fix(claude_sdk,conversations,rotation): context size from the last api call, new-day text by rotation reason

src/beaver_gateway/backends/claude_sdk.py

@@ -391,7 +391,7 @@ class ClaudeSdkBackend:
             live.session_id = turn.result.session_id
         if conversation_id is None:
             self._rekey(live.key, fingerprint([*history, *turn.synthesized]))
-        usage = _usage_of(turn.result)
+        usage = _usage_of(turn.result, context_tokens=turn.context_tokens)
         interrupted = live.interrupt_requested
         live.interrupt_requested = False
         if capture is not None:
@@ -455,6 +455,7 @@ class ClaudeSdkBackend:
                 event = message.event
                 if event.get("type") == "message_start":
                     offset = next_index
+                    turn.context_tokens = _context_of(event.get("message"))
                     continue
                 index = event.get("index")
                 if isinstance(index, int):
@@ -740,6 +741,9 @@ class _SessionSpec:
 class _Turn:
     events: int = 0
     """Wire events already yielded to the caller."""
+
+    context_tokens: int = 0
+    """Input size of the latest API call (``message_start`` usage)."""
     synthesized: list[dict[str, Any]] = field(default_factory=list)
     result: ResultMessage | None = None
     stop_reason: StopReason = "end_turn"
@@ -895,11 +899,23 @@ def _block_to_dict(block: Any) -> dict[str, Any]:
     raise TypeError(msg)
 
 
-def _usage_of(result: ResultMessage | None) -> TurnUsage:
+def _context_of(message: Any) -> int:
+    usage = message.get("usage") if isinstance(message, dict) else None
+    if not isinstance(usage, dict):
+        return 0
+    return (
+        _int(usage.get("input_tokens"))
+        + _int(usage.get("cache_read_input_tokens"))
+        + _int(usage.get("cache_creation_input_tokens"))
+    )
+
+
+def _usage_of(result: ResultMessage | None, *, context_tokens: int = 0) -> TurnUsage:
     if result is None:
-        return TurnUsage()
+        return TurnUsage(context_tokens=context_tokens)
     usage = result.usage or {}
     return TurnUsage(
+        context_tokens=context_tokens,
         input_tokens=_int(usage.get("input_tokens")),
         output_tokens=_int(usage.get("output_tokens")),
         cache_read_tokens=_int(usage.get("cache_read_input_tokens")),