feat: rework tool calling

dynamiq/callbacks/streaming.py

@@ -850,58 +850,32 @@ def _process_structured_output_mode(self, final_answer_only: bool) -> None:
                             action_value = buf[v_start + 1 : end_quote]
                             if action_value.strip().lower() == "finish":
                                 self._answer_started = True
-                                if self._current_state is None:
-                                    action_input_start = self._find_field_string_value_start(
-                                        buf, JSONStreamingField.ACTION_INPUT.value, end_quote + 1
-                                    )
-                                    if action_input_start != -1:
-                                        self._current_state = StreamingState.ANSWER
-                                        self._state_start_index = action_input_start
-                                        self._state_last_emit_index = max(
-                                            self._state_last_emit_index, action_input_start
-                                        )
                             else:
                                 self._tool_input_started = True
                                 self._current_action_name = action_value.strip()
                                 self.agent._streaming_tool_run_id = generate_uuid()
-                                if self._current_state is None:
-                                    action_input_start = self._find_field_string_value_start(
-                                        buf, JSONStreamingField.ACTION_INPUT.value, end_quote + 1
-                                    )
-                                    if action_input_start != -1:
-                                        self._current_state = StreamingState.TOOL_INPUT
-                                        self._state_start_index = action_input_start
-                                        self._state_last_emit_index = max(
-                                            self._state_last_emit_index, action_input_start
-                                        )
 
         if not self._state_has_emitted.get(StreamingState.REASONING, False):
             self._initialize_json_field_state(
                 buf, JSONStreamingField.THOUGHT.value, StreamingState.REASONING, final_answer_only
             )
 
-        if self._answer_started:
-            self._initialize_json_field_state(buf, JSONStreamingField.ACTION_INPUT.value, StreamingState.ANSWER)
-
-        if self._tool_input_started and not self._answer_started:
-            self._initialize_json_field_state(buf, JSONStreamingField.ACTION_INPUT.value, StreamingState.TOOL_INPUT)
+        # Initialize the action field (TOOL_INPUT for a tool call, ANSWER for finish).
+        # action_input is a JSON object in the current schema, so this falls back from the
+        # string form to the brace-delimited object form (see _so_initialize_action_field).
+        self._so_initialize_action_field(buf)
 
         if self._current_state == StreamingState.REASONING:
             if self._emit_json_field_content(buf, StreamingState.REASONING):
-                # Reasoning completed — immediately try to initialize ANSWER/TOOL_INPUT
-                # in the same call, in case this is the last chunk.
-                if self._answer_started:
-                    self._initialize_json_field_state(buf, JSONStreamingField.ACTION_INPUT.value, StreamingState.ANSWER)
-                elif self._tool_input_started:
-                    self._initialize_json_field_state(
-                        buf, JSONStreamingField.ACTION_INPUT.value, StreamingState.TOOL_INPUT
-                    )
+                # Reasoning completed — try to initialize ANSWER/TOOL_INPUT in the same
+                # call, in case this is the last chunk.
+                self._so_initialize_action_field(buf)
 
         if self._current_state == StreamingState.ANSWER:
-            if self._emit_json_field_content(buf, StreamingState.ANSWER):
+            if self._emit_answer_state(buf):
                 self._so_action_emitted = True
         elif self._current_state == StreamingState.TOOL_INPUT:
-            if self._emit_json_field_content(buf, StreamingState.TOOL_INPUT):
+            if self._emit_tool_input_state(buf):
                 self._so_action_emitted = True
 
     def _process_function_calling_mode(self, final_answer_only: bool) -> None:
@@ -1086,19 +1060,41 @@ def _try_initialize_next_json_field(self, buf: str, final_answer_only: bool) ->
                         buf, JSONStreamingField.ACTION_INPUT.value, StreamingState.TOOL_INPUT
                     )
 
