mojentic 0.8.3__py3-none-any.whl → 0.9.0__py3-none-any.whl

_examples/broker_as_tool.py

@@ -10,7 +10,7 @@ from mojentic.llm.tools.tool_wrapper import ToolWrapper
 #
 
 temporal_specialist = BaseLLMAgent(
-    llm=LLMBroker(model="qwen2.5:7b"),
+    llm=LLMBroker(model="qwen3:7b"),
     tools=[ResolveDateTool()],
     behaviour="You are a historian and sociologist who focuses on sorting out temporal events, determining what happened or will happen when."
 )
@@ -22,7 +22,7 @@ if not os.path.exists("local"):
 fs = FilesystemGateway(base_path="local")
 
 knowledge_specialist = BaseLLMAgent(
-    llm=LLMBroker(model="llama3.3-70b-32k"),
+    llm=LLMBroker(model="qwen3:32b"),
     tools=[
         ListFilesTool(fs),
         ReadFileTool(fs),
@@ -34,7 +34,7 @@ knowledge_specialist = BaseLLMAgent(
 
 
 coordinator = BaseLLMAgent(
-    llm=LLMBroker(model="llama3.3-70b-32k"),
+    llm=LLMBroker(model="qwen3:32b"),
     behaviour="You are a coordinator who can manage multiple agents and delegate tasks to them to solve problems.",
     tools=[
         ToolWrapper(temporal_specialist, "temporal_specialist", "A historian and sociologist who focuses on sorting out temporal events, figuring out dates, determining what happened or will happen when."),

_examples/broker_examples.py

@@ -21,7 +21,7 @@ def openai_llm(model="gpt-5"):
     return llm
 
 
-def ollama_llm(model="llama3.3-70b-32k"):
+def ollama_llm(model="qwen3:32b"):
     llm = LLMBroker(model=model)
     return llm
 

_examples/broker_image_examples.py

@@ -13,7 +13,7 @@ def openai_llm(model="gpt-4o"):
     return llm
 
 
-def ollama_llm(model="llama3.3-70b-32k"):
+def ollama_llm(model="qwen3:32b"):
     llm = LLMBroker(model=model)
     return llm
 

_examples/characterize_ollama.py

@@ -13,7 +13,7 @@ def check_ollama_gateway():
         label: str = Field(..., description="The label describing the feeling.")
 
     response = gateway.complete(
-        model="llama3.2:1b",
+        model="qwen3:7b",
         messages=[LLMMessage(content="Hello, how are you?")],
         object_model=Feeling,
         temperature=1.0,
@@ -25,7 +25,7 @@ def check_ollama_gateway():
 
 def check_tools_call():
     response = chat(
-        model="llama3.3-70b-32k",
+        model="qwen3:32b",
         messages=[
             # {
             #     'role': 'user',

_examples/chat_session.py

@@ -1,6 +1,6 @@
 from mojentic.llm import ChatSession, LLMBroker
 
-llm_broker = LLMBroker(model="llama3.3-70b-32k")
+llm_broker = LLMBroker(model="qwen3:32b")
 chat_session = ChatSession(llm_broker)
 
 while True:

_examples/chat_session_with_tool.py

@@ -1,7 +1,7 @@
 from mojentic.llm import ChatSession, LLMBroker
 from mojentic.llm.tools.date_resolver import ResolveDateTool
 
-llm_broker = LLMBroker(model="llama3.3-70b-32k")
+llm_broker = LLMBroker(model="qwen3:32b")
 chat_session = ChatSession(llm_broker, tools=[ResolveDateTool()])
 
 while True:

_examples/coding_file_tool.py

@@ -38,8 +38,8 @@ api_key = os.getenv("OPENAI_API_KEY")
 gateway = OpenAIGateway(api_key)
 llm = LLMBroker(model="o4-mini", gateway=gateway)
 
-# llm = LLMBroker("qwen2.5-coder:32b")
-# llm = LLMBroker("llama3.3")
+# llm = LLMBroker("qwen3-coder:30b")
+llm = LLMBroker("qwen3-coder:30b")
 # llm = LLMBroker(model="qwen3-128k:32b")
 
 # Create a filesystem gateway for the sandbox
@@ -75,28 +75,28 @@ solver = IterativeProblemSolver(
     system_prompt="""
 # 0 - Project Identity & Context
 
-You are an expert and principled software engineer, well versed in writing Python games. You work 
-carefully and purposefully and always check your work with an eye to testability and correctness. 
-You know that every line of code you write is a liability, and you take care that every line 
+You are an expert and principled software engineer, well versed in writing Python games. You work
+carefully and purposefully and always check your work with an eye to testability and correctness.
+You know that every line of code you write is a liability, and you take care that every line
 matters.
 
 # 1 - Universal Engineering Principles
 
 * **Code is communication** — optimise for the next human reader.
-* **Simple Design Heuristics** — guiding principles, not iron laws; consult the user when you 
+* **Simple Design Heuristics** — guiding principles, not iron laws; consult the user when you
 need to break them.
   1. **All tests pass** — correctness is non‑negotiable.
   2. **Reveals intent** — code should read like an explanation.
-  3. **No *****knowledge***** duplication** — avoid multiple spots that must change together; 
+  3. **No *****knowledge***** duplication** — avoid multiple spots that must change together;
   identical code is only a smell when it hides duplicate *decisions*.
   4. **Minimal entities** — remove unnecessary indirection, classes, or parameters.
 * **Small, safe increments** — single‑reason commits; avoid speculative work (**YAGNI**).
 * **Tests are the executable spec** — red first, green always; test behaviour not implementation.
 * **Compose over inherit**; favour pure functions where practical, avoid side-effects.
-* **Functional core, imperative shell** — isolate pure business logic from I/O and side effects; 
+* **Functional core, imperative shell** — isolate pure business logic from I/O and side effects;
 push mutations to the system boundaries, build mockable gateways at those boundaries.
 * **Psychological safety** — review code, not colleagues; critique ideas, not authors.
-* **Version‑control etiquette** — descriptive commit messages, branch from `main`, PRs require 
+* **Version‑control etiquette** — descriptive commit messages, branch from `main`, PRs require
 green CI.
 
 # 2 - Python‑Specific Conventions
@@ -115,21 +115,21 @@ green CI.
 
 ## 2.2 Core Libraries
 
-Mandatory: pydantic, structlog, pytest, pytest-spec, pytest-cov, pytest-mock, flake8, black, 
-pre‑commit, mkdocs‑material. Add new libs only when they eliminate **significant** boilerplate or 
+Mandatory: pydantic, structlog, pytest, pytest-spec, pytest-cov, pytest-mock, flake8, black,
+pre‑commit, mkdocs‑material. Add new libs only when they eliminate **significant** boilerplate or
 risk.
 
 ## 2.3 Project Structure & Imports
 
 * **src‑layout**: code in `src/<package_name>/`; tests live beside code as `*_spec.py`.
-* Import order: 1) stdlib, 2) third‑party, 3) first‑party — each group alphabetised with a blank 
+* Import order: 1) stdlib, 2) third‑party, 3) first‑party — each group alphabetised with a blank
 line.
 
 ## 2.4 Naming & Style
 
 * `snake_case` for functions & vars, `PascalCase` for classes, `UPPER_SNAKE` for constants.
 * Prefix intentionally unused vars/args with `_`.
-* **flake8** (with plugins) handles linting, and **black** auto‑formats code. Max line length 
+* **flake8** (with plugins) handles linting, and **black** auto‑formats code. Max line length
 **100**.
 * Cyclomatic complexity cap: **10** (flake8 `C901`).
 * Use **f‑strings**; avoid magic numbers.
@@ -144,14 +144,14 @@ line.
 
 * Configure **structlog** for JSON output by default.
 * Never use `print` for diagnostics; reserve for user‑facing CLI UX.
-* Log levels: `DEBUG` (dev detail) → `INFO` (lifecycle) → `WARNING` (recoverable) → `ERROR` (user 
+* Log levels: `DEBUG` (dev detail) → `INFO` (lifecycle) → `WARNING` (recoverable) → `ERROR` (user
 visible).
 
 ## 2.7 Testing Strategy
 
 * **pytest** with **pytest-spec** for specification-style output.
 * Test files end with `_spec.py` and live in the same folder as the code under test.
-* Use **Arrange / Act / Assert** blocks separated by a blank line (no comments) **or** BDD 
+* Use **Arrange / Act / Assert** blocks separated by a blank line (no comments) **or** BDD
 `describe/should` classes.
 * Function names: use `should_*` and BDD-style specifications.
 * Class names: use `Describe*` and BDD-style test suites.
@@ -162,7 +162,7 @@ visible).
 # 3 - Planning and Goal Tracking
 
 - Use the provided task manager tools to create your plans and work through them step by step.
-- Before declaring yourself finished list all tasks, ensure they are all complete, and that you 
+- Before declaring yourself finished list all tasks, ensure they are all complete, and that you
 have not missed any steps
 - If you've missed or forgotten some steps, add them to the task list and continue
 - When all tasks are complete, and you can think of no more to add, declare yourself finished.

_examples/current_datetime_tool_example.py

@@ -4,7 +4,7 @@ from mojentic.llm.tools.current_datetime import CurrentDateTimeTool
 
 # Create an LLM broker with a specified model
 # You can change the model to any supported model
-llm = LLMBroker(model="qwen2.5:7b")  # Using the same model as in simple_tool.py
+llm = LLMBroker(model="qwen3:7b")  # Using the same model as in simple_tool.py
 
 # Create our custom tool
 datetime_tool = CurrentDateTimeTool()

_examples/ephemeral_task_manager_example.py

@@ -26,8 +26,8 @@ from mojentic.llm.tools.tell_user_tool import TellUserTool
 
 # llm = LLMBroker(model="qwen3:30b-a3b-q4_K_M")
 # llm = LLMBroker(model="qwen3:32b")
-llm = LLMBroker(model="qwen2.5:7b")
-# llm = LLMBroker(model="qwen2.5:72b")
+llm = LLMBroker(model="qwen3:7b")
+# llm = LLMBroker(model="qwen3:72b")
 # llm = LLMBroker(model="o4-mini", gateway=OpenAIGateway(os.environ["OPENAI_API_KEY"]))
 message = LLMMessage(
     content="I want you to count from 1 to 10. Break that request down into individual tasks, track them using available tools, and perform them one by one until you're finished. Interrupt me to tell the user as you complete every task.")

_examples/file_tool.py

@@ -52,11 +52,11 @@ This is an unfinished story about Ernie, the most adorable and colourful caterpi
 #
 
 
-# llm = LLMBroker("llama3.3-70b-32k")
-# llm = LLMBroker("llama3.1:70b")
-# llm = LLMBroker("llama3.1:8b")
-llm = LLMBroker("qwen2.5:7b")
-# llm = LLMBroker("llama3.3")
+# llm = LLMBroker("qwen3:32b")
+# llm = LLMBroker("qwen3:32b")
+# llm = LLMBroker("qwen3:7b")
+llm = LLMBroker("qwen3:7b")
+# llm = LLMBroker("qwen3:32b")
 # api_key = os.getenv("OPENAI_API_KEY")
 # gateway = OpenAIGateway(api_key)
 # llm = LLMBroker(model="gpt-4o-mini", gateway=gateway)

_examples/iterative_solver.py

@@ -18,7 +18,7 @@ from mojentic.llm import LLMBroker
 def main():
     # Initialize the LLM broker with your preferred model
     # Uncomment one of the following lines or modify as needed:
-    # llm = LLMBroker(model="llama3.3-70b-32k")  # Ollama model
+    # llm = LLMBroker(model="qwen3:32b")  # Ollama model
     # llm = LLMBroker(model="gpt-4o")  # OpenAI model
     llm = LLMBroker(model="qwq")  # Default model for example
 

_examples/react.py

@@ -6,7 +6,7 @@ from mojentic import Router, Dispatcher
 from mojentic.agents import OutputAgent
 from mojentic.llm import LLMBroker
 
-# llm = LLMBroker("llama3.3-70b-32k")
+# llm = LLMBroker("qwen3:32b")
 llm = LLMBroker("deepseek-r1:70b")
 thinking_agent = ThinkingAgent(llm)
 decisioning_agent = DecisioningAgent(llm)

_examples/recursive_agent.py

@@ -21,7 +21,7 @@ async def demonstrate_async():
     3. Running multiple problem-solving tasks concurrently
     """
     # Initialize the LLM broker with your preferred model
-    # llm = LLMBroker(model="llama3.3-70b-32k")
+    # llm = LLMBroker(model="qwen3:32b")
     llm = LLMBroker(model="qwen3:30b-a3b-q4_K_M")
 
     # Create the agent with a maximum of 3 iterations

_examples/simple_llm.py

@@ -25,9 +25,9 @@ class RequestAgent(BaseLLMAgent):
 
 
 # llm = LLMBroker("deepseek-r1:70b")
-# llm = LLMBroker("llama3.1-instruct-8b-32k")
-llm = LLMBroker("qwen2.5:0.5b")
-# llm = LLMBroker("llama3.1:8b", gateway=OllamaGateway(host="http://odin.local:11434"))
+# llm = LLMBroker("qwen3:14b")
+llm = LLMBroker("qwen3:0.5b")
+# llm = LLMBroker("qwen3:7b", gateway=OllamaGateway(host="http://odin.local:11434"))
 request_agent = RequestAgent(llm)
 output_agent = OutputAgent()
 

_examples/simple_llm_repl.py

@@ -27,7 +27,7 @@ class ChatOutputAgent(BaseAgent):
         return []
 
 
-llm = LLMBroker("llama3.1-instruct-8b-32k")
+llm = LLMBroker("qwen3:14b")
 request_agent = ChatAgent(llm)
 output_agent = ChatOutputAgent()
 

_examples/simple_structured.py

@@ -32,7 +32,7 @@ class RequestAgent(BaseLLMAgent):
         return [ResponseEvent(source=type(self), correlation_id=event.correlation_id, capitol=response)]
 
 
-llm = LLMBroker("llama3.1-instruct-8b-32k")
+llm = LLMBroker("qwen3:14b")
 request_agent = RequestAgent(llm)
 output_agent = OutputAgent()
 

_examples/simple_tool.py

@@ -33,8 +33,8 @@ class RequestAgent(BaseLLMAgent):
 
 
 # llm = LLMBroker("deepseek-r1:70b")
-# llm = LLMBroker("llama3.3-70b-32k")
-llm = LLMBroker("qwen2.5:7b")
+# llm = LLMBroker("qwen3:32b")
+llm = LLMBroker("qwen3:7b")
 request_agent = RequestAgent(llm)
 output_agent = OutputAgent()
 

_examples/solver_chat_session.py

@@ -43,12 +43,12 @@ class IterativeProblemSolverTool(LLMTool):
 
 def main():
     # llm = LLMBroker(model="MFDoom/deepseek-r1-tool-calling:14b")
-    # llm = LLMBroker(model="qwen2.5:14b")
-    # llm = LLMBroker(model="qwen2.5:14b")
-    # llm = LLMBroker(model="qwen2.5:7b")
+    # llm = LLMBroker(model="qwen3:14b")
+    # llm = LLMBroker(model="qwen3:14b")
+    # llm = LLMBroker(model="qwen3:7b")
     llm = LLMBroker(model="qwq")
     # llm = LLMBroker(model="qwq:32b-fp16")
-    # llm = LLMBroker(model="llama3.3-70b-32k")
+    # llm = LLMBroker(model="qwen3:32b")
 
     tools = [
         ResolveDateTool(),

_examples/streaming.py

@@ -1,34 +1,49 @@
+import os
+from mojentic.llm.llm_broker import LLMBroker
 from mojentic.llm.gateways.models import LLMMessage
 from mojentic.llm.gateways.ollama import OllamaGateway
+from mojentic.llm.gateways.openai import OpenAIGateway
 from mojentic.llm.tools.date_resolver import ResolveDateTool
 
-#
-# This is here 2025-02-21 to demonstrate a deficiency in Ollama/llama tool calling
-# using the Stream option. We can't get chunk by chunk responses from the LLM
-# when using tools. This limits our ability to explore streaming capabilities
-# in the mojentic API, so I'm pausing this work for now until this is resolved.
-# https://github.com/ollama/ollama/issues/7886
-#
-
 
 def main():
-    ollama = OllamaGateway()
+    """
+    Demonstrates streaming text generation with tool calling support.
+
+    This example shows how generate_stream() handles tool calls seamlessly:
+    1. Streams content as it arrives
+    2. Detects tool calls in the stream
+    3. Executes tools
+    4. Recursively streams the LLM's response after tool execution
+    """
+    gateway = OllamaGateway()
+    # gateway = OpenAIGateway(api_key=os.getenv("OPENAI_API_KEY"))
+    broker = LLMBroker(
+        model="qwen3:32b",
+        # model="gpt-5",
+        gateway=gateway
+    )
+
     date_tool = ResolveDateTool()
-    
-    stream = ollama.complete_stream(
-        model="qwen2.5:7b",
+
+    print("Streaming response with tool calling enabled...\n")
+
+    stream = broker.generate_stream(
         messages=[
-            LLMMessage(content="Tell me a story about a dragon. In your story, reference several dates relative to today, "
+            LLMMessage(content="Tell me a short story about a dragon. In your story, reference several dates relative to today, "
                               "like 'three days from now' or 'last week'.")
         ],
         tools=[date_tool],
-        temperature=0.5,
+        temperature=0.7,
         num_ctx=32768,
         num_predict=-1
     )
-    
+
     for chunk in stream:
-        print(chunk.content, end='', flush=True)
+        print(chunk, end='', flush=True)
+
+    print("\n\nDone!")
+
 
 if __name__ == "__main__":
-    main()
+    main()

