legion-llm 0.8.26 → 0.8.27

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: 942f34663b8d915ee982996b5b2e63e26a7edf79a7aac17f8ce71ed1829dff01
-  data.tar.gz: dd78dd3bd79c9f1cf19d170f4ee2905fc92865dd3e21b107856c973eaf752fb5
+  metadata.gz: 2cccf9351fd9f4db59b1548197bf7b78c5947e85183535e86ede5c3359d71b89
+  data.tar.gz: 14e3a1b5b6648bea618941f84473e63aeccee7edc9520366d09dde8d27b00a7b
 SHA512:
-  metadata.gz: bfc1f55dce2a3eda78b5b6ab2405b6ce5d4e58fa841a81bb304af3bbe9a5b52851023c845d898713cfa87d9e292cd5fd1545464a7e0937eadde6f8668595ccc2
-  data.tar.gz: 4cad8eb9c6b6cfc79c1ffce687b7fddbb7b47d4e22ec9bca424f2dbb061ed83fff97d4ab2bbec441d3b319922316a238b057752210f1d7908e0d7169380485e9
+  metadata.gz: 31ec279fcb498e5cc3308bcefcb6adc94915c36867967fd08aa3f0422d4c583f83bc0ace1db9a47ecd45371a0ce0c82542fea7015c1474f5b7386409d789e5e0
+  data.tar.gz: '083bd8e581399a574b313a31784eacca4424fadb131f82cd17a5a7e840420da14ec3e5f589efa88b7be6be8a76916439c88bbf936a6d13c9a9435bd8fd04245c'

data/CHANGELOG.md

@@ -1,5 +1,16 @@
 # Legion LLM Changelog
 
+## [0.8.27] - 2026-04-24
+
+### Fixed
+- vLLM provider sent `developer` message role (OpenAI convention) which Qwen's chat template rejects. Added `Vllm::Chat` module that overrides `format_messages` and `format_role` to always send `system`.
+- vLLM provider called `OpenAI::Chat.render_payload` as a module function without provider instance context, causing `NoMethodError` on `openai_use_system_role`. Rewrote to use `super` with instance method overrides.
+- Audit events included the full conversation history in every message — quadratic payload growth. Now caps at the last 20 messages (configurable via `compliance.audit_max_messages`). Full conversation reconstructable via `conversation_id`.
+
+### Added
+- vLLM `chat_template_kwargs` with `enable_thinking` sent on every request so vLLM separates reasoning into the `reasoning` response field instead of inline `<think>` tags.
+- `providers.vllm.enable_thinking` setting (default: `true`). Controls whether thinking is enabled for vLLM requests. Per-request `thinking` param overrides.
+
 ## [0.8.26] - 2026-04-24
 
 ### Added

data/lib/legion/llm/inference/audit_publisher.rb

@@ -40,7 +40,7 @@ module Legion
             timeline:         compact_timeline(response.timeline),
             classification:   response.classification,
             tracing:          response.tracing,
-            messages:         request.messages,
+            messages:         current_turn_messages(request.messages),
             response_content: msg_content,
             tools_used:       tools_data,
             timestamp:        Time.now,
@@ -109,6 +109,23 @@ module Legion
           end
         end
 
+        def current_turn_messages(messages)
+          return messages unless messages.is_a?(Array)
+
+          max = audit_max_messages
+          return messages if messages.size <= max
+
+          messages.last(max)
+        end
+
+        def audit_max_messages
+          return 20 unless defined?(Legion::Settings)
+
+          Legion::Settings[:llm].dig(:compliance, :audit_max_messages) || 20
+        rescue StandardError
+          20
+        end
+
         def build_message_context(response:, **)
           {
             request_id:      response.request_id,

data/lib/legion/llm/patches/ruby_llm_vllm.rb

@@ -3,6 +3,47 @@
 module RubyLLM
   module Providers
     class Vllm < OpenAI
+      module Chat
+        def format_role(role)
+          role.to_s
+        end
+
+        def format_messages(messages)
+          messages.map do |msg|
+            {
+              role:         format_role(msg.role),
+              content:      OpenAI::Media.format_content(msg.content),
+              tool_calls:   format_tool_calls(msg.tool_calls),
+              tool_call_id: msg.tool_call_id
+            }.compact.merge(OpenAI::Chat.format_thinking(msg))
+          end
+        end
+
+        def render_payload(messages, tools:, temperature:, model:, stream: false, schema: nil,
+                           thinking: nil, tool_prefs: nil)
+          payload = super
+          enable = if thinking.nil?
+                     vllm_thinking_default
+                   else
+                     thinking ? true : false
+                   end
+          payload[:chat_template_kwargs] = { enable_thinking: enable }
+          payload
+        end
+
+        private
+
+        def vllm_thinking_default
+          return true unless defined?(Legion::Settings)
+
+          Legion::Settings[:llm].dig(:providers, :vllm, :enable_thinking) != false
+        rescue StandardError
+          true
+        end
+      end
+
+      include Vllm::Chat
+
       def api_base
         @config.vllm_api_base
       end

data/lib/legion/llm/settings.rb

@@ -377,10 +377,11 @@ module Legion
             base_url:      'http://localhost:11434'
           },
           vllm:      {
-            enabled:       false,
-            default_model: 'qwen3.6-27b',
-            base_url:      'http://localhost:8000/v1',
-            api_key:       nil
+            enabled:         false,
+            default_model:   'qwen3.6-27b',
+            base_url:        'http://localhost:8000/v1',
+            api_key:         nil,
+            enable_thinking: true
           }
         }
       end

data/lib/legion/llm/version.rb

@@ -2,6 +2,6 @@
 
 module Legion
   module LLM
-    VERSION = '0.8.26'
+    VERSION = '0.8.27'
   end
 end

metadata

@@ -1,7 +1,7 @@
 --- !ruby/object:Gem::Specification
 name: legion-llm
 version: !ruby/object:Gem::Version
-  version: 0.8.26
+  version: 0.8.27
 platform: ruby
 authors:
 - Esity