Made Jarvis more mobile friendly

server/providers/anthropic_provider.py

@@ -83,7 +83,12 @@ class AnthropicProvider(AIProvider):
             "max_tokens": max_tokens,
         }
         if system:
-            params["system"] = system
+            # Wrap in a content block with cache_control so the system prompt is
+            # cached across repeated calls (e.g. email handling runs). Cached tokens
+            # cost ~10% of normal input price after the first write.
+            params["system"] = [
+                {"type": "text", "text": system, "cache_control": {"type": "ephemeral"}}
+            ]
         if tools:
             # aide tool schemas ARE Anthropic format — pass through directly
             params["tools"] = tools
@@ -163,10 +168,17 @@ class AnthropicProvider(AIProvider):
                     arguments=block.input,
                 ))
 
-        usage = UsageStats(
-            input_tokens=response.usage.input_tokens,
-            output_tokens=response.usage.output_tokens,
-        ) if response.usage else UsageStats()
+        if response.usage:
+            cache_read = getattr(response.usage, "cache_read_input_tokens", 0) or 0
+            cache_write = getattr(response.usage, "cache_creation_input_tokens", 0) or 0
+            usage = UsageStats(
+                input_tokens=response.usage.input_tokens,
+                output_tokens=response.usage.output_tokens,
+                cache_read_tokens=cache_read,
+                cache_write_tokens=cache_write,
+            )
+        else:
+            usage = UsageStats()
 
         finish_reason = response.stop_reason or "stop"
         if tool_calls:

server/providers/base.py

@@ -24,6 +24,8 @@ class ToolCallResult:
 class UsageStats:
     input_tokens: int = 0
     output_tokens: int = 0
+    cache_read_tokens: int = 0    # Anthropic: tokens served from cache (billed at ~10%)
+    cache_write_tokens: int = 0   # Anthropic: tokens written to cache (billed at ~125%)
 
     @property
     def total_tokens(self) -> int: