Add llava vision: use Ollama llava:7b for all image analysis, Claude as fallback

- Add _ollama_vision_call() using Ollama /api/generate with images array
- handle_screenshot: try llava first (free, on-LAN), fall back to Claude on error;
  text-only snapshots now use ollama instead of groq
- handle_vision: same llava-first/Claude-fallback pattern; caller can force
  provider='ollama' or 'claude' explicitly; stored provider_used reflects actual
  provider that ran

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01X8tDRrQqgLjqXebMCBNcP3

deploy/reactor.py

@@ -175,6 +175,16 @@ async def _ollama_call(messages: list, system: str = "") -> str:
             data = await resp.json()
             return data.get("response", "")
 
+async def _ollama_vision_call(image_b64: str, prompt: str) -> str:
+    async with aiohttp.ClientSession() as session:
+        async with session.post(
+            f"{OLLAMA_HOST}/api/generate",
+            json={"model": "llava:7b", "prompt": prompt, "images": [image_b64], "stream": False},
+            timeout=aiohttp.ClientTimeout(total=120),
+        ) as resp:
+            data = await resp.json()
+            return data.get("response", "").strip()
+
 async def handle_llm(payload: dict) -> dict:
     message  = payload.get("message", "")
     system   = payload.get("system", "You are JARVIS, an Iron Man-style AI assistant.")
@@ -658,10 +668,16 @@ async def handle_screenshot(payload: dict) -> dict:
     height     = result.get("height", 0)
     file_size  = result.get("file_size", 0)
 
-    # Run Claude vision analysis if we have an image
+    # Run vision analysis if we have an image — llava first, Claude fallback
     analysis = ""
     provider_used = ""
     if do_analyze and image_b64:
+        try:
+            analysis = await _ollama_vision_call(image_b64, analyze_prompt)
+            provider_used = "ollama:llava"
+            log.info(f"[VISION] llava analysis complete ({len(analysis)} chars)")
+        except Exception as e:
+            log.warning(f"[VISION] llava failed: {e} — falling back to Claude")
             try:
                 import anthropic
                 client = anthropic.AsyncAnthropic(api_key=CLAUDE_API_KEY)
@@ -679,17 +695,17 @@ async def handle_screenshot(payload: dict) -> dict:
                 )
                 analysis = msg.content[0].text if msg.content else ""
                 provider_used = "claude"
-            log.info(f"[VISION] Claude analysis complete ({len(analysis)} chars)")
+                log.info(f"[VISION] Claude fallback analysis complete ({len(analysis)} chars)")
-        except Exception as e:
+            except Exception as e2:
-            log.warning(f"[VISION] Claude vision failed: {e}")
+                log.warning(f"[VISION] Claude fallback also failed: {e2}")
-            analysis = f"Vision analysis unavailable: {e}"
+                analysis = f"Vision analysis unavailable: {e2}"
     elif do_analyze and not image_b64 and result.get("snapshot_type") == "text":
-        # Text-only sysinfo snapshot — summarize with LLM
+        # Text-only sysinfo snapshot — summarize with Ollama
         try:
             snap_text = json.dumps(result, indent=2)[:3000]
             prompt = f"Summarize this server system snapshot for JARVIS. Highlight any concerns:\n\n{snap_text}"
-            analysis = await llm_call([{"role": "user", "content": prompt}], "groq")
+            analysis = await llm_call([{"role": "user", "content": prompt}], "ollama")
-            provider_used = "groq"
+            provider_used = "ollama"
         except Exception as e:
             analysis = f"Analysis unavailable: {e}"
 
@@ -742,8 +758,22 @@ async def handle_vision(payload: dict) -> dict:
     if not image_b64:
         raise ValueError("No image data provided")
 
-    log.info(f"[VISION] Analysis: screenshot_id={screenshot_id} agent={hostname}")
+    log.info(f"[VISION] Analysis: screenshot_id={screenshot_id} agent={hostname} provider={provider}")
 
+    analysis = ""
+    provider_used = provider
+
+    if provider == "ollama" or provider == "llava":
+        analysis = await _ollama_vision_call(image_b64, prompt)
+        provider_used = "ollama:llava"
+    else:
+        # Default: llava first, Claude fallback
+        try:
+            analysis = await _ollama_vision_call(image_b64, prompt)
+            provider_used = "ollama:llava"
+            log.info(f"[VISION] llava analysis complete ({len(analysis)} chars)")
+        except Exception as e:
+            log.warning(f"[VISION] llava failed: {e} — falling back to Claude")
             try:
                 import anthropic
                 client = anthropic.AsyncAnthropic(api_key=CLAUDE_API_KEY)
@@ -760,14 +790,15 @@ async def handle_vision(payload: dict) -> dict:
                     }],
                 )
                 analysis = msg.content[0].text if msg.content else ""
-    except Exception as e:
+                provider_used = "claude"
-        raise RuntimeError(f"Vision analysis failed: {e}")
+            except Exception as e2:
+                raise RuntimeError(f"Vision analysis failed (llava: {e}, claude: {e2})")
 
     # Update stored screenshot if we have an ID
     if screenshot_id:
         await db_execute(
             "UPDATE agent_screenshots SET vision_analysis=%s, vision_provider=%s WHERE id=%s",
-            (analysis, "claude", int(screenshot_id))
+            (analysis, provider_used, int(screenshot_id))
         )
 
     return {
@@ -775,7 +806,7 @@ async def handle_vision(payload: dict) -> dict:
         "screenshot_id": screenshot_id,
         "prompt":        prompt,
         "analysis":      analysis,
-        "provider":      "claude",
+        "provider":      provider_used,
     }