feat: add anthropic cached token values for local backends that support it

requests/anthropic.py

@@ -257,15 +257,46 @@ def finish_reason_to_stop_reason(finish_reason, has_tool_use=False):
     return _STOP_REASON_MAP.get(finish_reason, "end_turn")
 
 
+def _cached_prompt_tokens(usage):
+    """Read ``prompt_tokens_details.cached_tokens`` from a chat usage (dict or SDK obj).
+
+    OpenAI-compatible backends with automatic prefix caching (vLLM, recent
+    llama-server) report the reused-prefix token count here; backends that don't
+    populate it yield 0.
+    """
+    if not usage:
+        return 0
+    details = usage.get("prompt_tokens_details") if isinstance(usage, dict) \
+        else getattr(usage, "prompt_tokens_details", None)
+    if details is None:
+        return 0
+    if isinstance(details, dict):
+        return details.get("cached_tokens") or 0
+    return getattr(details, "cached_tokens", 0) or 0
+
+
 def usage_chat_to_anthropic(usage, cache_read_tokens=0, cache_creation_tokens=0):
-    """Map chat usage → Anthropic usage, folding in nomyo-cache attribution."""
-    prompt = (usage or {}).get("prompt_tokens") or 0
-    completion = (usage or {}).get("completion_tokens") or 0
+    """Map chat usage → Anthropic usage.
+
+    ``cached_tokens`` reported by the backend (automatic prefix-cache reuse) plus any
+    explicit ``cache_read_tokens`` become ``cache_read_input_tokens``; ``input_tokens``
+    is the remaining uncached prompt so ``input + cache_read`` equals the prompt total.
+
+    ``cache_creation_input_tokens`` stays at ``cache_creation_tokens`` (0 by default):
+    local backends do automatic KV-prefix reuse with no explicit ``cache_control``
+    write/breakpoint concept, so there is no honest "tokens written" figure to report —
+    that value is only real on the native Anthropic passthrough path.
+    """
+    prompt = (usage or {}).get("prompt_tokens") or 0 if isinstance(usage, dict) \
+        else (getattr(usage, "prompt_tokens", 0) or 0)
+    completion = (usage or {}).get("completion_tokens") or 0 if isinstance(usage, dict) \
+        else (getattr(usage, "completion_tokens", 0) or 0)
+    cache_read = _cached_prompt_tokens(usage) + cache_read_tokens
     return {
-        "input_tokens": prompt,
+        "input_tokens": max(prompt - cache_read, 0),
         "output_tokens": completion,
         "cache_creation_input_tokens": cache_creation_tokens,
-        "cache_read_input_tokens": cache_read_tokens,
+        "cache_read_input_tokens": cache_read,
     }
 
 
@@ -430,6 +461,7 @@ class ChatToMessagesStream:
                 self.usage = {
                     "prompt_tokens": getattr(usage, "prompt_tokens", 0) or 0,
                     "completion_tokens": getattr(usage, "completion_tokens", 0) or 0,
+                    "prompt_tokens_details": {"cached_tokens": _cached_prompt_tokens(usage)},
                 }
             choices = getattr(chunk, "choices", None)
             if not choices:
@@ -512,8 +544,15 @@ class ChatToMessagesStream:
                 "type": "tool_use", "id": state["id"], "name": state["name"], "input": parsed})
 
         self.stop_reason = finish_reason_to_stop_reason(finish_reason, has_tool_use=bool(tc_state))
-        out_tokens = (self.usage or {}).get("completion_tokens", 0)
+        final_usage = usage_chat_to_anthropic(
+            self.usage, cache_read_tokens=self.cache_read_tokens,
+            cache_creation_tokens=self.cache_creation_tokens)
         yield _sse("message_delta", {
             "delta": {"stop_reason": self.stop_reason, "stop_sequence": None},
-            "usage": {"output_tokens": out_tokens}})
+            "usage": {
+                "input_tokens": final_usage["input_tokens"],
+                "output_tokens": final_usage["output_tokens"],
+                "cache_read_input_tokens": final_usage["cache_read_input_tokens"],
+                "cache_creation_input_tokens": final_usage["cache_creation_input_tokens"],
+            }})
         yield _sse("message_stop", {})