CLI responding, but only to proprietary models

Need to investigate why self-hosted models via OLLAMA fail API-wise

Signed-off-by: Víctor Mayoral Vilches <v.mayoralv@gmail.com>

examples/cai/simple_test.py

@@ -0,0 +1,30 @@
+from cai.sdk.agents import Agent, Runner, AsyncOpenAI, OpenAIChatCompletionsModel
+import asyncio
+
+spanish_agent = Agent(
+    name="Spanish agent",
+    instructions="You only speak Spanish.",
+    model="o3-mini", 
+)
+
+english_agent = Agent(
+    name="English agent",
+    instructions="You only speak English",
+    model=OpenAIChatCompletionsModel( 
+        model="gpt-4o",
+        openai_client=AsyncOpenAI()
+    ),
+)
+
+triage_agent = Agent(
+    name="Triage agent",
+    instructions="Handoff to the appropriate agent based on the language of the request.",
+    handoffs=[spanish_agent, english_agent],
+    model="gpt-3.5-turbo",
+)
+
+async def main():
+    result = await Runner.run(english_agent, input="Hey there, what's up?")
+    print(result.final_output)
+
+asyncio.run(main())

src/cai/agents/__init__.py

@@ -139,7 +139,7 @@ def get_agent_module(agent_name: str) -> str:
 
     Returns:
         The full module name where the agent
-        is defined (e.g., 'cai.agents.basic')
+        is defined (e.g., 'cai.sdk.agents.basic')
     """
     # Try to import all agents from the agents folder
     for _, name, _ in pkgutil.iter_modules(__path__,
@@ -202,7 +202,7 @@ def get_agent_by_name(agent_name: str) -> Agent:
     
     # # Special handling for one_tool agent
     # if agent_name == "one_tool_agent":
-    #     from cai.agents.one_tool import one_tool_agent
+    #     from cai.sdk.agents.one_tool import one_tool_agent
         
     #     # Create handoffs between agents
     #     # Add a handoff from one_tool_agent to flag_discriminator

src/cai/agents/one_tool.py

@@ -2,46 +2,62 @@
 CTF Agent with one tool
 """
 import os
-from cai.sdk.agents import Agent
+from cai.sdk.agents import Agent, OpenAIChatCompletionsModel
 from cai.tools.reconnaissance.generic_linux_command import generic_linux_command  # noqa
-model = os.getenv('CAI_MODEL', "qwen2.5:14b")
+from openai import AsyncOpenAI
+
+# Get model from environment or use default
+model_name = os.getenv('CAI_MODEL', "qwen2.5:14b")
+
+# Create OpenAI client for the agent
+openai_client = AsyncOpenAI(
+    base_url = os.getenv('LITELLM_BASE_URL', 'http://localhost:4000'),
+    api_key=os.getenv('LITELLM_API_KEY', 'key')
+)
+
+# For Qwen models, we need to skip system instructions as they're not supported
+is_qwen = "qwen" in model_name.lower()
+instructions = None if is_qwen else """You are a Cybersecurity expert Leader facing a CTF
+                challenge.
+                INSTRUCTIONS:
+                1. Execute the generic_linux_command tool without any
+                explanation.
+                2. Never assume the flag format - it could be any string
+                3. YOU MUST USE THE flag_discriminator to check the flag
+                when you have it in DECODED 64-bit format
+
+                ## generic_linux_command usage
+                # 1. Regular command execution
+                # 2. Shell Session Management: create and manage
+                # interactive shell sessions for commands like netcat,
+                # SSH, telnet.
+
+                - To start a new session: Use `generic_linux_command` with
+                  commands like `ssh`
+                - To list active sessions:
+                  `generic_linux_command("session", "list")`
+                - To get output from a session:
+                  `generic_linux_command("session", "output <session_id>")`
+                - To send input to a session:
+                  `generic_linux_command("<command>", "<args>",
+                  session_id="<session_id>")`
+                - To terminate a session:
+                  `generic_linux_command("session", "kill <session_id>")`
+
+                """
 
 one_tool_agent = Agent(
-    model=model,
     name="CTF agent",
     description="""Agent focused on conquering security challenges using generic linux commands
                    Expert in cybersecurity and exploitation.""",
-    instructions="""You are a Cybersecurity expert Leader facing a CTF
+    instructions=instructions,
-                    challenge.
-                    INSTRUCTIONS:
-                    1. Execute the generic_linux_command tool without any
-                    explanation.
-                    2. Never assume the flag format - it could be any string
-                    3. YOU MUST USE THE flag_discriminator to check the flag
-                    when you have it in DECODED 64-bit format
-
-                    ## generic_linux_command usage
-                    # 1. Regular command execution
-                    # 2. Shell Session Management: create and manage
-                    # interactive shell sessions for commands like netcat,
-                    # SSH, telnet.
-
-                    - To start a new session: Use `generic_linux_command` with
-                      commands like `ssh`
-                    - To list active sessions:
-                      `generic_linux_command("session", "list")`
-                    - To get output from a session:
-                      `generic_linux_command("session", "output <session_id>")`
-                    - To send input to a session:
-                      `generic_linux_command("<command>", "<args>",
-                      session_id="<session_id>")`
-                    - To terminate a session:
-                      `generic_linux_command("session", "kill <session_id>")`
-
-                    """,
     tools=[
         generic_linux_command,
     ],
+    model=OpenAIChatCompletionsModel(
+        model=model_name,
+        openai_client=openai_client,
+    )
 )
 
