fix(mavvrik_focus): carry prompt/completion token counts in FOCUS Tags

FOCUS v1.2 has no standard column for LLM token counts, and the shared
FocusTransformer used by every destination (Mavvrik, Vantage, CloudZero)
drops prompt_tokens/completion_tokens even though the source query
already selects them. Merge the two counts into the existing Tags JSON
column, which is the spec's own escape hatch for non-standard fields,
inside the Mavvrik-only export path so no shared transformer changes.

litellm/integrations/mavvrik_focus/mavvrik_focus_logger.py

@@ -19,10 +19,13 @@ overwrite each other within the same day, producing incomplete data.
 
 from __future__ import annotations
 
+import json
 import os
 from datetime import datetime, timedelta, timezone
 from typing import TYPE_CHECKING, Any, List, Optional
 
+import polars as pl
+
 import litellm
 from litellm._logging import verbose_proxy_logger
 from litellm.constants import MAVVRIK_FOCUS_EXPORT_JOB_NAME
@@ -34,6 +37,39 @@ if TYPE_CHECKING:
 else:
     AsyncIOScheduler = Any
 
+# FOCUS v1.2 has no standard column for token counts; core's transformer
+# drops prompt_tokens/completion_tokens even though the source query selects
+# them. Mavvrik carries them through as extra keys in the existing Tags JSON
+# column (the spec's own escape hatch for non-standard fields), rather than
+# changing the shared transformer used by every FOCUS destination.
+_TOKEN_TAG_KEYS = ("prompt_tokens", "completion_tokens")
+
+
+def _with_token_tags(data: pl.DataFrame, normalized: pl.DataFrame) -> pl.DataFrame:
+    """Merge prompt/completion token counts from the pre-transform frame into
+    ``normalized``'s Tags column. Rows correspond 1:1 and in the same order
+    across both frames -- transform() only adds/renames columns, it never
+    filters or reorders rows.
+    """
+    available = [k for k in _TOKEN_TAG_KEYS if k in data.columns]
+    if not available or len(data) != len(normalized):
+        return normalized
+
+    token_rows = data.select(available).to_dicts()
+
+    def _merge(tags_json: str, row: dict) -> str:
+        tags = json.loads(tags_json) if tags_json else {}
+        for key in available:
+            value = row.get(key)
+            if value is not None:
+                tags[key] = str(value)
+        return json.dumps(tags)
+
+    merged_tags = pl.Series(
+        [_merge(tags_json, row) for tags_json, row in zip(normalized["Tags"].to_list(), token_rows)]
+    )
+    return normalized.with_columns(merged_tags.alias("Tags"))
+
 
 def _parse_metrics_marker(
     marker: Optional[object],
@@ -136,6 +172,7 @@ class MavvrikFocusLogger(FocusLogger):
         else:
             normalized = engine._transformer.transform(data)
             if not normalized.is_empty():
+                normalized = _with_token_tags(data, normalized)
                 payload = engine._serializer.serialize(normalized)
         await engine._destination.deliver(
             content=payload or b"",