-    def _emit_tool_input_state(self, buf: str) -> None:
-        """Emit content for the current TOOL_INPUT state."""
+    def _emit_tool_input_state(self, buf: str) -> bool:
+        """Emit content for the current TOOL_INPUT state. Returns True when complete."""
         if self._fc_object_tool_input:
-            self._emit_json_object_field_content(buf, StreamingState.TOOL_INPUT)
-        else:
-            self._emit_json_field_content(buf, StreamingState.TOOL_INPUT)
+            return self._emit_json_object_field_content(buf, StreamingState.TOOL_INPUT)
+        return self._emit_json_field_content(buf, StreamingState.TOOL_INPUT)
 
-    def _emit_answer_state(self, buf: str) -> None:
-        """Emit content for the current ANSWER state."""
+    def _emit_answer_state(self, buf: str) -> bool:
+        """Emit content for the current ANSWER state. Returns True when complete."""
         if self._fc_object_answer:
-            self._emit_json_object_field_content(buf, StreamingState.ANSWER)
-        else:
-            self._emit_json_field_content(buf, StreamingState.ANSWER)
+            return self._emit_json_object_field_content(buf, StreamingState.ANSWER)
+        return self._emit_json_field_content(buf, StreamingState.ANSWER)
+
+    def _so_initialize_action_field(self, buf: str) -> None:
+        """Initialize the ANSWER/TOOL_INPUT streaming state for structured output.
+
+        ``action_input`` is a JSON object in the current schema (tool args, or
+        ``{"answer": ...}`` for finish), though older models may still emit a plain
+        string. Try the string form first, then fall back to the brace-delimited
+        object form. For finish, prefer the nested ``answer`` string so only the
+        answer text streams, not the wrapping object.
+        """
+        if self._current_state is not None:
+            return
+        if self._answer_started:
+            if self._initialize_json_field_state(buf, JSONStreamingField.ANSWER.value, StreamingState.ANSWER):
+                return
+            if self._initialize_json_field_state(buf, JSONStreamingField.ACTION_INPUT.value, StreamingState.ANSWER):
+                return
+            self._initialize_json_object_field_state(buf, JSONStreamingField.ACTION_INPUT.value, StreamingState.ANSWER)
+        elif self._tool_input_started:
+            if self._initialize_json_field_state(buf, JSONStreamingField.ACTION_INPUT.value, StreamingState.TOOL_INPUT):
+                return
+            self._initialize_json_object_field_state(
+                buf, JSONStreamingField.ACTION_INPUT.value, StreamingState.TOOL_INPUT
+            )
 
     def _process_json_mode(self, final_answer_only: bool) -> None:
         """

dynamiq/nodes/agents/agent.py

@@ -1,6 +1,7 @@
 import json
+import types
 from concurrent.futures import as_completed
-from typing import Any, Callable, Literal, Mapping
+from typing import Any, Callable, Literal, Mapping, Union, get_args, get_origin
 
 from litellm import get_supported_openai_params, supports_function_calling, supports_response_schema
 from pydantic import BaseModel, Field, PrivateAttr, field_validator, model_validator
@@ -39,7 +40,7 @@
     StreamingMode,
 )
 from dynamiq.utils import generate_uuid, serialize_files_in_value
-from dynamiq.utils.json_parser import parse_llm_json_output
+from dynamiq.utils.json_parser import parse_llm_json_output, repair_truncated_json
 from dynamiq.utils.logger import logger
 
 
@@ -79,8 +80,11 @@ def parse_arguments(cls, v: Any) -> Any:
         if isinstance(v, str):
             try:
                 return json.loads(v, strict=False)
-            except json.JSONDecodeError as e:
-                raise ValueError(f"Tool call arguments are not valid JSON: {e}")
+            except json.JSONDecodeError:
+                try:
+                    return json.loads(repair_truncated_json(v), strict=False)
+                except json.JSONDecodeError as e:
+                    raise ValueError(f"Tool call arguments are not valid JSON: {e}")
         return v or {}
 
     def parse_as_tool_call(self) -> ToolCallArguments:
@@ -320,6 +324,93 @@ def _emit_tool_input_error(
         )
         self._streaming_tool_run_id = None
 
