legion-llm 0.14.15 → 0.14.16

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: 6720d66411b7d2125aab243354ca9f0bf19292b10895695c7e762fe6f23180d0
-  data.tar.gz: 87910fc8387d28d65759e207f79bf6d1a2b39e178a4fefe5c8b6bd5f17187f6b
+  metadata.gz: 3eec433896cf314b65ac64111f414a55dc7e9639db3d54cc76359c9fb01ed567
+  data.tar.gz: 659857c6aa71573f6fbb795a655b7a59f2936672b0538db0bc14c18ae06ecfbd
 SHA512:
-  metadata.gz: 83e7680e3083b13e523f66cdba62ff28680f85bc4e63d0e59ef15b4d7cb260c28c02d1325fa76566f5bd8a6606dc55d552bce9dbf097d45d197a02a2a486ac9d
-  data.tar.gz: 9cbc000d34b4afdc8e24dc677ea9a0a22a397868bf1e765d96e2bf7ea173016d293441dbdb6401edd2de30380dd0892cfd73e65550797aca1f4fdaa81e08db0d
+  metadata.gz: 648b2bd80386141477bf5b77b387814f768d5f64447a921d34ef828dde6d9469a15e63211e63c4173e7188e0a61c5d60eb366415ccb68c43c031a23cd4889ac3
+  data.tar.gz: 22f1a949a48470e5a6b543a683e2de4f9c86f44d1c21d706e704bb4a5f58672311ba632b03710647beb071987abed2947e6299e0ad69c06b73c72484398fb238

data/CHANGELOG.md

@@ -1,5 +1,15 @@
 # Legion LLM Changelog
 
+## [0.14.16] - 2026-07-14
+
+### Fixed
+
+- **Server tool results are now visible to Claude and Codex streaming clients.** When the daemon executes a LegionIO tool server-side during the streaming tool loop, the result must reach the client in a format its SDK can parse. Two bugs prevented this:
+  - **Anthropic Messages (`/v1/messages`):** The `server_tool_result` SSE emitted `content: []` in `content_block_start` and put the actual result in a `content_block_delta` with `delta.type = "content_block_delta"` — a type the Anthropic TypeScript/Python SDK does not process. The client saw an empty result block. Fixed: result content is now placed directly in the `content_block_start` event; no unrecognized delta needed.
+  - **OpenAI Responses (`/v1/responses`):** The `on_server_tool_result` emitter was a no-op — tool results were silently discarded. Codex saw `function_call` items with no matching `function_call_output`. Fixed: emits a proper `response.output_item.added` + `response.output_item.done` pair with `type: "function_call_output"`, `call_id`, and the result text.
+
+- **Empty assistant messages left after thinking extraction are now stripped.** `strip_thinking_from_history` can leave behind assistant messages with no content and no tool_calls (the thinking was their only payload). These empty messages confused some providers. A second `empty_assistant_message?` pass now runs after thinking stripping.
+
 ## [0.14.15] - 2026-07-09
 
 ### Fixed

data/lib/legion/llm/api/client_translators/anthropic_messages.rb

@@ -292,12 +292,8 @@ module Legion
               emit('content_block_start', {
                      type:          'content_block_start',
                      index:         block_index,
-                     content_block: { type: 'server_tool_result', id: tool_call_id, content: [] }
-                   })
-              emit('content_block_delta', {
-                     type:  'content_block_delta',
-                     index: block_index,
-                     delta: { type: 'content_block_delta', content: [{ type: 'text', text: result_text }] }
+                     content_block: { type: 'server_tool_result', id: tool_call_id,
+                                      content: [{ type: 'text', text: result_text }] }
                    })
               emit('content_block_stop', { type: 'content_block_stop', index: block_index })
             end

data/lib/legion/llm/api/client_translators/openai_responses.rb

@@ -401,11 +401,18 @@ module Legion
 
             def on_server_tool_result(block_index:, tool_call_id:, result_text:)
               _ = block_index
-              _ = tool_call_id
-              _ = result_text
-              # /v1/responses doesn't expose server tool results inline today —
-              # the function_call output item carries id/args; result is part
-              # of the assistant message that follows on the next turn.
+              idx = @output_items.length
+              item = { id: "fco_#{tool_call_id}", type: 'function_call_output',
+                       call_id: tool_call_id, output: result_text.to_s, status: 'completed' }
+              @output_items << item
+              emit('response.output_item.added', {
+                     type: 'response.output_item.added', sequence_number: next_seq,
+                     output_index: idx, item: item
+                   })
+              emit('response.output_item.done', {
+                     type: 'response.output_item.done', sequence_number: next_seq,
+                     output_index: idx, item: item
+                   })
             end
 
             def on_keep_alive

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

@@ -17,6 +17,11 @@ module Legion
               messages = messages.reject { |m| empty_assistant_message?(m) }
             end
             messages = strip_thinking_from_history(messages)
+            post_thinking_rejected = messages.count { |m| empty_assistant_message?(m) }
+            if post_thinking_rejected.positive?
+              log.warn "[llm][executor] action=strip_empty_assistants_post_thinking request_id=#{@request.id} removed=#{post_thinking_rejected}"
+              messages = messages.reject { |m| empty_assistant_message?(m) }
+            end
             messages = trim_oversized_tool_results(messages)
             enforce_context_window(messages)
           end

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

@@ -170,12 +170,12 @@ module Legion
           end
 
           def pipeline_escalation_enabled?
-            esc = Legion::Settings[:llm].dig(:routing, :escalation) || {}
+            esc = Legion::Settings[:llm][:routing][:escalation]
             esc[:enabled] == true && esc[:pipeline_enabled] == true
           end
 
           def pipeline_escalation_max_attempts
-            esc = Legion::Settings[:llm].dig(:routing, :escalation) || {}
+            esc = Legion::Settings[:llm][:routing][:escalation]
             esc[:max_attempts] || 3
           end
 

data/lib/legion/llm/version.rb

@@ -2,6 +2,6 @@
 
 module Legion
   module LLM
-    VERSION = '0.14.15'
+    VERSION = '0.14.16'
   end
 end

metadata

@@ -1,7 +1,7 @@
 --- !ruby/object:Gem::Specification
 name: legion-llm
 version: !ruby/object:Gem::Version
-  version: 0.14.15
+  version: 0.14.16
 platform: ruby
 authors:
 - Esity