 

src/cai/agents/thought.py

@@ -3,7 +3,7 @@ First prototype of a reasoner agent
 
 using reasoner as a tool call
 
-support meta agent may better @cai.agents.meta.reasoner_support
+support meta agent may better @cai.sdk.agents.meta.reasoner_support
 """
 from cai.tools.misc.reasoning import thought
 from cai.sdk.agents import Agent  # pylint: disable=import-error

src/cai/cli.py

@@ -99,7 +99,7 @@ Usage Examples:
 import os
 from dotenv import load_dotenv
 from openai import AsyncOpenAI
-from cai.sdk.agents import OpenAIChatCompletionsModel, Agent, Runner
+from cai.sdk.agents import OpenAIChatCompletionsModel, Agent, Runner, AsyncOpenAI
 from cai.sdk.agents import set_default_openai_client, set_tracing_disabled
 from openai.types.responses import ResponseTextDeltaEvent
 from rich.console import Console
@@ -126,22 +126,23 @@ external_client = AsyncOpenAI(
 set_default_openai_client(external_client)
 set_tracing_disabled(True)
 
-# # llm_model=os.getenv('LLM_MODEL', 'gpt-4o-mini')
+# llm_model=os.getenv('LLM_MODEL', 'gpt-4o-mini')
 # # llm_model=os.getenv('LLM_MODEL', 'claude-3-7')
-# llm_model=os.getenv('LLM_MODEL', 'qwen2.5:14b')
+llm_model=os.getenv('LLM_MODEL', 'qwen2.5:14b')
 
 
-# # For Qwen models, we need to skip system instructions as they're not supported
+# For Qwen models, we need to skip system instructions as they're not supported
-# instructions = None if "qwen" in llm_model.lower() else "You are a helpful assistant"
+instructions = None if "qwen" in llm_model.lower() else "You are a helpful assistant"
 
-# agent = Agent(
+agent = Agent(
-#     name="Assistant", 
+    name="Assistant", 
-#     instructions=instructions,
+    instructions=instructions,
-#     model=OpenAIChatCompletionsModel(
+    model=OpenAIChatCompletionsModel(
-#         model=llm_model,
+        model=llm_model,
-#         openai_client=external_client,
+        # openai_client=AsyncOpenAI()  # original OpenAI servers
-#     )
+        openai_client = external_client
-# )
+    )
+)
 
 def run_cai_cli(starting_agent, context_variables=None, stream=False, max_turns=float('inf')):
     """
@@ -226,35 +227,20 @@ def run_cai_cli(starting_agent, context_variables=None, stream=False, max_turns=
             turn_count += 1
         except KeyboardInterrupt:
             break
-        except Exception as e:
+        # except Exception as e:
-            import traceback
+        #     import traceback
-            import sys
+        #     import sys
-            exc_type, exc_value, exc_traceback = sys.exc_info()
+        #     exc_type, exc_value, exc_traceback = sys.exc_info()
-            tb_info = traceback.extract_tb(exc_traceback)
+        #     tb_info = traceback.extract_tb(exc_traceback)
-            filename, line, func, text = tb_info[-1]
+        #     filename, line, func, text = tb_info[-1]
-            console.print(f"[bold red]Error: {str(e)}[/bold red]")
+        #     console.print(f"[bold red]Error: {str(e)}[/bold red]")
-            console.print(f"[bold red]Traceback: {tb_info}[/bold red]")
+        #     console.print(f"[bold red]Traceback: {tb_info}[/bold red]")
 
 
 def main():
     # Get agent type from environment variables or use default
     agent_type = os.getenv('CAI_AGENT_TYPE', "one_tool_agent")
 
-    llm_model=os.getenv('LLM_MODEL', 'qwen2.5:14b')
-    # llm_model=os.getenv('LLM_MODEL', 'gpt-4o-mini')
-
-    # For Qwen models, we need to skip system instructions as they're not supported
-    instructions = None if "qwen" in llm_model.lower() else "You are a helpful assistant"
-
-    agent = Agent(
-        name="Assistant", 
-        instructions=instructions,
-        model=OpenAIChatCompletionsModel(
-            model=llm_model,
-            openai_client=external_client,
-        )
-    )
-
     # Get the agent instance by name
     agent = get_agent_by_name(agent_type)
 

tests/conftest.py

@@ -2,11 +2,11 @@ from __future__ import annotations
 
 import pytest
 
-from agents.models import _openai_shared
+from cai.sdk.agents.models import _openai_shared
-from agents.models.openai_chatcompletions import OpenAIChatCompletionsModel
+from cai.sdk.agents.models.openai_chatcompletions import OpenAIChatCompletionsModel
-from agents.models.openai_responses import OpenAIResponsesModel
+from cai.sdk.agents.models.openai_responses import OpenAIResponsesModel
-from agents.tracing import set_trace_processors
+from cai.sdk.agents.tracing import set_trace_processors
-from agents.tracing.setup import GLOBAL_TRACE_PROVIDER
+from cai.sdk.agents.tracing.setup import GLOBAL_TRACE_PROVIDER
 
 from .testing_processor import SPAN_PROCESSOR_TESTING
 

tests/testing_processor.py

@@ -4,7 +4,7 @@ import threading
 from datetime import datetime
 from typing import Any, Literal
 
-from agents.tracing import Span, Trace, TracingProcessor
+from cai.sdk.agents.tracing import Span, Trace, TracingProcessor
 
 TestSpanProcessorEvent = Literal["trace_start", "trace_end", "span_start", "span_end"]