iatoolkit 2.76.0__tar.gz → 2.76.2__tar.gz

{iatoolkit-2.76.0/src/iatoolkit.egg-info → iatoolkit-2.76.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: iatoolkit
-Version: 2.76.0
+Version: 2.76.2
 Summary: IAToolkit
 Author: Fernando Libedinsky
 License-Expression: MIT

{iatoolkit-2.76.0 → iatoolkit-2.76.2}/src/iatoolkit/__init__.py

@@ -3,7 +3,7 @@
 #
 # IAToolkit is open source software.
 
-__version__ = "2.76.0"
+__version__ = "2.76.2"
 
 # Expose main classes and functions at the top level of the package
 

{iatoolkit-2.76.0 → iatoolkit-2.76.2}/src/iatoolkit/services/history_manager_service.py

@@ -68,6 +68,11 @@ class HistoryManagerService:
     #: one a follow-up ("of those...") refers to; with two, one turn of 80k of
     #: results kept a compacted history at 130k.
     KEEP_TOOL_OUTPUT_TURNS = 1
+    #: ...unless what that turn carries for its tools (outputs, arguments and
+    #: reasoning) passes this: a turn of eight SQL rounds left 142k tokens for a
+    #: follow-up answered from the conversation in 5 seconds. Its answer stays,
+    #: and the model can call the tools again.
+    MAX_KEPT_TOOL_PAYLOAD_CHARS = 40000
     #: What is kept of a cut output: its beginning, enough to remember its shape.
     TRIMMED_TOOL_OUTPUT_CHARS = 1500
     TRIMMED_TOOL_OUTPUT_MARKER = "[tool output trimmed"
@@ -390,23 +395,43 @@ class HistoryManagerService:
 
     @classmethod
     def _trim_old_tool_outputs(cls, context_history: list):
-        """Cuts the long tool outputs of all but the last KEEP_TOOL_OUTPUT_TURNS turns, in place.
-
-        Each output keeps its message (a tool call without its output is a
-        history every provider rejects) and its beginning, plus a note naming the
-        tool so the model knows it can call it again. An output is cut once: the
-        stored history then stays the same from turn to turn, which is what a
-        provider's prompt cache reuses.
+        """Cuts what the turns before the last KEEP_TOOL_OUTPUT_TURNS carry for their tools, in place.
+
+        - Long tool outputs keep their beginning and a note naming the tool, so
+          the model knows it can call it again.
+        - Long tool call arguments (a whole dataset passed to an Excel, for
+          example) become a short valid JSON object: some adapters parse them.
+        - The reasoning of those turns is dropped: providers need it only within
+          the turn that produced it, and 8 rounds of it stayed in a compacted
+          history (110k tokens one turn after compacting, in production).
+
+        Every call keeps its message and its output message, so the history is
+        always valid for the provider. A message is cut once: the stored history
+        then stays the same from turn to turn, which is what a prompt cache reuses.
         """
         turn_starts = cls.turn_starts(context_history)
-        if len(turn_starts) <= cls.KEEP_TOOL_OUTPUT_TURNS:
+        if len(turn_starts) < cls.KEEP_TOOL_OUTPUT_TURNS or not turn_starts:
             return
         boundary = turn_starts[-cls.KEEP_TOOL_OUTPUT_TURNS]
+        kept_payload = cls._tool_payload_chars(context_history[boundary:])
+        if kept_payload > cls.MAX_KEPT_TOOL_PAYLOAD_CHARS:
+            logging.info(f"Last turn carries {kept_payload} characters for its tools: cut as an older turn.")
+            boundary = len(context_history)
         tool_names = {}
         for message in context_history[:boundary]:
-            for call in (message.get("tool_calls") or []) if isinstance(message, dict) else []:
-                if isinstance(call, dict):
-                    tool_names[call.get("id")] = (call.get("function") or {}).get("name")
+            if not isinstance(message, dict):
+                continue
+            message.pop("reasoning_content", None)
+            message.pop("reasoning_details", None)
+            for call in message.get("tool_calls") or []:
+                if not isinstance(call, dict):
+                    continue
+                function = call.get("function") if isinstance(call.get("function"), dict) else {}
+                tool_names[call.get("id")] = function.get("name")
+                arguments = function.get("arguments")
+                if isinstance(arguments, str) and len(arguments) > cls.TRIMMED_TOOL_OUTPUT_CHARS:
+                    function["arguments"] = json.dumps(
+                        {"trimmed": f"{len(arguments)} characters of arguments, from an earlier turn"})
         for message in context_history[:boundary]:
             if not isinstance(message, dict) or message.get("type") != "function_call_output":
                 continue
@@ -422,6 +447,24 @@ class HistoryManagerService:
                 f"Call {tool} again if you need all of it.]"
             )
 
+    @staticmethod
+    def _tool_payload_chars(messages: list) -> int:
+        """Characters of tool outputs, tool call arguments and reasoning in these messages."""
+        total = 0
+        for message in messages:
+            if not isinstance(message, dict):
+                continue
+            if message.get("type") == "function_call_output" and isinstance(message.get("output"), str):
+                total += len(message["output"])
+            total += len(str(message.get("reasoning_content") or ""))
+            if message.get("reasoning_details"):
+                total += len(json.dumps(message["reasoning_details"], default=str))
+            for call in message.get("tool_calls") or []:
+                function = call.get("function") if isinstance(call, dict) else None
+                if isinstance(function, dict) and isinstance(function.get("arguments"), str):
+                    total += len(function["arguments"])
+        return total
+
     def _cap_stored_history(self, company_short_name: str, model: Optional[str], context_history: list):
         """Keeps the stored history under STORED_HISTORY_MAX_BUDGETS budgets, whole turns at a time."""
         cap = self.context_budget_tokens(company_short_name, model) * self.STORED_HISTORY_MAX_BUDGETS

{iatoolkit-2.76.0 → iatoolkit-2.76.2/src/iatoolkit.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: iatoolkit
-Version: 2.76.0
+Version: 2.76.2
 Summary: IAToolkit
 Author: Fernando Libedinsky
 License-Expression: MIT