_examples/tell_user_example.py

@@ -17,7 +17,7 @@ from mojentic.llm import LLMBroker
 def main():
     # Initialize the LLM broker with your preferred model
     # Uncomment one of the following lines or modify as needed:
-    # llm = LLMBroker(model="llama3.3-70b-32k")  # Ollama model
+    # llm = LLMBroker(model="qwen3:32b")  # Ollama model
     # llm = LLMBroker(model="gpt-4o")  # OpenAI model
     llm = LLMBroker(model="qwq")  # Default model for example
 

_examples/tracer_demo.py

@@ -1,8 +1,8 @@
 """
 Example script demonstrating the tracer system with ChatSession and tools.
 
-This example shows how to use the tracer system to monitor an interactive 
-chat session with LLMBroker and tools. When the user exits the session, 
+This example shows how to use the tracer system to monitor an interactive
+chat session with LLMBroker and tools. When the user exits the session,
 the script displays a summary of all traced events.
 
 It also demonstrates how correlation_id is used to trace related events
@@ -35,7 +35,7 @@ def main():
     tracer = TracerSystem()
 
     # Create an LLM broker with the tracer
-    llm_broker = LLMBroker(model="llama3.3-70b-32k", tracer=tracer)
+    llm_broker = LLMBroker(model="gpt-oss:20b", tracer=tracer)
 
     # Create a date resolver tool that will also use the tracer
     date_tool = ResolveDateTool(llm_broker=llm_broker, tracer=tracer)

_examples/working_memory.py

@@ -51,7 +51,7 @@ memory = SharedWorkingMemory({
 })
 
 llm = LLMBroker("deepseek-r1:70b")
-# llm = LLMBroker("llama3.3-instruct-70b-32k")
+# llm = LLMBroker("qwen3:32b-instruct-70b-32k")
 request_agent = RequestAgent(llm, memory)
 output_agent = OutputAgent()
 

mojentic/llm/gateways/ollama.py

@@ -1,4 +1,4 @@
-from typing import List, Iterator
+from typing import List, Iterator, Optional
 import structlog
 from ollama import Client, Options, ChatResponse
 from pydantic import BaseModel
@@ -10,8 +10,18 @@ from mojentic.llm.gateways.ollama_messages_adapter import adapt_messages_to_olla
 logger = structlog.get_logger()
 
 class StreamingResponse(BaseModel):
-    """Simple wrapper for streaming response content"""
-    content: str
+    """
+    Wrapper for streaming response chunks.
+
+    Attributes
+    ----------
+    content : Optional[str]
+        Text content chunk from the LLM response.
+    tool_calls : Optional[List]
+        Tool calls from the LLM response (raw ollama format).
+    """
+    content: Optional[str] = None
+    tool_calls: Optional[List] = None
 
 class OllamaGateway(LLMGateway):
     """
@@ -144,28 +154,21 @@ class OllamaGateway(LLMGateway):
             'stream': True
         }
 
-        #
-        # This is here 2025-02-21 to demonstrate a deficiency in Ollama tool calling
-        # using the Stream option. We can't get chunk by chunk responses from the LLM
-        # when using tools. This limits our ability to explore streaming capabilities
-        # in the mojentic API, so I'm pausing this work for now until this is resolved.
-        # https://github.com/ollama/ollama/issues/7886
-        #
-
-        # if 'tools' in args and args['tools'] is not None:
-        #     ollama_args['tools'] = [t.descriptor for t in args['tools']]
+        # Enable tool support if tools are provided
+        if 'tools' in args and args['tools'] is not None:
+            ollama_args['tools'] = [t.descriptor for t in args['tools']]
 
         stream = self.client.chat(**ollama_args)
 
         for chunk in stream:
             if chunk.message:
+                # Yield content chunks as they arrive
                 if chunk.message.content:
                     yield StreamingResponse(content=chunk.message.content)
-                # if chunk.message.tool_calls:
-                #     for tool_call in chunk.message.tool_calls:
-                #         yield StreamingResponse(
-                #             content=f"\nTOOL CALL: {tool_call.function.name}({tool_call.function.arguments})\n"
-                #         )
+
+                # Yield tool calls when they arrive
+                if chunk.message.tool_calls:
+                    yield StreamingResponse(tool_calls=chunk.message.tool_calls)
 
     def get_available_models(self) -> List[str]:
         """