feat: Oracle Canvas, Revision History and Canvas Sharing (#33)

Co-authored-by: Sagnik <sagnik7896@gmail.com>
Reviewed-on: sagnik/Project_Velocity#33

comfy_engine/scripts/prompt_expander.py

@@ -1,11 +1,12 @@
 #!/usr/bin/env python3
-"""
-Dream Weaver — Local LLM Prompt Expander
-========================================
-Converts user keywords + room type into a photorealistic interior design prompt
-using a local Ollama model (default: qwen3.5:27b).
-Cloud API calls (Gemini, OpenAI) have been completely removed for data privacy
-and local inference requirements.
+"""
+Dream Weaver — Shared Runtime Prompt Expander
+============================================
+Converts user keywords + room type into a photorealistic interior design prompt
+using the shared OpenAI-compatible Desineuron runtime (default: SGLang-hosted
+Qwen 3.6 35B A3B).
+Cloud API calls (Gemini, OpenAI SaaS) have been removed in favor of the routed
+Desineuron inference path.
 
 Usage:
     from prompt_expander import expand_prompt
@@ -18,7 +19,7 @@ Usage:
 import os
 import json
 import logging
-import requests
+import requests
 import re
 
 logger = logging.getLogger(__name__)
@@ -126,26 +127,44 @@ class ExpandedPrompt:
         self.source = source
 
 
-def _call_ollama(user_message: str) -> str:
-    ollama_url = os.environ.get("OLLAMA_URL", "http://localhost:11434")
-    # Using Qwen 3.5 27B as requested
-    model = os.environ.get("OLLAMA_MODEL", "qwen3.5:27b") 
-    full_prompt = f"{SYSTEM_PROMPT}\n\nUSER REQUEST:\n{user_message}\n\nReturn JSON ONLY. No markdown wrapping."
-    
-    r = requests.post(
-        f"{ollama_url}/api/generate",
-        json={
-            "model": model, 
-            "prompt": full_prompt,
-            "stream": False, 
-            "format": "json",
-            "options": {"temperature": 0.5}
-        },
-        timeout=180 # Large models take time
-    )
-    r.raise_for_status()
-    resp_json = r.json()
-    return resp_json["response"]
+def _call_runtime(user_message: str) -> str:
+    runtime_base = os.environ.get(
+        "SGLANG_BASE_URL",
+        os.environ.get(
+            "LLM_BASE_URL",
+            os.environ.get("OLLAMA_URL", "https://llm.desineuron.in"),
+        ),
+    ).rstrip("/")
+    chat_url = os.environ.get("SGLANG_CHAT_URL", f"{runtime_base}/v1/chat/completions")
+    model = os.environ.get(
+        "SGLANG_MODEL",
+        os.environ.get("OLLAMA_MODEL", "qwen3.6:35b-a3b"),
+    )
+    api_token = os.environ.get("SGLANG_API_TOKEN", "")
+    full_prompt = (
+        f"{SYSTEM_PROMPT}\n\nUSER REQUEST:\n{user_message}\n\nReturn JSON ONLY. No markdown wrapping."
+    )
+
+    headers = {"Content-Type": "application/json"}
+    if api_token:
+        headers["Authorization"] = f"Bearer {api_token}"
+
+    r = requests.post(
+        chat_url,
+        json={
+            "model": model,
+            "messages": [{"role": "user", "content": full_prompt}],
+            "temperature": 0.5,
+            "response_format": {"type": "json_object"},
+            "max_tokens": 1200,
+        },
+        headers=headers,
+        timeout=180,
+    )
+    r.raise_for_status()
+    resp_json = r.json()
+    message = ((resp_json.get("choices") or [{}])[0].get("message") or {}).get("content", "")
+    return message if isinstance(message, str) else json.dumps(message)
 
 
 def expand_prompt(keywords: list[str], room_type: str = "living_room", additional_notes: str = "") -> ExpandedPrompt:
@@ -164,16 +183,16 @@ AVOID: {ctx['avoid']}
 {f'NOTES: {additional_notes}' if additional_notes else ''}"""
 
     try:
-        logger.info("Calling local Ollama LLM...")
-        raw = _call_ollama(user_message).strip()
+        logger.info("Calling shared Desineuron runtime LLM...")
+        raw = _call_runtime(user_message).strip()
         
         # Log the raw response for debugging
         logger.info(f"Raw Ollama response length: {len(raw)}")
         
         # Handle empty response
         if not raw:
-            logger.error("Empty response from Ollama")
-            raise ValueError("Ollama returned an empty response")
+            logger.error("Empty response from shared runtime")
+            raise ValueError("Shared runtime returned an empty response")
 
         # Clean string of common junk (control characters, leading/trailing non-bracket junk)
         raw_cleaned = re.sub(r'[\x00-\x08\x0b\x0c\x0e-\x1f\x7f]', '', raw)
@@ -215,7 +234,7 @@ AVOID: {ctx['avoid']}
             source="ollama_local"
         )
     except Exception as e:
-        logger.error(f"Ollama LLM expansion failed: {e}")
+        logger.error(f"Shared runtime LLM expansion failed: {e}")
         import traceback
         traceback.print_exc()
         # Full fallback if anything goes wrong