fix: preserve exploration tool history and cap context budget

artemis/services/token_meter.py

@@ -40,6 +40,7 @@ import threading
 from typing import Any
 
 from artemis.data_engine.context_vars import CURRENT_NODE_NAME
+from artemis.config.constants import MAX_CONTEXT_WINDOW_TOKENS
 from artemis.llm.usage import usage_from_message
 from artemis.utils.logger import get_logger
 
@@ -52,7 +53,7 @@ _METERS: dict[str, "SessionTokenMeter"] = {}
 # measured against when the UI reports "context used". The prompt size of the
 # last executor call (``context_base_tokens`` on its ``llm_usage`` trace) is
 # the numerator; this is the denominator.
-OPERATOR_CONTEXT_WINDOW_TOKENS = 1_000_000
+OPERATOR_CONTEXT_WINDOW_TOKENS = MAX_CONTEXT_WINDOW_TOKENS
 
 # Traced node names whose LLM calls carry the live executor context. Both the
 # Pro Operator and the Flash runner are "the operator" from the user's view.