Fix metadata token counting for Anthropic format responses

The _log_metadata method only supported OpenAI format usage keys
(prompt_tokens, completion_tokens) but Anthropic responses use
different keys (input_tokens, output_tokens). This caused null
token counts in metadata.json for dedaluslabs and firmware providers
when using the /v1/messages endpoint.

Changes:
- Add fallback from OpenAI to Anthropic format for token counts
- Use explicit None checks instead of 'or' to handle 0 values
- Calculate total_tokens if missing from Anthropic responses
- Handle stop_reason (Anthropic) as well as finish_reason (OpenAI)

Author: b3nw

src/rotator_library/anthropic_compat/streaming.py

@@ -128,7 +128,10 @@ async def anthropic_streaming_wrapper(
                 stop_reason_final = stop_reason
 
                 # Build final usage dict with cached tokens
-                final_usage = {"output_tokens": output_tokens}
+                final_usage = {
+                    "input_tokens": input_tokens - cached_tokens,
+                    "output_tokens": output_tokens,
+                }
                 if cached_tokens > 0:
                     final_usage["cache_read_input_tokens"] = cached_tokens
                     final_usage["cache_creation_input_tokens"] = 0
@@ -416,7 +419,10 @@ async def anthropic_streaming_wrapper(
         yield f'event: content_block_stop\ndata: {{"type": "content_block_stop", "index": {current_block_index}}}\n\n'
 
         # Build final usage with cached tokens
-        final_usage = {"output_tokens": 0}
+        final_usage = {
+            "input_tokens": input_tokens - cached_tokens,
+            "output_tokens": 0,
+        }
         if cached_tokens > 0:
             final_usage["cache_read_input_tokens"] = cached_tokens
             final_usage["cache_creation_input_tokens"] = 0

src/rotator_library/transaction_logger.py

@@ -265,8 +265,12 @@ def _log_metadata(
         model = response_data.get("model", self.model)
         finish_reason = "N/A"
 
+        # Handle OpenAI format (choices[0].finish_reason)
         if "choices" in response_data and response_data["choices"]:
             finish_reason = response_data["choices"][0].get("finish_reason", "N/A")
+        # Handle Anthropic format (stop_reason at top level)
+        elif "stop_reason" in response_data:
+            finish_reason = response_data.get("stop_reason", "N/A")
 
         # Check for provider subdirectory
         has_provider_logs = False
@@ -279,6 +283,19 @@ def _log_metadata(
             except OSError:
                 has_provider_logs = False
 
+        # Extract token counts - support both OpenAI and Anthropic formats
+        # Prefers OpenAI format if available: prompt_tokens, completion_tokens
+        # Falls back to Anthropic format: input_tokens, output_tokens
+        prompt_tokens = usage.get("prompt_tokens")
+        if prompt_tokens is None:
+            prompt_tokens = usage.get("input_tokens")
+        completion_tokens = usage.get("completion_tokens")
+        if completion_tokens is None:
+            completion_tokens = usage.get("output_tokens")
+        total_tokens = usage.get("total_tokens")
+        if total_tokens is None and prompt_tokens is not None and completion_tokens is not None:
+            total_tokens = prompt_tokens + completion_tokens
+
         metadata = {
             "request_id": self.request_id,
             "timestamp_utc": datetime.utcnow().isoformat(),
@@ -288,9 +305,9 @@ def _log_metadata(
             "model": model,
             "streaming": self.streaming,
             "usage": {
-                "prompt_tokens": usage.get("prompt_tokens"),
-                "completion_tokens": usage.get("completion_tokens"),
-                "total_tokens": usage.get("total_tokens"),
+                "prompt_tokens": prompt_tokens,
+                "completion_tokens": completion_tokens,
+                "total_tokens": total_tokens,
             },
             "finish_reason": finish_reason,
             "has_provider_logs": has_provider_logs,