feat: enhance OpenRouter support with additional reasoning handling and model entries

EvoScientist/llm/models.py

@@ -120,13 +120,18 @@ _MODEL_ENTRIES: list[tuple[str, str, str]] = [
     ("kimi-k2.5", "Pro/moonshotai/Kimi-K2.5", "siliconflow"),
     ("glm-4.7", "Pro/zai-org/GLM-4.7", "siliconflow"),
     # OpenRouter
+    ("claude-opus-4.6", "anthropic/claude-opus-4.6", "openrouter"),
+    ("claude-sonnet-4.6", "anthropic/claude-sonnet-4.6", "openrouter"),
     ("gpt-5.4", "openai/gpt-5.4", "openrouter"),
-    ("minimax-m2.5", "minimax/minimax-m2.5", "openrouter"),
+    ("gpt-5.3-codex", "openai/gpt-5.3-codex", "openrouter"),
+    ("gemini-3.1-pro", "google/gemini-3.1-pro-preview", "openrouter"),
+    ("gemini-3-flash", "google/gemini-3-flash-preview", "openrouter"),
+    ("kimi-k2.5", "moonshotai/kimi-k2.5", "openrouter"),
+    ("glm-5v-turbo", "z-ai/glm-5v-turbo", "openrouter"),
+    ("minimax-m2.7", "minimax/minimax-m2.7", "openrouter"),
+    ("mimo-v2-pro", "xiaomi/mimo-v2-pro", "openrouter"),
     ("grok-4.1-fast", "x-ai/grok-4.1-fast", "openrouter"),
     ("qwen3.5-122b", "qwen/qwen3.5-122b-a10b", "openrouter"),
-    ("gemini-3-flash", "google/gemini-3-flash-preview", "openrouter"),
-    ("claude-sonnet-4.6", "anthropic/claude-sonnet-4.6", "openrouter"),
-    ("glm-5-turbo", "z-ai/glm-5-turbo", "openrouter"),
     # Zhipu CodePlan (智谱代码计划 — coding-only endpoint)
     ("glm-5.1", "glm-5.1", "zhipu-code"),
     ("glm-5", "glm-5", "zhipu-code"),

EvoScientist/stream/events.py

@@ -735,10 +735,21 @@ def _process_chunk_content(
     """Process content blocks from an AI message chunk."""
     content = chunk.content
 
+    # OpenRouter (langchain-openrouter) stores reasoning in
+    # additional_kwargs["reasoning_content"] instead of content blocks.
+    _additional = getattr(chunk, "additional_kwargs", None) or {}
+    _rc = _additional.get("reasoning_content")
+    _emitted_thinking = False
+    if _rc and isinstance(_rc, str):
+        yield emitter.thinking(_rc)
+        _emitted_thinking = True
+
     if isinstance(content, str):
         if content:
             cleaned = _strip_legacy_thinking_tags(content)
-            if cleaned:
+            # Skip whitespace-only text (OpenRouter may send '\n\n' before
+            # reasoning chunks, which would prematurely trigger is_responding).
+            if cleaned and not cleaned.isspace():
                 yield emitter.text(cleaned)
             return
 
@@ -771,7 +782,8 @@ def _process_chunk_content(
 
         if block_type in ("thinking", "reasoning"):
             thinking_text = block.get("thinking") or block.get("reasoning") or ""
-            if thinking_text:
+            # Skip if already emitted from additional_kwargs (avoid duplicates)
+            if thinking_text and not _emitted_thinking:
                 yield emitter.thinking(thinking_text)
 
         elif block_type == "text":