+    @staticmethod
+    def _annotation_accepts_none(annotation: Any) -> bool:
+        """Return True if a Pydantic field annotation includes ``NoneType``."""
+        if annotation is type(None):
+            return True
+        origin = get_origin(annotation)
+        if origin in (Union, types.UnionType):
+            return type(None) in get_args(annotation)
+        return False
+
+    @staticmethod
+    def _annotation_is_dict_like(annotation: Any) -> bool:
+        """Return True if the annotation is ``dict`` / ``dict[...]`` or a union including one."""
+        if annotation is dict:
+            return True
+        origin = get_origin(annotation)
+        if origin is dict:
+            return True
+        if origin in (Union, types.UnionType):
+            return any(Agent._annotation_is_dict_like(arg) for arg in get_args(annotation))
+        return False
+
+    @staticmethod
+    def _extract_basemodel(annotation: Any) -> type[BaseModel] | None:
+        """Return the BaseModel subclass in an annotation (handles ``Model | None``), else None."""
+        if isinstance(annotation, type) and issubclass(annotation, BaseModel):
+            return annotation
+        origin = get_origin(annotation)
+        if origin in (Union, types.UnionType):
+            for arg in get_args(annotation):
+                if isinstance(arg, type) and issubclass(arg, BaseModel):
+                    return arg
+        return None
+
+    def _coerce_json_fields(self, tool: Node, action_input: dict) -> dict:
+        """Parse stringified free-form dict fields back into dicts.
+
+        Strict mode can't express a free-form ``dict[str, Any]`` as an object, so
+        the schema transforms ship those fields as JSON-encoded strings (see the
+        provider converters). Here we reverse that: if the tool declares a
+        dict-typed field and the model supplied a JSON string for it, parse it
+        back so the tool's Pydantic schema validates the real dict.
+        """
+        fields = tool.input_schema.model_fields
+        for name, field in fields.items():
+            value = action_input.get(name)
+            if isinstance(value, str) and self._annotation_is_dict_like(field.annotation):
+                stripped = value.strip()
+                if stripped.startswith("{") and stripped.endswith("}"):
+                    try:
+                        action_input[name] = json.loads(stripped)
+                    except json.JSONDecodeError:
+                        pass  # leave as string; Pydantic will surface the error
+        return action_input
+
+    def _strip_protocol_nulls(self, tool: Node, action_input: dict) -> dict:
+        """Drop ``None`` values for fields whose Pydantic annotation rejects None.
+
+        OpenAI strict mode requires every property in ``required`` and uses
+        ``"null"`` in the type union as the signal for "leave it at the default."
+        Fields with a non-nullable default (``encoding: str = "utf-8"``) can't
+        accept that ``None`` directly — so we drop the key, letting the tool's
+        Pydantic default apply. Fields that genuinely accept ``None``
+        (``encoding: str | None = None``) keep it.
+
+        Recurses into nested ``BaseModel`` fields so the same applies at depth
+        (e.g. ``config.port`` where ``DBConfig.port: int = 8080``).
+        """
+        self._strip_nulls_for_fields(tool.input_schema.model_fields, action_input)
+        return action_input
+
+    def _strip_nulls_for_fields(self, fields: Mapping[str, Any], data: Any) -> None:
+        if not isinstance(data, dict):
+            return
+        for name in list(data):
+            field = fields.get(name)
+            if field is None:
+                continue
+            value = data[name]
+            if value is None:
+                if not self._annotation_accepts_none(field.annotation):
+                    del data[name]
+            elif isinstance(value, dict):
+                nested_model = self._extract_basemodel(field.annotation)
+                if nested_model is not None:
+                    self._strip_nulls_for_fields(nested_model.model_fields, value)
+
     def _should_delegate_final(
         self,
         tool: Node | None,
@@ -791,6 +882,7 @@ def _handle_function_calling_mode(
         if len(actual_tool_calls) > 1 and self.parallel_tool_calls_enabled:
             tool_items = []
             for tc in actual_tool_calls:
+                tc_name = tc.function.name.strip()
                 args = tc.function.parse_as_tool_call()
                 tc_input = args.action_input
                 if isinstance(tc_input, str):
@@ -800,9 +892,20 @@ def _handle_function_calling_mode(
                         raise ActionParsingException(f"Error parsing action_input string. {e}", recoverable=True)
                 if not isinstance(tc_input, dict):
                     tc_input = {"input": tc_input}
+                tc_tool = self.tool_by_names.get(self.sanitize_tool_name(tc_name))
+                if tc_tool is not None:
+                    self._coerce_json_fields(tc_tool, tc_input)
+                    self._strip_protocol_nulls(tc_tool, tc_input)
+                    try:
+                        tc_tool.input_schema.model_validate(tc_input)
+                    except Exception as e:
+                        raise ActionParsingException(
+                            f"Tool call for '{tc_name}' has invalid arguments: {e}",
+                            recoverable=True,
+                        )
                 tool_items.append(
                     ToolCallItem(
-                        name=tc.function.name.strip(),
+                        name=tc_name,
                         input=tc_input,
                         thought=args.thought,
                     )
@@ -826,6 +929,18 @@ def _handle_function_calling_mode(
         if not isinstance(action_input, dict):
             action_input = {"input": action_input}
 
