending on caller.
-            calls = tool_calls.is_a?(Hash) ? tool_calls.values : Array(tool_calls)
-
-            calls.map do |tool_call|
+            # Array<Canonical::ToolCall> only (the legacy Hash shape is deleted).
+            tool_calls.map do |tool_call|
               {
                 id: tool_call.id,
                 type: 'function',
@@ -94,7 +125,7 @@ module Legion
           end
 
           def format_openai_tools(tools)
-            return nil if tools.empty?
+            return nil if tools.nil? || tools.empty?
 
             tools.values.map do |tool|
               {
@@ -123,176 +154,128 @@ module Legion
             { type: 'json_schema', json_schema: schema_hash }
           end
 
+          # Transpose canonical thinking -> OpenAI `reasoning_effort` (edge, O03a).
+          # Omit the key when there is no thinking or it is explicitly disabled
+          # (a reasoning model then uses its own default). resolved_effort fills
+          # the effort axis from budget when the client supplied only a budget,
+          # so a budget-only request is never silently dropped.
           def openai_reasoning_effort(thinking)
-            return nil unless thinking.is_a?(Hash)
+            return nil unless thinking.is_a?(Canonical::Thinking::Config)
+            return nil unless thinking.enabled
 
-            thinking[:effort] || thinking['effort']
+            thinking.resolved_effort
           end
 
+          # One response-parse boundary (08 R2): returns Canonical::Response.
           def parse_completion_response(response)
             body = response.body
             choice = Array(body['choices']).first || {}
             message = choice['message'] || {}
-            usage = body['usage'] || {}
-            content, thinking = extract_thinking_from_completion(message)
-
-            Legion::Extensions::Llm::Message.new(
-              role: :assistant,
-              content: content,
-              model_id: body['model'],
-              tool_calls: parse_tool_calls(message['tool_calls']),
-              thinking: thinking,
-              input_tokens: usage['prompt_tokens'],
-              output_tokens: usage['completion_tokens'],
-              reasoning_tokens: usage.dig('completion_tokens_details', 'reasoning_tokens'),
-              raw: body
-            )
-          end
-
-          def extract_thinking_from_completion(message)
             extraction = Responses::ThinkingExtractor.extract(
               message['content'],
               metadata: thinking_metadata(message)
             )
 
-            [
-              extraction.content,
-              Thinking.build(
-                text: extraction.thinking,
-                signature: extraction.signature
-              )
-            ]
+            Canonical::Response.build(
+              text: extraction.content,
+              thinking: thinking_value(extraction),
+              tool_calls: parse_tool_calls(message['tool_calls']),
+              usage: openai_usage_to_canonical(body['usage']),
+              stop_reason: stop_reason_lookup(body.dig('choices', 0, 'finish_reason')),
+              model: body['model']
+            )
           end
 
-          def thinking_metadata(message)
-            {
-              reasoning_content: message['reasoning_content'],
-              reasoning: message['reasoning'],
-              thinking: message['thinking'],
-              thinking_text: message['thinking_text'],
-              thinking_signature: message['thinking_signature'],
-              reasoning_signature: message['reasoning_signature']
-            }.compact
+          # One thinking-metadata key list (10 U3): both the sync and chunk
+          # paths extract through the shared ThinkingExtractor core.
+          def thinking_metadata(wire_message)
+            Responses::ThinkingExtractor::THINKING_METADATA_KEYS.each_with_object({}) do |key, hash|
+              value = wire_message[key.to_s]
+              hash[key] = value unless value.nil?
+            end
           end
 
-          def build_chunk(data)
-            choice = Array(data['choices']).first || {}
-            delta = choice['delta'] || {}
-            usage = data['usage'] || {}
-            content, thinking = extract_thinking_from_chunk(delta)
-
-            Legion::Extensions::Llm::Chunk.new(
-              role: :assistant,
-              content: content,
-              model_id: data['model'],
-              tool_calls: parse_streaming_tool_calls(delta['tool_calls']),
-              thinking: thinking,
+          def thinking_value(extraction)
+            return nil if extraction.thinking.nil? && extraction.signature.nil?
+
+            Canonical::Thinking.build(content: extraction.thinking, signature: extraction.signature)
+          end
+
+          # OpenAI wire usage → canonical keys (edge translation, O03a).
+          def openai_usage_to_canonical(usage)
+            return nil if usage.to_h.empty?
+
+            Canonical::Usage.build(
               input_tokens: usage['prompt_tokens'],
               output_tokens: usage['completion_tokens'],
-              raw: data
+              cache_read_tokens: usage.dig('prompt_tokens_details', 'cached_tokens') || usage.dig('input_tokens_details', 'cached_tokens'),
+              thinking_tokens: usage.dig('completion_tokens_details', 'reasoning_tokens') || usage.dig('output_tokens_details', 'reasoning_tokens')
             )
           end
 
-          def parse_streaming_tool_calls(tool_calls)
-            return nil unless tool_calls&.any?
+          # One chunk-parse boundary (08 R2): returns a Canonical::Chunk or an
+          # Array of them. In-band think tags are NOT split here — tags split
+          # across chunks are the accumulator's stateful job (U1); per-chunk
+          # metadata (e.g. reasoning_content) is separated here, statelessly.
+          def build_chunk(data)
+            choice = Array(data['choices']).first || {}
+            delta = choice['delta'] || {}
+            usage = data['usage'] || {}
 
-            tool_calls.to_h do |call|
-              function = call.fetch('function', {})
-              name = function['name']
-              id = call['id']
-              key = (name || id || call['index']).to_s.to_sym
-              [
-                key,
-                Legion::Extensions::Llm::ToolCall.new(id: id, name: name, arguments: function['arguments'] || '')
-              ]
+            chunks = []
+            metadata = thinking_metadata(delta)
+            if metadata.any?
+              extraction = Responses::ThinkingExtractor.extract(nil, metadata:)
+              if extraction.thinking
+                chunks << Canonical::Chunk.thinking_delta(
+                  delta: extraction.thinking, request_id: nil, signature: extraction.signature
+                )
+              end
             end
-          end
-
-          def extract_thinking_from_chunk(delta)
-            reasoning = delta['reasoning_content'] || delta['reasoning']
             content = delta['content']
+            chunks << Canonical::Chunk.text_delta(delta: content.to_s, request_id: nil) if content
+            chunks.concat(parse_streaming_tool_calls(delta['tool_calls']))
+            canonical_usage = openai_usage_to_canonical(usage)
+            chunks << Canonical::Chunk.usage_chunk(usage: canonical_usage, request_id: nil) if canonical_usage
 
-            if reasoning
-              [content, Thinking.build(text: reasoning)]
+            if chunks.empty?
+              nil
             else
-              [content, nil]
+              (chunks.size == 1 ? chunks.first : chunks)
             end
           end
 
-          def parse_tool_calls(tool_calls)
-            return nil unless tool_calls&.any?
+          def parse_streaming_tool_calls(tool_calls)
+            return [] unless tool_calls&.any?
 
-            tool_calls.to_h do |call|
+            tool_calls.filter_map do |call|
               function = call.fetch('function', {})
               name = function['name']
-              id = call['id'] || name || call['index']
-              key = name || id
-              [
-                key.to_s.to_sym,
-                Legion::Extensions::Llm::ToolCall.new(
-                  id: id&.to_s,
-                  name: name,
-                  arguments: parse_tool_arguments(function['arguments'])
-                )
-              ]
-            end
-          end
-
-          def parse_tool_arguments(arguments)
-            return {} if arguments.nil? || arguments == ''
-            return arguments if arguments.is_a?(Hash)
-
-            Legion::JSON.parse(arguments, symbolize_names: false)
-          rescue Legion::JSON::ParseError => e
-            handle_exception(e, level: :warn, handled: true, operation: 'llm.provider.parse_tool_arguments')
-            {}
-          end
+              id = call['id']
+              index = call['index']
+              arguments_fragment = function['arguments'].to_s
+              next nil if id.nil? && name.nil? && index.nil? && arguments_fragment.empty?
 
-          def parse_list_models_response(response, provider, capabilities)
-            response.body.fetch('data', []).map do |model|
-              critical_capabilities = critical_capabilities_for(capabilities, model)
-              Legion::Extensions::Llm::Model::Info.from_hash(
-                id: model.fetch('id'),
-                name: model['id'],
-                provider: provider,
-                created_at: model_created_at(model['created']),
-                capabilities: critical_capabilities,
-                modalities: modalities_for_capabilities(critical_capabilities),
-                metadata: model
+              Canonical::Chunk.tool_call_delta(
+                tool_call: { id: id, name: name, arguments: arguments_fragment, index: index },
+                request_id: nil
               )
             end
           end
 
