feat: refine metadata catalog workflows

harness/tht/internal/litellm_completion.py

@@ -265,6 +265,36 @@ def _response_content(response: object) -> str:
     return content
 
 
+def _response_usage(response: object) -> dict[str, int] | None:
+    """Normalize LiteLLM usage, tolerating providers that omit it."""
+    try:
+        usage = _field(response, "usage")
+        prompt = _field(usage, "prompt_tokens")
+        output = _field(usage, "completion_tokens")
+        if type(prompt) is not int or prompt < 0 or type(output) is not int or output < 0:
+            return None
+        cached = 0
+        for source, name in ((usage, "cache_read_input_tokens"), (usage, "cached_tokens")):
+            try:
+                value = _field(source, name)
+            except _InvalidResponse:
+                continue
+            if type(value) is int and value >= 0:
+                cached = value
+                break
+        try:
+            details = _field(usage, "prompt_tokens_details")
+            value = _field(details, "cached_tokens")
+            if type(value) is int and value >= 0:
+                cached = value
+        except _InvalidResponse:
+            pass
+        cached = min(cached, prompt)
+        return {"input": prompt - cached, "cacheRead": cached, "output": output}
+    except _InvalidResponse:
+        return None
+
+
 def handle_request(request: object, *, completion: Completion | None = None) -> Result:
     """Validate one request and normalize one injected or LiteLLM completion."""
     try:
@@ -284,7 +314,11 @@ def handle_request(request: object, *, completion: Completion | None = None) ->
             content = _response_content(response)
     except Exception:  # noqa: BLE001 - arbitrary provider objects are untrusted response data
         return {"ok": False, "error": "invalid_response"}
-    return {"ok": True, "content": content}
+    result: dict[str, object] = {"ok": True, "content": content}
+    usage = _response_usage(response)
+    if usage is not None:
+        result["usage"] = usage
+    return result
 
 
 def _write_json(payload: Mapping[str, object]) -> None: