legion-llm 0.10.1 → 0.10.2

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: 94a90f69b3940eb805b6b7bf0b15eb30aa8c616e66562560df01d3f0d5b24ca2
-  data.tar.gz: 4a90c910e19cec4b3da0e5e44a00d1076495d41d653ad4f7161b3b38892f8d8b
+  metadata.gz: f5c1d01be312fb6e070b3e75236a8ecd6c6841653e42eae391cb5b2519a668b6
+  data.tar.gz: f6703f5c706dd134e85993c11da4222fc68edced3ca88e09a841f04ca2fb2474
 SHA512:
-  metadata.gz: e5f444b505e4b75937223475300cd4bae1e2f292f93148fbb3f0a06446ae6b90b0bf482a7d308a3e580e55219ddc4eb6f0e9274f6479c6e96619d54946cc0ef6
-  data.tar.gz: 6a9dfedae0e30d61f1bf8c4dc825764e4e9174829cc2d1ea6ee65a00ab7dea2cec5299e5988117aacd24adba93d956970124e00c11a1360bacfab04aa9cf6044
+  metadata.gz: 06aef6de56965e58876fa98bfcdf8b411996d00ea3faff44f9aa0f9035d385d9da3591300ac2663a31a2820cf56610572fff3b801ec417af20f20fd377ad0e8a
+  data.tar.gz: 77b720ac764fbb899dbbea7d372edfb7a6a8c2671d2b8eef4f92680439d4572e9e8986519c68d4452232ede7c7c87ae74b3b38fbb6383bd9e3dcc1aa0f30670a

data/CHANGELOG.md

@@ -1,5 +1,10 @@
 # Legion LLM Changelog
 
+## [0.10.2] - 2026-05-30
+
+### Fixed
+- **Registry tool injection disabled for client passthrough** — When `client_tool_passthrough` is enabled and the request has client-sourced tools (Claude Code, Codex, Kai), Legion no longer injects internal registry tools (`exec_filesystem_mv`, `legion-apollo-*`, etc.) into the payload. This fixes the first-turn garbage output where Qwen3 would output raw tool definition JSON instead of responding normally.
+- **Explicit tool_choice scanning disabled for passthrough** — The vLLM-specific heuristic that scans user message text for tool names and forces `tool_choice` was matching tool names (e.g., "Workflow") inside Claude Code's system prompt (CLAUDE.md embedded in user content). Now skips for client passthrough and messages longer than 500 chars.
 
 ## [0.10.1] - 2026-05-29
 

data/lib/legion/llm/api/namespaces/anthropic/messages.rb

@@ -63,6 +63,15 @@ module Legion
               log.info "[llm][api][anthropic] action=request request_id=#{request_id} " \
                        "messages=#{msg_count} chars=#{msg_chars} est_tokens=#{est_tokens} " \
                        "tools=#{tool_defs.size} stream=#{streaming}"
+              normalized[:messages].each_with_index do |m, i|
+                role = m[:role]
+                content = m[:content].is_a?(String) ? m[:content] : m[:content].to_s
+                tc = m[:tool_calls]
+                tcid = m[:tool_call_id]
+                log.debug "[llm][api][anthropic] action=request_msg idx=#{i} role=#{role} " \
+                          "content_length=#{content.length} content_preview=#{content[0, 120]} " \
+                          "tool_calls=#{tc&.size || 0} tool_call_id=#{tcid}"
+              end
 
               executor = Legion::LLM::Inference::Executor.new(pipeline_request)
               model = body[:model]
@@ -109,6 +118,21 @@ module Legion
                   stop_reason = tool_calls.any? ? 'tool_use' : translator.format_stop_reason(pipeline_response)
                   content_index = 0
 
+                  if !text_block_opened && tool_calls.empty?
+                    fallback_text = extract_fallback_text(pipeline_response)
+                    unless fallback_text.empty?
+                      out << "event: content_block_start\ndata: #{Legion::JSON.dump({
+                                                                                      type: 'content_block_start', index: 0,
+                        content_block: { type: 'text', text: '' }
+                                                                                    })}\n\n"
+                      out << "event: content_block_delta\ndata: #{Legion::JSON.dump({
+                                                                                      type: 'content_block_delta', index: 0,
+                        delta: { type: 'text_delta', text: fallback_text }
+                                                                                    })}\n\n"
+                      text_block_opened = true
+                    end
+                  end
+
                   log.info "[llm][api][anthropic] action=stream_post request_id=#{request_id} " \
                            "tool_calls=#{tool_calls.size} stop_reason=#{stop_reason} " \
                            "text_block_opened=#{text_block_opened} full_text_length=#{full_text.length}"
@@ -181,6 +205,21 @@ module Legion
                 halt 400, { 'Content-Type' => 'application/json' },
                      Legion::JSON.dump({ type: 'error', error: { type: 'invalid_request_error', message: "missing required fields: #{missing.join(', ')}" } })
               end
+
+              def extract_fallback_text(pipeline_response)
+                msg = pipeline_response.message
+                text = msg.is_a?(Hash) ? (msg[:content] || msg['content']).to_s : ''
+                return text unless text.empty?
+                return '' unless pipeline_response.respond_to?(:thinking) && pipeline_response.thinking
+
+                thinking_data = pipeline_response.thinking
+                thinking_content = if thinking_data.is_a?(Hash)
+                                     thinking_data[:content] || thinking_data['content']
+                                   elsif thinking_data.respond_to?(:content)
+                                     thinking_data.content
+                                   end
+                thinking_content.to_s.strip
+              end
             end
           end
         end

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

@@ -983,7 +983,19 @@ module Legion
         end
 
         def registry_tool_injection_requested?
-          !(@request.respond_to?(:suppress_tools) && @request.suppress_tools)
+          return false if @request.respond_to?(:suppress_tools) && @request.suppress_tools
+          return false if client_tools_only?
+
+          true
+        end
+
+        def client_tools_only?
+          return false unless client_tool_passthrough_enabled?
+
+          Array(@request.tools).any? do |tool|
+            source = tool.respond_to?(:source) ? tool.source : {}
+            source.is_a?(Hash) && source[:type] == :client
+          end
         end
 
         def client_tool_passthrough_enabled?

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

@@ -110,9 +110,10 @@ module Legion
 
         def explicit_native_tool_choice
           return unless @resolved_provider.to_s == 'vllm'
+          return if client_tools_only?
 
           text = latest_user_text.to_s.downcase
-          return if text.empty?
+          return if text.empty? || text.length > 500
 
           native_dispatch_tools.keys.map(&:to_s).sort_by { |tool_name| -tool_name.length }.find do |tool_name|
             explicit_tool_name_mentioned?(text, tool_name)

data/lib/legion/llm/version.rb

@@ -2,6 +2,6 @@
 
 module Legion
   module LLM
-    VERSION = '0.10.1'
+    VERSION = '0.10.2'
   end
 end

metadata

@@ -1,7 +1,7 @@
 --- !ruby/object:Gem::Specification
 name: legion-llm
 version: !ruby/object:Gem::Version
-  version: 0.10.1
+  version: 0.10.2
 platform: ruby
 authors:
 - Esity