-          def model_created_at(value)
-            value.is_a?(Numeric) ? Time.at(value).utc : value
-          end
-
-          def critical_capabilities_for(capabilities, model)
-            return [] unless capabilities
-            return capabilities.critical_capabilities_for(model) if capabilities.respond_to?(:critical_capabilities_for)
-
-            {
-              'streaming' => :streaming?,
-              'function_calling' => :functions?,
-              'vision' => :vision?,
-              'embeddings' => :embeddings?,
-              'moderation' => :moderation?,
-              'image' => :images?,
-              'audio_transcription' => :audio_transcription?
-            }.filter_map do |capability, predicate|
-              capability if capabilities.respond_to?(predicate) && capabilities.public_send(predicate, model)
-            end
-          end
+          # Sync tool calls: the ONE strict arguments parser (10 U2).
+          def parse_tool_calls(tool_calls)
+            return [] unless tool_calls&.any?
 
-          def modalities_for_capabilities(capabilities)
-            if capabilities.include?('embeddings') && (capabilities - ['embeddings']).empty?
-              { input: %w[text], output: %w[embeddings] }
-            elsif capabilities.include?('image')
-              { input: %w[text image], output: %w[image] }
-            elsif capabilities.include?('audio_transcription')
-              { input: %w[audio], output: %w[text] }
-            else
-              { input: %w[text image], output: %w[text] }
+            tool_calls.map do |call|
+              function = call.fetch('function', {})
+              name = function['name']
+              id = call['id'] || name || call['index']
+              Canonical::ToolCall.build(
+                name: name.to_s,
+                id: id&.to_s,
+                arguments: Responses::ToolArguments.parse!(function['arguments'])
+              )
             end
           end
 
@@ -300,51 +283,64 @@ module Legion
             { model: model.respond_to?(:id) ? model.id : model, input: text, dimensions: dimensions }.compact
           end
 
+          # 05 §3 documented artifact: { text:, model:, embedding: Array<Float>,
+          # usage: Canonical::Usage }.
           def parse_embedding_response(response, model:, text:)
             vectors = response.body.fetch('data', []).map { |item| item['embedding'] }
             vectors = vectors.first unless text.is_a?(Array)
             usage = response.body['usage'] || {}
 
-            Legion::Extensions::Llm::Embedding.new(vectors: vectors, model: model,
-                                                   input_tokens: usage['prompt_tokens'].to_i)
+            {
+              text: text,
+              model: model.respond_to?(:id) ? model.id : model,
+              embedding: vectors,
+              usage: openai_usage_to_canonical(usage)
+            }.compact
           end
 
           def render_moderation_payload(input, model:)
-            { model: model, input: input }.compact
+            input_text = input.is_a?(::Array) ? input.map(&:text).join("\n") : input
+            { model: model, input: input_text }.compact
           end
 
+          # 05 §3 documented artifact: { model:, result: { flagged:, categories: } }.
           def parse_moderation_response(response, model:)
-            Legion::Extensions::Llm::Moderation.new(id: response.body['id'], model: response.body['model'] || model,
-                                                    results: response.body.fetch('results', []))
+            result = Array(response.body['results']).first || {}
+            {
+              model: response.body['model'] || model,
+              result: {
+                flagged: result['flagged'] == true,
+                categories: result['categories'] || {}
+              }
+            }
           end
 
           def render_image_payload(prompt, model:, size:, with:, mask:, params:) # rubocop:disable Metrics/ParameterLists
             { model: model, prompt: prompt, size: size, image: with, mask: mask }.merge(params).compact
           end
 
+          # 05 §3 documented artifact: { model:, image:, size: }.
           def parse_image_response(response, model:)
             image = response.body.fetch('data', []).first || {}
-            Legion::Extensions::Llm::Image.new(
-              url: image['url'],
-              data: image['b64_json'],
-              revised_prompt: image['revised_prompt'],
-              model_id: model,
-              usage: response.body['usage'] || {}
-            )
+            {
+              model: model,
+              image: image['url'] || image['b64_json'],
+              size: image['size']
+            }.compact
           end
 
           def render_transcription_payload(file_part, model:, language:, **options)
             { model: model, file: file_part, language: language }.merge(options).compact
           end
 
+          # 05 §3 documented artifact: { model:, text:, language:, duration: }.
           def parse_transcription_response(response, model:)
-            Legion::Extensions::Llm::Transcription.new(
-              text: response.body['text'],
+            {
               model: model,
+              text: response.body['text'],
               language: response.body['language'],
-              duration: response.body['duration'],
-              segments: response.body['segments']
-            )
+              duration: response.body['duration']
+            }.compact
           end
         end
       end