+        tool = self.tool_by_names.get(self.sanitize_tool_name(action))
+        if tool is not None:
+            self._coerce_json_fields(tool, action_input)
+            self._strip_protocol_nulls(tool, action_input)
+            try:
+                tool.input_schema.model_validate(action_input)
+            except Exception as e:
+                raise ActionParsingException(
+                    f"Tool call for '{action}' has invalid arguments: {e}",
+                    recoverable=True,
+                )
+
         self.log_reasoning(thought, action, action_input, loop_num)
         return thought, action, action_input
 
@@ -866,8 +981,14 @@ def _handle_structured_output_mode(
             self._requested_output_files = self._parse_output_files_csv(
                 llm_generated_output_json.get("output_files") or ""
             )
-            self.log_final_output(thought, action_input, loop_num)
-            return thought, "final_answer", action_input
+            # action_input is now an object (per schema); the final answer lives
+            # under the ``answer`` key. Fall back to the raw value for backward
+            # compatibility with older models that still emit a plain string.
+            final_answer: Any = action_input
+            if isinstance(action_input, dict) and "answer" in action_input:
+                final_answer = action_input["answer"]
+            self.log_final_output(thought, final_answer, loop_num)
+            return thought, "final_answer", final_answer
 
         try:
             if isinstance(action_input, str):
@@ -1504,12 +1625,24 @@ def _run_react_llm_step(self, config: RunnableConfig | None, loop_num: int, **kw
 
         try:
             native_parallel = self.parallel_tool_calls_enabled and self.inference_mode == InferenceMode.FUNCTION_CALLING
+            # In FUNCTION_CALLING mode with tools present, force a tool call so
+            # the model cannot bail out with a text-only response. Honour any
+            # explicit caller override (kwargs / self.llm.tool_choice).
+            forced_tool_choice = None
+            if (
+                self.inference_mode == InferenceMode.FUNCTION_CALLING
+                and self._tools
+                and "tool_choice" not in kwargs
+                and getattr(self.llm, "tool_choice", None) is None
+            ):
+                forced_tool_choice = "required"
             llm_result = self._run_llm(
                 messages=messages,
                 tools=self._tools,
                 response_format=self._response_format,
                 config=llm_config,
                 parallel_tool_calls=True if native_parallel else None,
+                **({"tool_choice": forced_tool_choice} if forced_tool_choice else {}),
                 **kwargs,
             )
         finally: