From 78084110188677b0180d7d4ac37d595ea82d3ed6 Mon Sep 17 00:00:00 2001 From: luijait Date: Tue, 6 May 2025 16:21:02 +0200 Subject: [PATCH 01/10] Qwen executes tool calls --- .../agents/models/openai_chatcompletions.py | 485 +++++++++++++++--- src/cai/sdk/agents/run.py | 81 +++ src/cai/tools/common.py | 99 +++- 3 files changed, 576 insertions(+), 89 deletions(-) diff --git a/src/cai/sdk/agents/models/openai_chatcompletions.py b/src/cai/sdk/agents/models/openai_chatcompletions.py index 0134fcbf..c5138d14 100644 --- a/src/cai/sdk/agents/models/openai_chatcompletions.py +++ b/src/cai/sdk/agents/models/openai_chatcompletions.py @@ -8,6 +8,8 @@ import litellm import tiktoken import inspect import hashlib +import re +import asyncio from collections.abc import AsyncIterator, Iterable from dataclasses import dataclass, field @@ -540,7 +542,27 @@ class OpenAIChatCompletionsModel(Model): streaming_text_buffer = "" # For tool call streaming, accumulate tool_calls to add to message_history at the end streamed_tool_calls = [] + + # Ollama specific: accumulate full content to check for function calls at the end + # Some Ollama models output the function call as JSON in the text content + ollama_full_content = "" + is_ollama = False + + model_str = str(self.model).lower() + is_ollama = self.is_ollama or "ollama" in model_str or ":" in model_str or "qwen" in model_str + # Add a small delay to make sure any previous tool outputs are fully rendered + # This helps prevent overlapping of panels in the terminal + await asyncio.sleep(0.2) + + # Add visual separation before agent output + if streaming_context and should_show_rich_stream: + # If we're using rich context, we'll add separation through that + pass + else: + # Print clear visual separator + print("\n") + async for chunk in stream: if not state.started: state.started = True @@ -593,6 +615,10 @@ class OpenAIChatCompletionsModel(Model): content = delta['content'] if content: + # For Ollama, we need to accumulate the full content to check for function calls + if is_ollama: + ollama_full_content += content + # Add to the streaming text buffer streaming_text_buffer += content @@ -776,6 +802,320 @@ class OpenAIChatCompletionsModel(Model): streamed_tool_calls.append(tool_call_msg) add_to_message_history(tool_call_msg) + # Special handling for Ollama - check if accumulated text contains a valid function call + if is_ollama and ollama_full_content and len(state.function_calls) == 0: + # Look for JSON object that might be a function call + try: + # Try to extract a JSON object from the content + json_start = ollama_full_content.find('{') + json_end = ollama_full_content.rfind('}') + 1 + + logger.debug(f"Ollama content length: {len(ollama_full_content)}, JSON start: {json_start}, JSON end: {json_end}") + + if json_start >= 0 and json_end > json_start: + json_str = ollama_full_content[json_start:json_end] + logger.debug(f"Extracted potential JSON: {json_str[:100]}...") + + # Special check for generic_linux_command format + if "generic_linux_command" in json_str and "arguments" in json_str: + logger.debug("Detected generic_linux_command pattern - using special handling") + + # Try regex pattern matching to handle potentially malformed JSON + cmd_pattern = re.search(r'"command"\s*:\s*"([^"]+)"', json_str) + args_pattern = re.search(r'"args"\s*:\s*"([^"]*)"', json_str) + ctf_pattern = re.search(r'"ctf"\s*:\s*"([^"]*)"', json_str) + + if cmd_pattern: + # Create a tool call specifically for generic_linux_command + command = cmd_pattern.group(1) + args = args_pattern.group(1) if args_pattern else "" + ctf = ctf_pattern.group(1) if ctf_pattern else "" + + tool_call_id = f"call_{hashlib.md5(('generic_linux_command' + str(time.time())).encode()).hexdigest()[:8]}" + + # Create a proper arguments object + command_args = { + "command": command, + "args": args, + "ctf": ctf, + "async_mode": False, + "session_id": "" + } + + arguments_str = json.dumps(command_args) + logger.debug(f"Created generic_linux_command with args: {arguments_str}") + + # Add it to our function_calls state + state.function_calls[0] = ResponseFunctionToolCall( + id=FAKE_RESPONSES_ID, + arguments=arguments_str, + name="generic_linux_command", + type="function_call", + call_id=tool_call_id, + ) + + # Display the tool call in CLI + from cai.util import cli_print_agent_messages + try: + # Create a message-like object to display the function call + tool_msg = type('ToolCallWrapper', (), { + 'content': None, + 'tool_calls': [ + type('ToolCallDetail', (), { + 'function': type('FunctionDetail', (), { + 'name': "generic_linux_command", + 'arguments': arguments_str + }), + 'id': tool_call_id, + 'type': 'function' + }) + ] + }) + + # Print the tool call using the CLI utility + cli_print_agent_messages( + agent_name=getattr(self, 'agent_name', 'Agent'), + message=tool_msg, + counter=getattr(self, 'interaction_counter', 0), + model=str(self.model), + debug=False, + interaction_input_tokens=estimated_input_tokens, + interaction_output_tokens=estimated_output_tokens, + interaction_reasoning_tokens=0, + total_input_tokens=getattr(self, 'total_input_tokens', 0) + estimated_input_tokens, + total_output_tokens=getattr(self, 'total_output_tokens', 0) + estimated_output_tokens, + total_reasoning_tokens=0, + interaction_cost=None, + total_cost=None + ) + except Exception as e: + logger.error(f"Error displaying tool call in CLI: {e}") + + # Add to message history + tool_call_msg = { + "role": "assistant", + "content": None, + "tool_calls": [ + { + "id": tool_call_id, + "type": "function", + "function": { + "name": "generic_linux_command", + "arguments": arguments_str + } + } + ] + } + + streamed_tool_calls.append(tool_call_msg) + add_to_message_history(tool_call_msg) + logger.debug(f"Added generic_linux_command with args: {arguments_str}") + + + # Standard JSON parsing for other function calls + parsed = json.loads(json_str) + + # Check if it looks like a function call + if ('name' in parsed and 'arguments' in parsed and + isinstance(parsed['arguments'], dict)): + + logger.debug(f"Found valid function call in Ollama output: {json_str}") + + # Create a tool call from the JSON object + tool_call_id = f"call_{hashlib.md5((parsed['name'] + str(time.time())).encode()).hexdigest()[:8]}" + + # Add it to our function_calls state + state.function_calls[0] = ResponseFunctionToolCall( + id=FAKE_RESPONSES_ID, + arguments=json.dumps(parsed['arguments']), + name=parsed['name'], + type="function_call", + call_id=tool_call_id, + ) + + # Ensure arguments is a valid JSON string + arguments_str = "" + if isinstance(parsed['arguments'], dict): + arguments_str = json.dumps(parsed['arguments']) + elif isinstance(parsed['arguments'], str): + # If it's already a string, check if it's valid JSON + try: + # Try parsing to validate + json.loads(parsed['arguments']) + arguments_str = parsed['arguments'] + except: + # If not valid JSON, encode it as a JSON string + arguments_str = json.dumps(parsed['arguments']) + else: + # For any other type, convert to string and then JSON + arguments_str = json.dumps(str(parsed['arguments'])) + + logger.debug(f"Final arguments string: {arguments_str}") + + # Update with the properly formatted arguments + state.function_calls[0].arguments = arguments_str + + # Log the tool call for development purposes + logger.debug(f"Adding tool call: {parsed['name']} with arguments: {arguments_str}") + + # Display the tool call in CLI + from cai.util import cli_print_agent_messages + try: + # Create a message-like object to display the function call + tool_msg = type('ToolCallWrapper', (), { + 'content': None, + 'tool_calls': [ + type('ToolCallDetail', (), { + 'function': type('FunctionDetail', (), { + 'name': parsed['name'], + 'arguments': arguments_str + }), + 'id': tool_call_id, + 'type': 'function' + }) + ] + }) + + # Print the tool call using the CLI utility + cli_print_agent_messages( + agent_name=getattr(self, 'agent_name', 'Agent'), + message=tool_msg, + counter=getattr(self, 'interaction_counter', 0), + model=str(self.model), + debug=False, + interaction_input_tokens=estimated_input_tokens, + interaction_output_tokens=estimated_output_tokens, + interaction_reasoning_tokens=0, # Not available for Ollama + total_input_tokens=getattr(self, 'total_input_tokens', 0) + estimated_input_tokens, + total_output_tokens=getattr(self, 'total_output_tokens', 0) + estimated_output_tokens, + total_reasoning_tokens=getattr(self, 'total_reasoning_tokens', 0), + interaction_cost=None, + total_cost=None, + tool_output=None # Will be shown once the tool is executed + ) + except Exception as e: + logger.error(f"Error displaying tool call in CLI: {e}") + + # Add to message history + tool_call_msg = { + "role": "assistant", + "content": None, + "tool_calls": [ + { + "id": tool_call_id, + "type": "function", + "function": { + "name": parsed['name'], + "arguments": arguments_str + } + } + ] + } + + streamed_tool_calls.append(tool_call_msg) + add_to_message_history(tool_call_msg) + + logger.debug(f"Added function call: {parsed['name']} with args: {json.dumps(parsed['arguments'])}") + except Exception as e: + logger.debug(f"Failed to parse potential Ollama function call: {e}") + + # Even if JSON parsing fails, try to extract generic_linux_command with regex + if "generic_linux_command" in ollama_full_content: + logger.debug("JSON parsing failed but detected generic_linux_command - trying regex fallback") + try: + # Use regex to extract command components even from malformed output + cmd_pattern = re.search(r'"command"\s*:\s*"([^"]+)"', ollama_full_content) + args_pattern = re.search(r'"args"\s*:\s*"([^"]*)"', ollama_full_content) + ctf_pattern = re.search(r'"ctf"\s*:\s*"([^"]*)"', ollama_full_content) + + if cmd_pattern: + # Create a tool call specifically for generic_linux_command + command = cmd_pattern.group(1) + args = args_pattern.group(1) if args_pattern else "" + ctf = ctf_pattern.group(1) if ctf_pattern else "" + + tool_call_id = f"call_{hashlib.md5(('generic_linux_command' + str(time.time())).encode()).hexdigest()[:8]}" + + # Create a proper arguments object + command_args = { + "command": command, + "args": args, + "ctf": ctf, + "async_mode": False, + "session_id": "" + } + + arguments_str = json.dumps(command_args) + logger.debug(f"Fallback created generic_linux_command with args: {arguments_str}") + + # Add it to our function_calls state + state.function_calls[0] = ResponseFunctionToolCall( + id=FAKE_RESPONSES_ID, + arguments=arguments_str, + name="generic_linux_command", + type="function_call", + call_id=tool_call_id, + ) + + # Display the tool call in CLI + from cai.util import cli_print_agent_messages + try: + # Create a message-like object for display + tool_msg = type('ToolCallWrapper', (), { + 'content': None, + 'tool_calls': [ + type('ToolCallDetail', (), { + 'function': type('FunctionDetail', (), { + 'name': "generic_linux_command", + 'arguments': arguments_str + }), + 'id': tool_call_id, + 'type': 'function' + }) + ] + }) + + cli_print_agent_messages( + agent_name=getattr(self, 'agent_name', 'Agent'), + message=tool_msg, + counter=getattr(self, 'interaction_counter', 0), + model=str(self.model), + debug=False, + interaction_input_tokens=estimated_input_tokens, + interaction_output_tokens=estimated_output_tokens, + interaction_reasoning_tokens=0, + total_input_tokens=getattr(self, 'total_input_tokens', 0) + estimated_input_tokens, + total_output_tokens=getattr(self, 'total_output_tokens', 0) + estimated_output_tokens, + total_reasoning_tokens=0, + interaction_cost=None, + total_cost=None + ) + except Exception as e: + logger.error(f"Error displaying fallback tool call in CLI: {e}") + + # Add to message history + tool_call_msg = { + "role": "assistant", + "content": None, + "tool_calls": [ + { + "id": tool_call_id, + "type": "function", + "function": { + "name": "generic_linux_command", + "arguments": arguments_str + } + } + ] + } + + streamed_tool_calls.append(tool_call_msg) + add_to_message_history(tool_call_msg) + + logger.debug(f"Added fallback generic_linux_command") + except Exception as regex_err: + logger.error(f"Regex fallback also failed: {regex_err}") + function_call_starting_index = 0 if state.text_content_index_and_output: function_call_starting_index += 1 @@ -952,13 +1292,26 @@ class OpenAIChatCompletionsModel(Model): direct_stats = final_stats.copy() direct_stats["interaction_cost"] = float(interaction_cost) direct_stats["total_cost"] = float(total_cost) + + # Add a small delay to avoid overlapping outputs + await asyncio.sleep(0.3) + # Use the direct copy with guaranteed float costs finish_agent_streaming(streaming_context, direct_stats) + + # Add visual separation after agent output completes + print("\n") # If we're not using rich streaming and not suppressing output, use old method elif not self.suppress_final_output and final_response.output and any(isinstance(item, ResponseOutputMessage) for item in final_response.output): # Find the assistant message to print for item in final_response.output: if isinstance(item, ResponseOutputMessage) and item.role == 'assistant': + # Add a small delay to avoid overlapping outputs + await asyncio.sleep(0.3) + + # Print clear visual separator before message + print("\n") + cli_print_agent_messages( agent_name=getattr(self, 'agent_name', 'Agent'), message=item, @@ -974,6 +1327,9 @@ class OpenAIChatCompletionsModel(Model): interaction_cost=interaction_cost, total_cost=total_cost, ) + + # Add visual separation after message + print("\n") break # --- Add assistant tool call(s) to message_history at the end of streaming --- @@ -1194,6 +1550,7 @@ class OpenAIChatCompletionsModel(Model): # Use the specialized Qwen approach first return await self._fetch_response_litellm_ollama(kwargs, model_settings, tool_choice, stream, parallel_tool_calls) except Exception as qwen_e: + print(qwen_e) # If that fails, try our direct OpenAI approach qwen_params = kwargs.copy() qwen_params["api_base"] = get_ollama_api_base() @@ -1351,117 +1708,80 @@ class OpenAIChatCompletionsModel(Model): # Standard OpenAI handling for non-streaming ret = litellm.completion(**kwargs) return ret - + async def _fetch_response_litellm_ollama( self, kwargs: dict, model_settings: ModelSettings, tool_choice: ChatCompletionToolChoiceOptionParam | NotGiven, stream: bool, - parallel_tool_calls: bool + parallel_tool_calls: bool, + provider="ollama" ) -> ChatCompletion | tuple[Response, AsyncStream[ChatCompletionChunk]]: + # Extract only supported parameters for Ollama ollama_supported_params = { "model": kwargs.get("model", ""), "messages": kwargs.get("messages", []), + "stream": kwargs.get("stream", False) } - # Safely add optional parameters only if they exist in kwargs - if "temperature" in kwargs: - ollama_supported_params["temperature"] = kwargs["temperature"] if kwargs["temperature"] is not NOT_GIVEN else None - if "top_p" in kwargs: - ollama_supported_params["top_p"] = kwargs["top_p"] if kwargs["top_p"] is not NOT_GIVEN else None - if "max_tokens" in kwargs: - ollama_supported_params["max_tokens"] = kwargs["max_tokens"] if kwargs["max_tokens"] is not NOT_GIVEN else None - - # Add stream parameter - default to False if not present - ollama_supported_params["stream"] = kwargs.get("stream", False) + # Add optional parameters if they exist and are not NOT_GIVEN + for param in ["temperature", "top_p", "max_tokens"]: + if param in kwargs and kwargs[param] is not NOT_GIVEN: + ollama_supported_params[param] = kwargs[param] # Add extra headers if available if "extra_headers" in kwargs: ollama_supported_params["extra_headers"] = kwargs["extra_headers"] - # IMPORTANT: For tool calls with Ollama (especially with Qwen), - # we need to pass the tools and tool_choice as part of the request - # This is needed for both streaming and non-streaming modes + # Add tools and tool_choice for compatibility with Qwen if "tools" in kwargs and kwargs.get("tools") and kwargs.get("tools") is not NOT_GIVEN: ollama_supported_params["tools"] = kwargs.get("tools") - # Include tool_choice if present and not NOT_GIVEN if "tool_choice" in kwargs and kwargs.get("tool_choice") is not NOT_GIVEN: ollama_supported_params["tool_choice"] = kwargs.get("tool_choice") - # Modify the messages to remove system message for Ollama - if (ollama_supported_params["messages"] and - len(ollama_supported_params["messages"]) > 0 and - ollama_supported_params["messages"][0].get("role") == "system"): - # Extract the system message - system_content = ollama_supported_params["messages"][0].get("content", "") - # Remove it from the messages - ollama_supported_params["messages"] = ollama_supported_params["messages"][1:] - # If there are user messages, prepend system to first user - if (ollama_supported_params["messages"] and - len(ollama_supported_params["messages"]) > 0 and - ollama_supported_params["messages"][0].get("role") == "user"): - # Prepend the system instruction to the first user message, with a separator - user_content = ollama_supported_params["messages"][0].get("content", "") - if isinstance(user_content, str): - ollama_supported_params["messages"][0]["content"] = f"System: {system_content}\n\nUser: {user_content}" - # Remove None values ollama_kwargs = {k: v for k, v in ollama_supported_params.items() if v is not None} - # Detect if this is a Qwen model + # Check if this is a Qwen model model_str = str(self.model).lower() is_qwen = "qwen" in model_str - + + api_base = get_ollama_api_base() + if "ollama" in provider: + api_base = api_base.rstrip('/v1') + # Create response object for streaming if stream: - # For streaming with Ollama, we need to create a Response object first response = Response( id=FAKE_RESPONSES_ID, created_at=time.time(), model=self.model, object="response", output=[], - tool_choice="auto" if tool_choice is None or tool_choice == NOT_GIVEN else cast(Literal["auto", "required", "none"], tool_choice), + tool_choice="auto" if tool_choice is None or tool_choice == NOT_GIVEN else + cast(Literal["auto", "required", "none"], tool_choice), top_p=model_settings.top_p, temperature=model_settings.temperature, tools=[], parallel_tool_calls=parallel_tool_calls or False, ) - - # Special handling for Qwen when streaming to ensure tool calls are properly handled - if is_qwen and os.getenv('CAI_STREAM', 'false').lower() == 'true': - # Make sure we have the right custom provider in streaming mode - stream_obj = await litellm.acompletion( - **ollama_kwargs, - api_base=get_ollama_api_base(), - custom_llm_provider="openai" # Use openai provider for best compatibility with Qwen - ) - else: - # Standard Ollama streaming - stream_obj = await litellm.acompletion( - **ollama_kwargs, - api_base=get_ollama_api_base().rstrip('/v1'), - custom_llm_provider="ollama" - ) + # Get streaming response + stream_obj = await litellm.acompletion( + **ollama_kwargs, + api_base=api_base, + custom_llm_provider=provider, + ) return response, stream_obj else: - # Non-streaming mode - # For Qwen models, use the openai provider when tools are present - if is_qwen and "tools" in ollama_kwargs: - ret = litellm.completion( - **ollama_kwargs, - api_base=get_ollama_api_base(), - custom_llm_provider="openai" # Use openai provider for better tool handling - ) - else: - # Standard Ollama completion - ret = litellm.completion( - **ollama_kwargs, - api_base=get_ollama_api_base().rstrip('/v1'), - custom_llm_provider="ollama" - ) - return ret + + + # Get completion response + return litellm.completion( + **ollama_kwargs, + api_base=api_base, + custom_llm_provider=provider, + ) def _get_client(self) -> AsyncOpenAI: if self._client is None: @@ -1494,6 +1814,33 @@ class OpenAIChatCompletionsModel(Model): 'arguments': function_call.get('arguments', '') } }] + + # Handle special Ollama generic_linux_command format + if isinstance(delta, dict) and 'content' in delta: + content = delta['content'] + # Try to detect if the content is a JSON string with function call format + try: + if isinstance(content, str) and '{' in content and '}' in content: + # Try to extract JSON from the content (it might be embedded in text) + json_start = content.find('{') + json_end = content.rfind('}') + 1 + if json_start >= 0 and json_end > json_start: + json_str = content[json_start:json_end] + parsed = json.loads(json_str) + if 'name' in parsed and 'arguments' in parsed: + # This looks like a function call in JSON format + return [{ + 'index': 0, + 'id': f"call_{time.time_ns()}", # Generate a unique ID + 'type': 'function', + 'function': { + 'name': parsed['name'], + 'arguments': json.dumps(parsed['arguments']) if isinstance(parsed['arguments'], dict) else parsed['arguments'] + } + }] + except Exception: + # If JSON parsing fails, just continue with normal processing + pass # Anthropic-style tool_use format if hasattr(delta, 'tool_use') and delta.tool_use: diff --git a/src/cai/sdk/agents/run.py b/src/cai/sdk/agents/run.py index 84bb664b..a0937d28 100644 --- a/src/cai/sdk/agents/run.py +++ b/src/cai/sdk/agents/run.py @@ -4,6 +4,7 @@ import asyncio import copy from dataclasses import dataclass, field from typing import Any, cast +import os from openai.types.responses import ResponseCompletedEvent @@ -778,6 +779,23 @@ class Runner: run_config: RunConfig, tool_use_tracker: AgentToolUseTracker, ) -> SingleStepResult: + # Log the raw model response output, focusing on tool calls + logger.debug(f"[_get_single_step_result_from_response] Raw new_response.id: {new_response.referenceable_id}") + if new_response.output: + for i, item in enumerate(new_response.output): + item_type = type(item).__name__ + logger.debug(f"[_get_single_step_result_from_response] Raw output item [{i}] type: {item_type}") + if hasattr(item, "name") and hasattr(item, "arguments"): # For ResponseFunctionToolCall + logger.debug( + f"[_get_single_step_result_from_response] Raw ResponseFunctionToolCall item [{i}]: " + f"Name='{getattr(item, 'name', 'N/A')}', " + f"Args='{getattr(item, 'arguments', 'N/A')}', " + f"CallID='{getattr(item, 'call_id', 'N/A')}'" + ) + elif hasattr(item, "text"): # For ResponseOutputText + logger.debug(f"[_get_single_step_result_from_response] Raw ResponseOutputText item [{i}]: Text='{getattr(item, 'text', '')[:100]}...'") + + processed_response = RunImpl.process_model_response( agent=agent, all_tools=all_tools, @@ -786,6 +804,69 @@ class Runner: handoffs=handoffs, ) + # Log the processed response, focusing on tools_used + logger.debug(f"[_get_single_step_result_from_response] Processed response type: {type(processed_response).__name__}") + if hasattr(processed_response, 'is_final_output'): + logger.debug(f"[_get_single_step_result_from_response] Processed response: is_final_output={processed_response.is_final_output}") + else: + logger.debug(f"[_get_single_step_result_from_response] Processed response does not have is_final_output attribute") + + if hasattr(processed_response, 'final_output_from_llm') and processed_response.final_output_from_llm is not None: + logger.debug(f"[_get_single_step_result_from_response] Processed response: final_output_from_llm='{str(processed_response.final_output_from_llm)[:100]}...'") + + # Log tools used with robust type checking + if hasattr(processed_response, 'tools_used') and processed_response.tools_used: + # Log summarizing number of tools used first + logger.debug(f"[_get_single_step_result_from_response] Found {len(processed_response.tools_used)} tools used") + + # Add spacing between blocks in terminal output + if os.environ.get('CAI_STREAM', 'false').lower() == 'true': + print("") # Visual separator only when streaming is enabled + + for i, tool_call in enumerate(processed_response.tools_used): + try: + # Safely extract tool name with multiple fallbacks + tool_name = "Unknown" + try: + if hasattr(tool_call, 'tool'): + if isinstance(tool_call.tool, str): + tool_name = tool_call.tool + elif hasattr(tool_call.tool, 'name'): + tool_name = tool_call.tool.name + else: + tool_name = str(tool_call.tool) + except Exception: + pass + + # Safely extract call_id + call_id = "Unknown" + try: + if hasattr(tool_call, 'call_id'): + call_id = str(tool_call.call_id) + except Exception: + pass + + # Safely extract parsed_args + parsed_args = "Unknown" + try: + if hasattr(tool_call, 'parsed_args'): + parsed_args = str(tool_call.parsed_args) + except Exception: + pass + + logger.debug( + f"[_get_single_step_result_from_response] Processed tool_call [{i}]: " + f"Name='{tool_name}', " + f"CallID='{call_id}', " + f"ParsedArgs='{parsed_args}'" + ) + except Exception as e: + # Last resort fallback for any unexpected structure + logger.debug(f"[_get_single_step_result_from_response] Could not process tool_call [{i}]: {str(e)}") + else: + logger.debug("[_get_single_step_result_from_response] Processed response: No tools_used.") + + tool_use_tracker.add_tool_use(agent, processed_response.tools_used) return await RunImpl.execute_tools_and_side_effects( diff --git a/src/cai/tools/common.py b/src/cai/tools/common.py index e1e58778..95b0f223 100644 --- a/src/cai/tools/common.py +++ b/src/cai/tools/common.py @@ -222,18 +222,41 @@ def terminate_session(session_id): def _run_ctf(ctf, command, stdout=False, timeout=100, stream=False, call_id=None): try: - # Ensure the command is executed in a shell that supports command - # chaining - output = ctf.get_shell(command, timeout=timeout) - # exploit_logger.log_ok() - - if stdout: - print("\033[32m" + output + "\033[0m") - return output # output if output else result.stder + # Check the type of ctf object to handle various formats properly + if ctf is None: + return f"Error: CTF environment is None" + + # Handle string CTF values (like "CTF_ENV") + if isinstance(ctf, str): + # If in local mode, fallback to running local command + return _run_local(command, stdout, timeout, stream, call_id) + + # Handle dict format that might come from Ollama/Qwen models + if isinstance(ctf, dict): + # Check if this dict has a get_shell key or function + if callable(getattr(ctf, 'get_shell', None)): + # This is a proper CTF object + output = ctf.get_shell(command, timeout=timeout) + if stdout: + print("\033[32m" + output + "\033[0m") + return output + else: + # This is a dict but not a proper CTF object, fallback to local + return _run_local(command, stdout, timeout, stream, call_id) + + # Original code path for proper CTF objects + if hasattr(ctf, 'get_shell') and callable(ctf.get_shell): + output = ctf.get_shell(command, timeout=timeout) + if stdout: + print("\033[32m" + output + "\033[0m") + return output + + # Fallback for any other case + return _run_local(command, stdout, timeout, stream, call_id) except Exception as e: # pylint: disable=broad-except print(color(f"Error executing CTF command: {e}", fg="red")) - # exploit_logger.log_error(str(e)) - return f"Error executing CTF command: {str(e)}" + # Fallback to local execution + return _run_local(command, stdout, timeout, stream, call_id) def _run_local(command, stdout=False, timeout=100, stream=False, call_id=None, tool_name=None): @@ -328,7 +351,7 @@ def _run_local_streamed(command, call_id, timeout=100, tool_name=None): header.append(")", style="yellow") tool_time = 0 start_time = time.time() - total_time = time.time() - START_TIME + total_time = time.time() - START_TIME if START_TIME else 0 timing_info = [] if total_time: timing_info.append(f"Total: {format_time(total_time)}") @@ -339,15 +362,24 @@ def _run_local_streamed(command, call_id, timeout=100, tool_name=None): content = Text() + # Get console width and calculate appropriate panel width + console_width = console.width if hasattr(console, 'width') else 100 + panel_width = min(int(console_width * 0.95), console_width - 4) # Use 95% of width or leave 4 chars margin + + # Create the panel with a specific width to avoid overflow panel = Panel( Text.assemble(header, "\n\n", content), title="[bold green]Tool Execution[/bold green]", subtitle="[bold green]Live Output[/bold green]", border_style="green", padding=(1, 2), - box=ROUNDED + box=ROUNDED, + width=panel_width ) + # Print clear separator before panel to avoid overlap with previous content + console.print("\n\n") + # Start Live display with Live(panel, console=console, refresh_per_second=4) as live: # Stream stdout in real-time @@ -364,7 +396,7 @@ def _run_local_streamed(command, call_id, timeout=100, tool_name=None): # Update tool_time and header with new timing info tool_time = time.time() - start_time - total_time = time.time() - START_TIME + total_time = time.time() - START_TIME if START_TIME else 0 # Remove any previous timing info from header (rebuild header) timing_info = [] if total_time: @@ -386,7 +418,8 @@ def _run_local_streamed(command, call_id, timeout=100, tool_name=None): subtitle="[bold green]Live Output[/bold green]", border_style="green", padding=(1, 2), - box=ROUNDED + box=ROUNDED, + width=panel_width ) live.update(panel) # Check if process is done @@ -405,7 +438,8 @@ def _run_local_streamed(command, call_id, timeout=100, tool_name=None): subtitle="[bold green]Live Output[/bold green]", border_style="green", padding=(1, 2), - box=ROUNDED + box=ROUNDED, + width=panel_width ) live.update(panel) @@ -415,12 +449,16 @@ def _run_local_streamed(command, call_id, timeout=100, tool_name=None): title="[bold green]Tool Execution[/bold green]", border_style="green", padding=(1, 2), - box=ROUNDED + box=ROUNDED, + width=panel_width ) live.update(panel) # Wait a moment for the panel to be displayed properly time.sleep(0.5) + + # Print clear separator after panel to avoid overlap with next content + console.print("\n\n") else: # Fallback to simpler streaming with cli_print_tool_output # Parse command into command and args (same as rich mode) @@ -436,6 +474,9 @@ def _run_local_streamed(command, call_id, timeout=100, tool_name=None): tool_args["args"] = args # Note: Omitted empty values and async_mode=False as it's default + # Print separator to avoid overlap with previous content + print("\n") + # Initial notification - just once cli_print_tool_output(tool_name, tool_args, "Command started...", call_id=call_id) @@ -473,6 +514,9 @@ def _run_local_streamed(command, call_id, timeout=100, tool_name=None): final_output += f"\nCommand exited with code {return_code}" cli_print_tool_output(tool_name, tool_args, final_output, call_id=call_id) + + # Print separator to avoid overlap with next content + print("\n") # Return the full output return ''.join(output_buffer) @@ -547,7 +591,22 @@ def run_command(command, ctf=None, stdout=False, # pylint: disable=too-many-arg if stream and not call_id: call_id = str(uuid.uuid4())[:8] - # Otherwise, run command normally - if ctf: - return _run_ctf(ctf, command, stdout, timeout, stream, call_id) - return _run_local(command, stdout, timeout, stream, call_id, tool_name) + # Determine whether to use CTF or local execution + use_ctf = False + if ctf is not None: + # Check if ctf is a proper object that can handle shell commands + if (hasattr(ctf, 'get_shell') and callable(ctf.get_shell)) or isinstance(ctf, dict) or isinstance(ctf, str): + use_ctf = True + + # Run the command using the appropriate method + if use_ctf: + try: + # Try with CTF first + return _run_ctf(ctf, command, stdout, timeout, stream, call_id) + except Exception as e: # pylint: disable=broad-except + # Fallback to local if CTF fails + print(color(f"CTF execution failed, falling back to local: {str(e)}", fg="yellow")) + return _run_local(command, stdout, timeout, stream, call_id, tool_name) + else: + # Use local execution + return _run_local(command, stdout, timeout, stream, call_id, tool_name) From 42aed4ea7055880cd9e4c9164dc5ddab61dc714e Mon Sep 17 00:00:00 2001 From: luijait Date: Tue, 6 May 2025 16:23:10 +0200 Subject: [PATCH 02/10] Qwen executes tool calls --- src/cai/sdk/agents/run.py | 10 ---------- 1 file changed, 10 deletions(-) diff --git a/src/cai/sdk/agents/run.py b/src/cai/sdk/agents/run.py index a0937d28..3c45b72f 100644 --- a/src/cai/sdk/agents/run.py +++ b/src/cai/sdk/agents/run.py @@ -853,16 +853,6 @@ class Runner: parsed_args = str(tool_call.parsed_args) except Exception: pass - - logger.debug( - f"[_get_single_step_result_from_response] Processed tool_call [{i}]: " - f"Name='{tool_name}', " - f"CallID='{call_id}', " - f"ParsedArgs='{parsed_args}'" - ) - except Exception as e: - # Last resort fallback for any unexpected structure - logger.debug(f"[_get_single_step_result_from_response] Could not process tool_call [{i}]: {str(e)}") else: logger.debug("[_get_single_step_result_from_response] Processed response: No tools_used.") From 2c6fa34cd6dba2aaf2c0809dc7e1163fd968746e Mon Sep 17 00:00:00 2001 From: luijait Date: Tue, 6 May 2025 16:47:39 +0200 Subject: [PATCH 03/10] Qwen executes tool calls --- src/cai/sdk/agents/run.py | 5 ++--- 1 file changed, 2 insertions(+), 3 deletions(-) diff --git a/src/cai/sdk/agents/run.py b/src/cai/sdk/agents/run.py index 3c45b72f..30d82d2f 100644 --- a/src/cai/sdk/agents/run.py +++ b/src/cai/sdk/agents/run.py @@ -853,9 +853,8 @@ class Runner: parsed_args = str(tool_call.parsed_args) except Exception: pass - else: - logger.debug("[_get_single_step_result_from_response] Processed response: No tools_used.") - + except Exception: + pass tool_use_tracker.add_tool_use(agent, processed_response.tools_used) From a943938e493ad8b1fc0f94224b779d58d5ef2da7 Mon Sep 17 00:00:00 2001 From: luijait Date: Tue, 6 May 2025 17:11:18 +0200 Subject: [PATCH 04/10] generalice tool calls, remove ctf from arguments --- .../agents/models/openai_chatcompletions.py | 251 ++---------------- 1 file changed, 24 insertions(+), 227 deletions(-) diff --git a/src/cai/sdk/agents/models/openai_chatcompletions.py b/src/cai/sdk/agents/models/openai_chatcompletions.py index c5138d14..21c5ef81 100644 --- a/src/cai/sdk/agents/models/openai_chatcompletions.py +++ b/src/cai/sdk/agents/models/openai_chatcompletions.py @@ -809,154 +809,48 @@ class OpenAIChatCompletionsModel(Model): # Try to extract a JSON object from the content json_start = ollama_full_content.find('{') json_end = ollama_full_content.rfind('}') + 1 - - logger.debug(f"Ollama content length: {len(ollama_full_content)}, JSON start: {json_start}, JSON end: {json_end}") - + if json_start >= 0 and json_end > json_start: - json_str = ollama_full_content[json_start:json_end] - logger.debug(f"Extracted potential JSON: {json_str[:100]}...") - - # Special check for generic_linux_command format - if "generic_linux_command" in json_str and "arguments" in json_str: - logger.debug("Detected generic_linux_command pattern - using special handling") - - # Try regex pattern matching to handle potentially malformed JSON - cmd_pattern = re.search(r'"command"\s*:\s*"([^"]+)"', json_str) - args_pattern = re.search(r'"args"\s*:\s*"([^"]*)"', json_str) - ctf_pattern = re.search(r'"ctf"\s*:\s*"([^"]*)"', json_str) - - if cmd_pattern: - # Create a tool call specifically for generic_linux_command - command = cmd_pattern.group(1) - args = args_pattern.group(1) if args_pattern else "" - ctf = ctf_pattern.group(1) if ctf_pattern else "" - - tool_call_id = f"call_{hashlib.md5(('generic_linux_command' + str(time.time())).encode()).hexdigest()[:8]}" - - # Create a proper arguments object - command_args = { - "command": command, - "args": args, - "ctf": ctf, - "async_mode": False, - "session_id": "" - } - - arguments_str = json.dumps(command_args) - logger.debug(f"Created generic_linux_command with args: {arguments_str}") - - # Add it to our function_calls state - state.function_calls[0] = ResponseFunctionToolCall( - id=FAKE_RESPONSES_ID, - arguments=arguments_str, - name="generic_linux_command", - type="function_call", - call_id=tool_call_id, - ) - - # Display the tool call in CLI - from cai.util import cli_print_agent_messages - try: - # Create a message-like object to display the function call - tool_msg = type('ToolCallWrapper', (), { - 'content': None, - 'tool_calls': [ - type('ToolCallDetail', (), { - 'function': type('FunctionDetail', (), { - 'name': "generic_linux_command", - 'arguments': arguments_str - }), - 'id': tool_call_id, - 'type': 'function' - }) - ] - }) - - # Print the tool call using the CLI utility - cli_print_agent_messages( - agent_name=getattr(self, 'agent_name', 'Agent'), - message=tool_msg, - counter=getattr(self, 'interaction_counter', 0), - model=str(self.model), - debug=False, - interaction_input_tokens=estimated_input_tokens, - interaction_output_tokens=estimated_output_tokens, - interaction_reasoning_tokens=0, - total_input_tokens=getattr(self, 'total_input_tokens', 0) + estimated_input_tokens, - total_output_tokens=getattr(self, 'total_output_tokens', 0) + estimated_output_tokens, - total_reasoning_tokens=0, - interaction_cost=None, - total_cost=None - ) - except Exception as e: - logger.error(f"Error displaying tool call in CLI: {e}") - - # Add to message history - tool_call_msg = { - "role": "assistant", - "content": None, - "tool_calls": [ - { - "id": tool_call_id, - "type": "function", - "function": { - "name": "generic_linux_command", - "arguments": arguments_str - } - } - ] - } - - streamed_tool_calls.append(tool_call_msg) - add_to_message_history(tool_call_msg) - logger.debug(f"Added generic_linux_command with args: {arguments_str}") - - - # Standard JSON parsing for other function calls + json_str = ollama_full_content[json_start:json_end] + # Try to parse the JSON parsed = json.loads(json_str) # Check if it looks like a function call - if ('name' in parsed and 'arguments' in parsed and - isinstance(parsed['arguments'], dict)): - + if ('name' in parsed and 'arguments' in parsed): logger.debug(f"Found valid function call in Ollama output: {json_str}") - # Create a tool call from the JSON object + # Create a tool call ID tool_call_id = f"call_{hashlib.md5((parsed['name'] + str(time.time())).encode()).hexdigest()[:8]}" - # Add it to our function_calls state - state.function_calls[0] = ResponseFunctionToolCall( - id=FAKE_RESPONSES_ID, - arguments=json.dumps(parsed['arguments']), - name=parsed['name'], - type="function_call", - call_id=tool_call_id, - ) - # Ensure arguments is a valid JSON string arguments_str = "" if isinstance(parsed['arguments'], dict): + # Remove 'ctf' field if it exists + if 'ctf' in parsed['arguments']: + del parsed['arguments']['ctf'] arguments_str = json.dumps(parsed['arguments']) elif isinstance(parsed['arguments'], str): # If it's already a string, check if it's valid JSON try: - # Try parsing to validate - json.loads(parsed['arguments']) - arguments_str = parsed['arguments'] + # Try parsing to validate and remove 'ctf' if present + args_dict = json.loads(parsed['arguments']) + if isinstance(args_dict, dict) and 'ctf' in args_dict: + del args_dict['ctf'] + arguments_str = json.dumps(args_dict) except: # If not valid JSON, encode it as a JSON string arguments_str = json.dumps(parsed['arguments']) else: # For any other type, convert to string and then JSON - arguments_str = json.dumps(str(parsed['arguments'])) - - logger.debug(f"Final arguments string: {arguments_str}") - - # Update with the properly formatted arguments - state.function_calls[0].arguments = arguments_str - - # Log the tool call for development purposes - logger.debug(f"Adding tool call: {parsed['name']} with arguments: {arguments_str}") + arguments_str = json.dumps(str(parsed['arguments'])) + # Add it to our function_calls state + state.function_calls[0] = ResponseFunctionToolCall( + id=FAKE_RESPONSES_ID, + arguments=arguments_str, + name=parsed['name'], + type="function_call", + call_id=tool_call_id, + ) # Display the tool call in CLI from cai.util import cli_print_agent_messages @@ -1015,106 +909,9 @@ class OpenAIChatCompletionsModel(Model): streamed_tool_calls.append(tool_call_msg) add_to_message_history(tool_call_msg) - logger.debug(f"Added function call: {parsed['name']} with args: {json.dumps(parsed['arguments'])}") + logger.debug(f"Added function call: {parsed['name']} with args: {arguments_str}") except Exception as e: - logger.debug(f"Failed to parse potential Ollama function call: {e}") - - # Even if JSON parsing fails, try to extract generic_linux_command with regex - if "generic_linux_command" in ollama_full_content: - logger.debug("JSON parsing failed but detected generic_linux_command - trying regex fallback") - try: - # Use regex to extract command components even from malformed output - cmd_pattern = re.search(r'"command"\s*:\s*"([^"]+)"', ollama_full_content) - args_pattern = re.search(r'"args"\s*:\s*"([^"]*)"', ollama_full_content) - ctf_pattern = re.search(r'"ctf"\s*:\s*"([^"]*)"', ollama_full_content) - - if cmd_pattern: - # Create a tool call specifically for generic_linux_command - command = cmd_pattern.group(1) - args = args_pattern.group(1) if args_pattern else "" - ctf = ctf_pattern.group(1) if ctf_pattern else "" - - tool_call_id = f"call_{hashlib.md5(('generic_linux_command' + str(time.time())).encode()).hexdigest()[:8]}" - - # Create a proper arguments object - command_args = { - "command": command, - "args": args, - "ctf": ctf, - "async_mode": False, - "session_id": "" - } - - arguments_str = json.dumps(command_args) - logger.debug(f"Fallback created generic_linux_command with args: {arguments_str}") - - # Add it to our function_calls state - state.function_calls[0] = ResponseFunctionToolCall( - id=FAKE_RESPONSES_ID, - arguments=arguments_str, - name="generic_linux_command", - type="function_call", - call_id=tool_call_id, - ) - - # Display the tool call in CLI - from cai.util import cli_print_agent_messages - try: - # Create a message-like object for display - tool_msg = type('ToolCallWrapper', (), { - 'content': None, - 'tool_calls': [ - type('ToolCallDetail', (), { - 'function': type('FunctionDetail', (), { - 'name': "generic_linux_command", - 'arguments': arguments_str - }), - 'id': tool_call_id, - 'type': 'function' - }) - ] - }) - - cli_print_agent_messages( - agent_name=getattr(self, 'agent_name', 'Agent'), - message=tool_msg, - counter=getattr(self, 'interaction_counter', 0), - model=str(self.model), - debug=False, - interaction_input_tokens=estimated_input_tokens, - interaction_output_tokens=estimated_output_tokens, - interaction_reasoning_tokens=0, - total_input_tokens=getattr(self, 'total_input_tokens', 0) + estimated_input_tokens, - total_output_tokens=getattr(self, 'total_output_tokens', 0) + estimated_output_tokens, - total_reasoning_tokens=0, - interaction_cost=None, - total_cost=None - ) - except Exception as e: - logger.error(f"Error displaying fallback tool call in CLI: {e}") - - # Add to message history - tool_call_msg = { - "role": "assistant", - "content": None, - "tool_calls": [ - { - "id": tool_call_id, - "type": "function", - "function": { - "name": "generic_linux_command", - "arguments": arguments_str - } - } - ] - } - - streamed_tool_calls.append(tool_call_msg) - add_to_message_history(tool_call_msg) - - logger.debug(f"Added fallback generic_linux_command") - except Exception as regex_err: - logger.error(f"Regex fallback also failed: {regex_err}") + pass function_call_starting_index = 0 if state.text_content_index_and_output: From c378d7375b0e09aa62ba3d1a1c17970ccb430fb5 Mon Sep 17 00:00:00 2001 From: luijait Date: Tue, 6 May 2025 17:14:01 +0200 Subject: [PATCH 05/10] rebase common.py --- src/cai/tools/common.py | 601 +++++++++++++++++++++++++++++----------- 1 file changed, 439 insertions(+), 162 deletions(-) diff --git a/src/cai/tools/common.py b/src/cai/tools/common.py index 95b0f223..e475d5f4 100644 --- a/src/cai/tools/common.py +++ b/src/cai/tools/common.py @@ -24,14 +24,70 @@ except ImportError: # Global dictionary to store active sessions ACTIVE_SESSIONS = {} +def _get_workspace_dir() -> str: + """Determines the target workspace directory based on env vars for host.""" + base_dir_env = os.getenv("CAI_WORKSPACE_DIR") + workspace_name = os.getenv("CAI_WORKSPACE") + + # Determine the base directory + if base_dir_env: + base_dir = os.path.abspath(base_dir_env) + else: # Default base directory is 'workspaces' + if workspace_name: + base_dir = os.path.join(os.getcwd(), "workspaces") + else: # If no workspace name is set, the workspace IS the CWD. + return os.getcwd() + + # If a workspace name is provided, append it to the base directory + if workspace_name: + if not all(c.isalnum() or c in ['_', '-'] for c in workspace_name): + print(color(f"Invalid CAI_WORKSPACE name '{workspace_name}'. " + f"Using directory '{base_dir}' instead.", fg="yellow")) + target_dir = base_dir + else: + target_dir = os.path.join(base_dir, workspace_name) + else: + target_dir = base_dir + + # Ensure the final target directory exists on the host + try: + abs_target_dir = os.path.abspath(target_dir) + os.makedirs(abs_target_dir, exist_ok=True) + return abs_target_dir + except OSError as e: + print(color(f"Error creating/accessing host workspace directory '{abs_target_dir}': {e}", + fg="red")) + print(color(f"Falling back to current directory: {os.getcwd()}", fg="yellow")) + return os.getcwd() + +def _get_container_workspace_path() -> str: + """Determines the target workspace path inside the container.""" + workspace_name = os.getenv("CAI_WORKSPACE") + if workspace_name: + if not all(c.isalnum() or c in ['_', '-'] for c in workspace_name): + print(color(f"Invalid CAI_WORKSPACE name '{workspace_name}' for container. " + f"Using '/workspace'.", fg="yellow")) + return "/" + # Standard path inside CAI containers + return f"/workspace/workspaces/{workspace_name}" + else: + return "/" class ShellSession: # pylint: disable=too-many-instance-attributes """Class to manage interactive shell sessions""" - def __init__(self, command, session_id=None, ctf=None): + def __init__(self, command, session_id=None, ctf=None, workspace_dir=None, container_id=None): # noqa E501 self.session_id = session_id or str(uuid.uuid4())[:8] - self.command = command + self.command = command self.ctf = ctf + self.container_id = container_id + # Determine workspace based on context (container, ctf or local host) + if self.container_id: + self.workspace_dir = _get_container_workspace_path() + elif self.ctf: + self.workspace_dir = workspace_dir or _get_workspace_dir() + else: + self.workspace_dir = _get_workspace_dir() self.process = None self.master = None self.slave = None @@ -40,9 +96,40 @@ class ShellSession: # pylint: disable=too-many-instance-attributes self.last_activity = time.time() def start(self): - """Start the shell session""" + """Start the shell session in the appropriate environment.""" + start_message_cmd = self.command + + # --- Start in Container --- + if self.container_id: + try: + self.master, self.slave = pty.openpty() + docker_cmd_list = [ + "docker", "exec", "-i", + "-w", self.workspace_dir, + self.container_id, + "sh", "-c", # Use shell to handle complex commands if needed + self.command # The actual command to run + ] + self.process = subprocess.Popen( + docker_cmd_list, + stdin=self.slave, + stdout=self.slave, + stderr=self.slave, + preexec_fn=os.setsid, + universal_newlines=True + ) + self.is_running = True + self.output_buffer.append( + f"[Session {self.session_id}] Started in container {self.container_id[:12]}: " + f"{start_message_cmd} in {self.workspace_dir}") + threading.Thread(target=self._read_output, daemon=True).start() + except Exception as e: + self.output_buffer.append(f"Error starting container session: {str(e)}") + self.is_running = False + return + + # --- Start in CTF --- if self.ctf: - # For CTF environments self.is_running = True self.output_buffer.append( f"[Session { @@ -52,81 +139,100 @@ class ShellSession: # pylint: disable=too-many-instance-attributes output = self.ctf.get_shell(self.command) self.output_buffer.append(output) except Exception as e: # pylint: disable=broad-except - self.output_buffer.append(f"Error: {str(e)}") - self.is_running = False + self.output_buffer.append(f"Error executing CTF command: {str(e)}") + self.is_running = False return - # For local environment + # --- Start Locally (Host) --- try: - # Create a pseudo-terminal self.master, self.slave = pty.openpty() - - # Start the process self.process = subprocess.Popen( # pylint: disable=subprocess-popen-preexec-fn, consider-using-with # noqa: E501 - self.command, - shell=True, # nosec B602 + self.command, + shell=True, # nosec B602 stdin=self.slave, stdout=self.slave, stderr=self.slave, - preexec_fn=os.setsid, # Create a new process group + cwd=self.workspace_dir, + preexec_fn=os.setsid, universal_newlines=True ) - self.is_running = True self.output_buffer.append( f"[Session { self.session_id}] Started: { self.command}") - # Start a thread to read output threading.Thread(target=self._read_output, daemon=True).start() except Exception as e: # pylint: disable=broad-except - self.output_buffer.append(f"Error starting session: {str(e)}") + self.output_buffer.append(f"Error starting local session: {str(e)}") self.is_running = False def _read_output(self): """Read output from the process""" try: - while self.is_running: + while self.is_running and self.master is not None: try: + # Check if process has exited before reading + if self.process and self.process.poll() is not None: + self.is_running = False + break + # Read the output output = os.read(self.master, 1024).decode() if output: self.output_buffer.append(output) self.last_activity = time.time() - except OSError: - # No data available or terminal closed - time.sleep(0.1) - if not self.is_process_running(): + else: self.is_running = False break - except Exception as e: # pylint: disable=broad-except - self.output_buffer.append(f"Error reading output: {str(e)}") + except Exception as e: + self.output_buffer.append(f"Error reading output buffer: {str(read_err)}") + self.is_running = False + break + # Add a small sleep to prevent busy-waiting if no output + if is_process_running(self): + time.sleep(0.05) + except Exception as e: + self.output_buffer.append(f"Error in read_output loop: {str(e)}") self.is_running = False + def is_process_running(self): """Check if the process is still running""" + # For CTF or container + if self.container_id or self.ctf: + return self.is_running + # For local host if not self.process: return False return self.process.poll() is None def send_input(self, input_data): - """Send input to the process""" - if not self.is_running: - return "Session is not running" + """Send input to the process (local or container)""" + if not self.is_running: # For CTF or container + if self.process and self.process.poll() is None: + self.is_running = True + else: # For local host + return "Session is not running" try: + # --- Send to CTF --- if self.ctf: - # For CTF environments output = self.ctf.get_shell(input_data) self.output_buffer.append(output) return "Input sent to CTF session" - # For local environment - input_data = input_data.rstrip() + "\n" - os.write(self.master, input_data.encode()) - self.last_activity = time.time() - return "Input sent to session" + # --- Send to Local or Container PTY --- + if self.master is not None: + input_data_bytes = (input_data.rstrip() + "\n").encode() + bytes_written = os.write(self.master, input_data_bytes) + if bytes_written != len(input_data_bytes): + self.output_buffer.append(f"[Session {self.session_id}] Warning: Partial input write.") + self.last_activity = time.time() + return "Input sent to session" + else: + return "Session PTY not available for input" except Exception as e: # pylint: disable=broad-except + self.output_buffer.append(f"Error sending input: {str(e)}") return f"Error sending input: {str(e)}" def get_output(self, clear=True): @@ -138,8 +244,12 @@ class ShellSession: # pylint: disable=too-many-instance-attributes def terminate(self): """Terminate the session""" + session_id_short = self.session_id[:8] if not self.is_running: - return "Session already terminated" + if self.process and self.process.poll() is None: + pass # Process is running, proceed with termination + else: + return f"Session {session_id_short} already terminated or finished." try: self.is_running = False @@ -147,28 +257,64 @@ class ShellSession: # pylint: disable=too-many-instance-attributes if self.process: # Try to terminate the process group try: - os.killpg(os.getpgid(self.process.pid), signal.SIGTERM) - except BaseException: # pylint: disable=bare-except,broad-except # noqa: E501 - # If that fails, try to terminate just the process - self.process.terminate() + os.killpg(os.getpgid(self.process.pid), signal.SIGTERM) + except ProcessLookupError: + pass # Process already gone + except subprocess.TimeoutExpired: + print(color(f"Session {session_id_short} did not terminate gracefully, sending SIGKILL...", fg="yellow")) # noqa E501 + try: + if pgid: + os.killpg(pgid, signal.SIGKILL) # Force kill + else: + self.process.kill() + except ProcessLookupError: + pass # Already gone + except Exception as kill_err: + termination_message = f" (Error during SIGKILL: {kill_err})" + except Exception as term_err: # Catch other errors during SIGTERM + termination_message = f" (Error during SIGTERM: {term_err})" + try: + self.process.kill() + except Exception: pass # Ignore nested errors - # Clean up resources - if self.master: - os.close(self.master) - if self.slave: - os.close(self.slave) - return f"Session {self.session_id} terminated" + # Final check + if self.process.poll() is None: + print(color(f"Session {session_id_short} process {self.process.pid} may still be running after termination attempts.", fg="red")) # noqa E501 + termination_message += " (Warning: Process may still be running)" + + + # Clean up PTY resources if they exist + if self.master: + try: os.close(self.master) + except OSError: pass + self.master = None + if self.slave: + try: os.close(self.slave) + except OSError: pass + self.slave = None + + return termination_message or f"Session {self.session_id} terminated" except Exception as e: # pylint: disable=broad-except - return f"Error terminating session: {str(e)}" + return f"Error terminating session {session_id_short}: {str(e)}" -def create_shell_session(command, ctf=None): - """Create a new shell session""" - session = ShellSession(command, ctf=ctf) +def create_shell_session(command, ctf=None, container_id=None, **kwargs): + """Create a new shell session in the correct workspace/environment.""" + if container_id: + session = ShellSession(command, ctf=ctf, container_id=container_id) + else: + workspace_dir = _get_workspace_dir() + session = ShellSession(command, ctf=ctf, workspace_dir=workspace_dir) + session.start() - ACTIVE_SESSIONS[session.session_id] = session - return session.session_id + if session.is_running or (ctf and not session.is_running): + ACTIVE_SESSIONS[session.session_id] = session + return session.session_id + else: + error_msg = session.get_output(clear=True) + print(color(f"Failed to start session: {error_msg}", fg="red")) + return f"Failed to start session: {error_msg}" def list_shell_sessions(): @@ -212,70 +358,100 @@ def get_session_output(session_id, clear=True): def terminate_session(session_id): """Terminate a specific session""" if session_id not in ACTIVE_SESSIONS: - return f"Session {session_id} not found" + return f"Session {session_id} not found or already terminated." session = ACTIVE_SESSIONS[session_id] result = session.terminate() - del ACTIVE_SESSIONS[session_id] + if session_id in ACTIVE_SESSIONS: + del ACTIVE_SESSIONS[session_id] return result -def _run_ctf(ctf, command, stdout=False, timeout=100, stream=False, call_id=None): +def _run_ctf(ctf, command, stdout=False, timeout=100, workspace_dir=None): + """Runs command in CTF env, changing to workspace_dir first.""" + target_dir = workspace_dir or _get_workspace_dir() + full_command = f"cd '{target_dir}' && {command}" + original_cmd_for_msg = command # For logging + context_msg = f"(ctf:{target_dir})" try: - # Check the type of ctf object to handle various formats properly - if ctf is None: - return f"Error: CTF environment is None" - - # Handle string CTF values (like "CTF_ENV") - if isinstance(ctf, str): - # If in local mode, fallback to running local command - return _run_local(command, stdout, timeout, stream, call_id) - - # Handle dict format that might come from Ollama/Qwen models - if isinstance(ctf, dict): - # Check if this dict has a get_shell key or function - if callable(getattr(ctf, 'get_shell', None)): - # This is a proper CTF object - output = ctf.get_shell(command, timeout=timeout) - if stdout: - print("\033[32m" + output + "\033[0m") - return output - else: - # This is a dict but not a proper CTF object, fallback to local - return _run_local(command, stdout, timeout, stream, call_id) - - # Original code path for proper CTF objects - if hasattr(ctf, 'get_shell') and callable(ctf.get_shell): - output = ctf.get_shell(command, timeout=timeout) - if stdout: - print("\033[32m" + output + "\033[0m") - return output - - # Fallback for any other case - return _run_local(command, stdout, timeout, stream, call_id) + output = ctf.get_shell(full_command, timeout=timeout) + if stdout: + print(f"\033[32m{context_msg} $ {original_cmd_for_msg}\n{output}\033[0m") # noqa E501 + return output except Exception as e: # pylint: disable=broad-except - print(color(f"Error executing CTF command: {e}", fg="red")) - # Fallback to local execution - return _run_local(command, stdout, timeout, stream, call_id) + error_msg = f"Error executing CTF command '{original_cmd_for_msg}' in '{target_dir}': {e}" # noqa E501 + print(color(error_msg, fg="red")) + return error_msg + +def _run_ssh(command, stdout=False, timeout=100, workspace_dir=None): + """Runs command via SSH. Assumes SSH agent or passwordless setup unless sshpass is used externally.""" # noqa E501 + ssh_user = os.environ.get('SSH_USER') + ssh_host = os.environ.get('SSH_HOST') + ssh_pass = os.environ.get('SSH_PASS') + remote_command = command + original_cmd_for_msg = command + context_msg = f"({ssh_user}@{ssh_host})" + + # Construct base SSH command list + if ssh_pass: + ssh_cmd_list = ["sshpass", "-p", ssh_pass, "ssh", f"{ssh_user}@{ssh_host}"] # noqa E501 + else: + ssh_cmd_list = ["ssh", f"{ssh_user}@{ssh_host}"] + ssh_cmd_list.append(remote_command) + + try: + # Use subprocess.run with list of args for better security than shell=True + result = subprocess.run( + ssh_cmd_list, + capture_output=True, + text=True, + check=False, # Don't raise exception on non-zero exit code + timeout=timeout + ) + output = result.stdout if result.stdout else result.stderr + if stdout: + print(f"\033[32m{context_msg} $ {original_cmd_for_msg}\n{output}\033[0m") # noqa E501 + # Return combined output, potentially including errors + return output.strip() + except subprocess.TimeoutExpired as e: + error_output = e.stdout if e.stdout else str(e) + timeout_msg = f"Timeout executing SSH command: {error_output}" + if stdout: + print(f"\033[33m{context_msg} $ {original_cmd_for_msg}\nTIMEOUT\n{error_output}\033[0m") # noqa E501 + return timeout_msg + except FileNotFoundError: + # Handle case where ssh or sshpass isn't installed + error_msg = f"'sshpass' or 'ssh' command not found. Ensure they are installed and in PATH." # noqa E501 + print(color(error_msg, fg="red")) + return error_msg + except Exception as e: # pylint: disable=broad-except + error_msg = f"Error executing SSH command '{original_cmd_for_msg}' on {ssh_host}: {e}" # noqa E501 + print(color(error_msg, fg="red")) + return error_msg -def _run_local(command, stdout=False, timeout=100, stream=False, call_id=None, tool_name=None): +def _run_local(command, stdout=False, timeout=100, stream=False, call_id=None, tool_name=None, workspace_dir=None): + """Runs command locally in the specified workspace_dir.""" # If streaming is enabled and we have a call_id if stream and call_id: - return _run_local_streamed(command, call_id, timeout, tool_name) + return _run_local_streamed(command, call_id, timeout, tool_name, workspace_dir) + target_dir = workspace_dir or _get_workspace_dir() + original_cmd_for_msg = command # For logging + context_msg = f"(local:{target_dir})" try: - # nosec B602 - shell=True is required for command chaining result = subprocess.run( command, shell=True, # nosec B602 capture_output=True, text=True, - check=False, - timeout=timeout) + check=False, + timeout=timeout, + cwd=target_dir + ) output = result.stdout if result.stdout else result.stderr if stdout: - print("\033[32m" + output + "\033[0m") + print(f"\033[32m{context_msg} $ {original_cmd_for_msg}\n{output}\033[0m") # noqa E501 # Skip passing output to cli_print_tool_output when CAI_STREAM=true # This prevents duplicate output in streaming mode @@ -284,20 +460,21 @@ def _run_local(command, stdout=False, timeout=100, stream=False, call_id=None, t # Optional: Add cli_print_tool_output call here if needed for non-streaming pass - return output + return output.strip() except subprocess.TimeoutExpired as e: - error_output = e.stdout.decode() if e.stdout else str(e) + error_output = e.stdout if e.stdout else str(e) if stdout: print("\033[32m" + error_output + "\033[0m") - return error_output + return error_output except Exception as e: # pylint: disable=broad-except - error_msg = f"Error executing local command: {e}" - print(color(error_msg, fg="red")) - return error_msg + error_msg = f"Error executing local command: {e}" + print(color(error_msg, fg="red")) + return error_msg -def _run_local_streamed(command, call_id, timeout=100, tool_name=None): - """Run a local command with streaming output to the Tool output panel""" +def _run_local_streamed(command, call_id, timeout=100, tool_name=None, workspace_dir=None): + """Run a local command with streaming output to the Tool output panel.""" + target_dir = workspace_dir or _get_workspace_dir() try: # Try to import Rich for nice display try: @@ -321,7 +498,8 @@ def _run_local_streamed(command, call_id, timeout=100, tool_name=None): stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True, - bufsize=1 + bufsize=1, + cwd=target_dir # Set CWD for local process ) # If tool_name is not provided, derive it from the command @@ -351,7 +529,7 @@ def _run_local_streamed(command, call_id, timeout=100, tool_name=None): header.append(")", style="yellow") tool_time = 0 start_time = time.time() - total_time = time.time() - START_TIME if START_TIME else 0 + total_time = time.time() - START_TIME timing_info = [] if total_time: timing_info.append(f"Total: {format_time(total_time)}") @@ -362,24 +540,15 @@ def _run_local_streamed(command, call_id, timeout=100, tool_name=None): content = Text() - # Get console width and calculate appropriate panel width - console_width = console.width if hasattr(console, 'width') else 100 - panel_width = min(int(console_width * 0.95), console_width - 4) # Use 95% of width or leave 4 chars margin - - # Create the panel with a specific width to avoid overflow panel = Panel( Text.assemble(header, "\n\n", content), title="[bold green]Tool Execution[/bold green]", subtitle="[bold green]Live Output[/bold green]", border_style="green", padding=(1, 2), - box=ROUNDED, - width=panel_width + box=ROUNDED ) - # Print clear separator before panel to avoid overlap with previous content - console.print("\n\n") - # Start Live display with Live(panel, console=console, refresh_per_second=4) as live: # Stream stdout in real-time @@ -396,7 +565,7 @@ def _run_local_streamed(command, call_id, timeout=100, tool_name=None): # Update tool_time and header with new timing info tool_time = time.time() - start_time - total_time = time.time() - START_TIME if START_TIME else 0 + total_time = time.time() - START_TIME # Remove any previous timing info from header (rebuild header) timing_info = [] if total_time: @@ -418,8 +587,7 @@ def _run_local_streamed(command, call_id, timeout=100, tool_name=None): subtitle="[bold green]Live Output[/bold green]", border_style="green", padding=(1, 2), - box=ROUNDED, - width=panel_width + box=ROUNDED ) live.update(panel) # Check if process is done @@ -438,8 +606,7 @@ def _run_local_streamed(command, call_id, timeout=100, tool_name=None): subtitle="[bold green]Live Output[/bold green]", border_style="green", padding=(1, 2), - box=ROUNDED, - width=panel_width + box=ROUNDED ) live.update(panel) @@ -449,16 +616,12 @@ def _run_local_streamed(command, call_id, timeout=100, tool_name=None): title="[bold green]Tool Execution[/bold green]", border_style="green", padding=(1, 2), - box=ROUNDED, - width=panel_width + box=ROUNDED ) live.update(panel) # Wait a moment for the panel to be displayed properly time.sleep(0.5) - - # Print clear separator after panel to avoid overlap with next content - console.print("\n\n") else: # Fallback to simpler streaming with cli_print_tool_output # Parse command into command and args (same as rich mode) @@ -474,9 +637,6 @@ def _run_local_streamed(command, call_id, timeout=100, tool_name=None): tool_args["args"] = args # Note: Omitted empty values and async_mode=False as it's default - # Print separator to avoid overlap with previous content - print("\n") - # Initial notification - just once cli_print_tool_output(tool_name, tool_args, "Command started...", call_id=call_id) @@ -514,9 +674,6 @@ def _run_local_streamed(command, call_id, timeout=100, tool_name=None): final_output += f"\nCommand exited with code {return_code}" cli_print_tool_output(tool_name, tool_args, final_output, call_id=call_id) - - # Print separator to avoid overlap with next content - print("\n") # Return the full output return ''.join(output_buffer) @@ -549,7 +706,8 @@ def run_command(command, ctf=None, stdout=False, # pylint: disable=too-many-arg async_mode=False, session_id=None, timeout=100, stream=False, call_id=None, tool_name=None): """ - Run command either in CTF container or on the local attacker machine + Run command in the appropriate environment (Docker, CTF, SSH, Local) + and workspace. Args: command: The command to execute @@ -564,49 +722,168 @@ def run_command(command, ctf=None, stdout=False, # pylint: disable=too-many-arg If None, the tool name will be derived from the command. Returns: - str: Command output, status message, or session ID + str: Command output, status message, or session ID. """ # If session_id is provided, send command to that session if session_id: if session_id not in ACTIVE_SESSIONS: return f"Session {session_id} not found" - - result = send_to_session(session_id, command) + session = ACTIVE_SESSIONS[session_id] + result = session.send_input(command) # Send the raw command string if stdout: output = get_session_output(session_id, clear=False) - print("\033[32m" + output + "\033[0m") - return result - - # If async_mode, create a new session - if async_mode: - session_id = create_shell_session(command, ctf) - if stdout: - # Wait a moment for initial output - time.sleep(0.5) - output = get_session_output(session_id, clear=False) - print("\033[32m" + output + "\033[0m") - return f"Created session {session_id}. Use this ID to interact with the session." + env_type = "Local" + if session.container_id: + env_type = f"Container({session.container_id[:12]})" + elif session.ctf: + env_type = "CTF" + print(f"\033[32m(Session {session_id} in {env_type}:{session.workspace_dir}) >> {command}\n{output}\033[0m") # noqa E501 + return result # Return the result of sending input ("Input sent..." or error) # Generate a call_id if we're streaming and one wasn't provided if stream and not call_id: call_id = str(uuid.uuid4())[:8] - - # Determine whether to use CTF or local execution - use_ctf = False - if ctf is not None: - # Check if ctf is a proper object that can handle shell commands - if (hasattr(ctf, 'get_shell') and callable(ctf.get_shell)) or isinstance(ctf, dict) or isinstance(ctf, str): - use_ctf = True - - # Run the command using the appropriate method - if use_ctf: + + # 2. Determine Execution Environment (Container > CTF > SSH > Local) + active_container = os.getenv("CAI_ACTIVE_CONTAINER", "") + is_ssh_env = all(os.getenv(var) for var in ['SSH_USER', 'SSH_HOST']) + + # --- Docker Container Execution --- + if active_container and not ctf and not is_ssh_env: + container_id = active_container + container_workspace = _get_container_workspace_path() + context_msg = f"(docker:{container_id[:12]}:{container_workspace})" + + # Handle Async Session Creation in Container + if async_mode: + # Create a session specifically for the container environment + new_session_id = create_shell_session(command, container_id=container_id) # noqa E501 + if "Failed" in new_session_id: # Check if session creation failed + return new_session_id + if stdout: + # Wait a moment for initial output + time.sleep(0.2) + output = get_session_output(new_session_id, clear=False) + print(f"\033[32m(Started Session {new_session_id} in {context_msg})\n{output}\033[0m") # noqa E501 + return f"Started async session {new_session_id} in container {container_id[:12]}. Use this ID to interact." # noqa E501 + + # Handle Streaming Container Execution - not yet implemented for containers + if stream: + # For now, display that streaming isn't supported for containers + from cai.util import cli_print_tool_output + if call_id and tool_name: + tool_args = {"command": command, "container": container_id[:12]} + cli_print_tool_output( + tool_name, + tool_args, + "Streaming not yet supported for container execution. Running normally...", + call_id=call_id + ) + + # Handle Synchronous Execution in Container try: - # Try with CTF first - return _run_ctf(ctf, command, stdout, timeout, stream, call_id) + # Ensure container workspace exists (best effort) + # Consider moving this to workspace set/container activation + mkdir_cmd = ["docker", "exec", container_id, "mkdir", "-p", container_workspace] # noqa E501 + subprocess.run(mkdir_cmd, capture_output=True, text=True, check=False, timeout=10) # noqa E501 + + # Construct the docker exec command with workspace context + cmd_list = [ + "docker", "exec", + "-w", container_workspace, # Set working directory + container_id, + "sh", "-c", command # Execute command via shell + ] + result = subprocess.run( + cmd_list, + capture_output=True, + text=True, + check=False, # Don't raise exception on non-zero exit + timeout=timeout + ) + + output = result.stdout if result.stdout else result.stderr + output = output.strip() # Clean trailing newline + + if stdout: + print(f"\033[32m{context_msg} $ {command}\n{output}\033[0m") # noqa E501 + + # Check if command failed specifically because container isn't running + if result.returncode != 0 and "is not running" in result.stderr: + print(color(f"{context_msg} Container is not running. Attempting execution on host instead.", fg="yellow")) # noqa E501 + # Fallback to local execution, preserving workspace context + return _run_local(command, stdout, timeout, stream, call_id, tool_name, _get_workspace_dir()) # noqa E501 + + return output # Return combined stdout/stderr + + except subprocess.TimeoutExpired: + timeout_msg = "Timeout executing command in container." + if stdout: + print(f"\033[33m{context_msg} $ {command}\nTIMEOUT\033[0m") # noqa E501 + print(color("Attempting execution on host instead.", fg="yellow")) + # Fallback to local execution on timeout + return _run_local(command, stdout, timeout, stream, call_id, tool_name, _get_workspace_dir()) # noqa E501 except Exception as e: # pylint: disable=broad-except - # Fallback to local if CTF fails - print(color(f"CTF execution failed, falling back to local: {str(e)}", fg="yellow")) - return _run_local(command, stdout, timeout, stream, call_id, tool_name) - else: - # Use local execution - return _run_local(command, stdout, timeout, stream, call_id, tool_name) + error_msg = f"Error executing command in container: {str(e)}" + print(color(f"{context_msg} {error_msg}", fg="red")) + print(color("Attempting execution on host instead.", fg="yellow")) + # Fallback to local execution on other errors + return _run_local(command, stdout, timeout, stream, call_id, tool_name, _get_workspace_dir()) # noqa E501 + + # --- CTF Execution --- + if ctf: + # Handling streaming for CTF - not fully implemented yet + if stream: + from cai.util import cli_print_tool_output + if call_id and tool_name: + tool_args = {"command": command, "ctf": True} + cli_print_tool_output( + tool_name, + tool_args, + "Streaming not yet supported for CTF execution. Running normally...", + call_id=call_id + ) + + # _run_ctf handles workspace internally using _get_workspace_dir() default + return _run_ctf(ctf, command, stdout, timeout) # Pass None for workspace_dir + + # --- SSH Execution --- + if is_ssh_env: + # Async for SSH would require session management via SSH client features + if async_mode: + return "Async mode not fully supported for SSH environment via this function yet." + + # Handling streaming for SSH - not fully implemented yet + if stream: + from cai.util import cli_print_tool_output + if call_id and tool_name: + tool_args = {"command": command, "ssh": True} + cli_print_tool_output( + tool_name, + tool_args, + "Streaming not yet supported for SSH execution. Running normally...", + call_id=call_id + ) + + # _run_ssh handles command execution, workspace is relative to remote home + return _run_ssh(command, stdout, timeout) # Workspace dir less relevant here + + # --- Local Execution (Default Fallback) --- + # Let _run_local handle determining the host workspace + # Handle Async Session Creation Locally + if async_mode: + # create_shell_session uses _get_workspace_dir() when container_id is None + new_session_id = create_shell_session(command) + if isinstance(new_session_id, str) and "Failed" in new_session_id: # Check failure + return new_session_id + # Retrieve the actual workspace dir the session is using + session = ACTIVE_SESSIONS.get(new_session_id) + actual_workspace = session.workspace_dir if session else "unknown" + if stdout: + time.sleep(0.2) # Allow session buffer to populate + output = get_session_output(new_session_id, clear=False) + print(f"\033[32m(Started Session {new_session_id} in local:{actual_workspace})\n{output}\033[0m") + return f"Started async session {new_session_id} locally. Use this ID to interact." + + # Handle Synchronous Execution Locally using _run_local default with streaming support + return _run_local(command, stdout, timeout, stream, call_id, tool_name, None) From 069ec7f9b1d97d8306c1227d10990ef53e14432e Mon Sep 17 00:00:00 2001 From: luijait Date: Tue, 6 May 2025 17:16:21 +0200 Subject: [PATCH 06/10] rebase common.py --- src/cai/tools/common.py | 532 ++++++++-------------------------------- 1 file changed, 98 insertions(+), 434 deletions(-) diff --git a/src/cai/tools/common.py b/src/cai/tools/common.py index e475d5f4..e1e58778 100644 --- a/src/cai/tools/common.py +++ b/src/cai/tools/common.py @@ -24,70 +24,14 @@ except ImportError: # Global dictionary to store active sessions ACTIVE_SESSIONS = {} -def _get_workspace_dir() -> str: - """Determines the target workspace directory based on env vars for host.""" - base_dir_env = os.getenv("CAI_WORKSPACE_DIR") - workspace_name = os.getenv("CAI_WORKSPACE") - - # Determine the base directory - if base_dir_env: - base_dir = os.path.abspath(base_dir_env) - else: # Default base directory is 'workspaces' - if workspace_name: - base_dir = os.path.join(os.getcwd(), "workspaces") - else: # If no workspace name is set, the workspace IS the CWD. - return os.getcwd() - - # If a workspace name is provided, append it to the base directory - if workspace_name: - if not all(c.isalnum() or c in ['_', '-'] for c in workspace_name): - print(color(f"Invalid CAI_WORKSPACE name '{workspace_name}'. " - f"Using directory '{base_dir}' instead.", fg="yellow")) - target_dir = base_dir - else: - target_dir = os.path.join(base_dir, workspace_name) - else: - target_dir = base_dir - - # Ensure the final target directory exists on the host - try: - abs_target_dir = os.path.abspath(target_dir) - os.makedirs(abs_target_dir, exist_ok=True) - return abs_target_dir - except OSError as e: - print(color(f"Error creating/accessing host workspace directory '{abs_target_dir}': {e}", - fg="red")) - print(color(f"Falling back to current directory: {os.getcwd()}", fg="yellow")) - return os.getcwd() - -def _get_container_workspace_path() -> str: - """Determines the target workspace path inside the container.""" - workspace_name = os.getenv("CAI_WORKSPACE") - if workspace_name: - if not all(c.isalnum() or c in ['_', '-'] for c in workspace_name): - print(color(f"Invalid CAI_WORKSPACE name '{workspace_name}' for container. " - f"Using '/workspace'.", fg="yellow")) - return "/" - # Standard path inside CAI containers - return f"/workspace/workspaces/{workspace_name}" - else: - return "/" class ShellSession: # pylint: disable=too-many-instance-attributes """Class to manage interactive shell sessions""" - def __init__(self, command, session_id=None, ctf=None, workspace_dir=None, container_id=None): # noqa E501 + def __init__(self, command, session_id=None, ctf=None): self.session_id = session_id or str(uuid.uuid4())[:8] - self.command = command + self.command = command self.ctf = ctf - self.container_id = container_id - # Determine workspace based on context (container, ctf or local host) - if self.container_id: - self.workspace_dir = _get_container_workspace_path() - elif self.ctf: - self.workspace_dir = workspace_dir or _get_workspace_dir() - else: - self.workspace_dir = _get_workspace_dir() self.process = None self.master = None self.slave = None @@ -96,40 +40,9 @@ class ShellSession: # pylint: disable=too-many-instance-attributes self.last_activity = time.time() def start(self): - """Start the shell session in the appropriate environment.""" - start_message_cmd = self.command - - # --- Start in Container --- - if self.container_id: - try: - self.master, self.slave = pty.openpty() - docker_cmd_list = [ - "docker", "exec", "-i", - "-w", self.workspace_dir, - self.container_id, - "sh", "-c", # Use shell to handle complex commands if needed - self.command # The actual command to run - ] - self.process = subprocess.Popen( - docker_cmd_list, - stdin=self.slave, - stdout=self.slave, - stderr=self.slave, - preexec_fn=os.setsid, - universal_newlines=True - ) - self.is_running = True - self.output_buffer.append( - f"[Session {self.session_id}] Started in container {self.container_id[:12]}: " - f"{start_message_cmd} in {self.workspace_dir}") - threading.Thread(target=self._read_output, daemon=True).start() - except Exception as e: - self.output_buffer.append(f"Error starting container session: {str(e)}") - self.is_running = False - return - - # --- Start in CTF --- + """Start the shell session""" if self.ctf: + # For CTF environments self.is_running = True self.output_buffer.append( f"[Session { @@ -139,100 +52,81 @@ class ShellSession: # pylint: disable=too-many-instance-attributes output = self.ctf.get_shell(self.command) self.output_buffer.append(output) except Exception as e: # pylint: disable=broad-except - self.output_buffer.append(f"Error executing CTF command: {str(e)}") - self.is_running = False + self.output_buffer.append(f"Error: {str(e)}") + self.is_running = False return - # --- Start Locally (Host) --- + # For local environment try: + # Create a pseudo-terminal self.master, self.slave = pty.openpty() + + # Start the process self.process = subprocess.Popen( # pylint: disable=subprocess-popen-preexec-fn, consider-using-with # noqa: E501 - self.command, - shell=True, # nosec B602 + self.command, + shell=True, # nosec B602 stdin=self.slave, stdout=self.slave, stderr=self.slave, - cwd=self.workspace_dir, - preexec_fn=os.setsid, + preexec_fn=os.setsid, # Create a new process group universal_newlines=True ) + self.is_running = True self.output_buffer.append( f"[Session { self.session_id}] Started: { self.command}") + # Start a thread to read output threading.Thread(target=self._read_output, daemon=True).start() except Exception as e: # pylint: disable=broad-except - self.output_buffer.append(f"Error starting local session: {str(e)}") + self.output_buffer.append(f"Error starting session: {str(e)}") self.is_running = False def _read_output(self): """Read output from the process""" try: - while self.is_running and self.master is not None: + while self.is_running: try: - # Check if process has exited before reading - if self.process and self.process.poll() is not None: - self.is_running = False - break - # Read the output output = os.read(self.master, 1024).decode() if output: self.output_buffer.append(output) self.last_activity = time.time() - else: + except OSError: + # No data available or terminal closed + time.sleep(0.1) + if not self.is_process_running(): self.is_running = False break - except Exception as e: - self.output_buffer.append(f"Error reading output buffer: {str(read_err)}") - self.is_running = False - break - # Add a small sleep to prevent busy-waiting if no output - if is_process_running(self): - time.sleep(0.05) - except Exception as e: - self.output_buffer.append(f"Error in read_output loop: {str(e)}") + except Exception as e: # pylint: disable=broad-except + self.output_buffer.append(f"Error reading output: {str(e)}") self.is_running = False - def is_process_running(self): """Check if the process is still running""" - # For CTF or container - if self.container_id or self.ctf: - return self.is_running - # For local host if not self.process: return False return self.process.poll() is None def send_input(self, input_data): - """Send input to the process (local or container)""" - if not self.is_running: # For CTF or container - if self.process and self.process.poll() is None: - self.is_running = True - else: # For local host - return "Session is not running" + """Send input to the process""" + if not self.is_running: + return "Session is not running" try: - # --- Send to CTF --- if self.ctf: + # For CTF environments output = self.ctf.get_shell(input_data) self.output_buffer.append(output) return "Input sent to CTF session" - # --- Send to Local or Container PTY --- - if self.master is not None: - input_data_bytes = (input_data.rstrip() + "\n").encode() - bytes_written = os.write(self.master, input_data_bytes) - if bytes_written != len(input_data_bytes): - self.output_buffer.append(f"[Session {self.session_id}] Warning: Partial input write.") - self.last_activity = time.time() - return "Input sent to session" - else: - return "Session PTY not available for input" + # For local environment + input_data = input_data.rstrip() + "\n" + os.write(self.master, input_data.encode()) + self.last_activity = time.time() + return "Input sent to session" except Exception as e: # pylint: disable=broad-except - self.output_buffer.append(f"Error sending input: {str(e)}") return f"Error sending input: {str(e)}" def get_output(self, clear=True): @@ -244,12 +138,8 @@ class ShellSession: # pylint: disable=too-many-instance-attributes def terminate(self): """Terminate the session""" - session_id_short = self.session_id[:8] if not self.is_running: - if self.process and self.process.poll() is None: - pass # Process is running, proceed with termination - else: - return f"Session {session_id_short} already terminated or finished." + return "Session already terminated" try: self.is_running = False @@ -257,64 +147,28 @@ class ShellSession: # pylint: disable=too-many-instance-attributes if self.process: # Try to terminate the process group try: - os.killpg(os.getpgid(self.process.pid), signal.SIGTERM) - except ProcessLookupError: - pass # Process already gone - except subprocess.TimeoutExpired: - print(color(f"Session {session_id_short} did not terminate gracefully, sending SIGKILL...", fg="yellow")) # noqa E501 - try: - if pgid: - os.killpg(pgid, signal.SIGKILL) # Force kill - else: - self.process.kill() - except ProcessLookupError: - pass # Already gone - except Exception as kill_err: - termination_message = f" (Error during SIGKILL: {kill_err})" - except Exception as term_err: # Catch other errors during SIGTERM - termination_message = f" (Error during SIGTERM: {term_err})" - try: - self.process.kill() - except Exception: pass # Ignore nested errors + os.killpg(os.getpgid(self.process.pid), signal.SIGTERM) + except BaseException: # pylint: disable=bare-except,broad-except # noqa: E501 + # If that fails, try to terminate just the process + self.process.terminate() + # Clean up resources + if self.master: + os.close(self.master) + if self.slave: + os.close(self.slave) - # Final check - if self.process.poll() is None: - print(color(f"Session {session_id_short} process {self.process.pid} may still be running after termination attempts.", fg="red")) # noqa E501 - termination_message += " (Warning: Process may still be running)" - - - # Clean up PTY resources if they exist - if self.master: - try: os.close(self.master) - except OSError: pass - self.master = None - if self.slave: - try: os.close(self.slave) - except OSError: pass - self.slave = None - - return termination_message or f"Session {self.session_id} terminated" + return f"Session {self.session_id} terminated" except Exception as e: # pylint: disable=broad-except - return f"Error terminating session {session_id_short}: {str(e)}" + return f"Error terminating session: {str(e)}" -def create_shell_session(command, ctf=None, container_id=None, **kwargs): - """Create a new shell session in the correct workspace/environment.""" - if container_id: - session = ShellSession(command, ctf=ctf, container_id=container_id) - else: - workspace_dir = _get_workspace_dir() - session = ShellSession(command, ctf=ctf, workspace_dir=workspace_dir) - +def create_shell_session(command, ctf=None): + """Create a new shell session""" + session = ShellSession(command, ctf=ctf) session.start() - if session.is_running or (ctf and not session.is_running): - ACTIVE_SESSIONS[session.session_id] = session - return session.session_id - else: - error_msg = session.get_output(clear=True) - print(color(f"Failed to start session: {error_msg}", fg="red")) - return f"Failed to start session: {error_msg}" + ACTIVE_SESSIONS[session.session_id] = session + return session.session_id def list_shell_sessions(): @@ -358,100 +212,47 @@ def get_session_output(session_id, clear=True): def terminate_session(session_id): """Terminate a specific session""" if session_id not in ACTIVE_SESSIONS: - return f"Session {session_id} not found or already terminated." + return f"Session {session_id} not found" session = ACTIVE_SESSIONS[session_id] result = session.terminate() - if session_id in ACTIVE_SESSIONS: - del ACTIVE_SESSIONS[session_id] + del ACTIVE_SESSIONS[session_id] return result -def _run_ctf(ctf, command, stdout=False, timeout=100, workspace_dir=None): - """Runs command in CTF env, changing to workspace_dir first.""" - target_dir = workspace_dir or _get_workspace_dir() - full_command = f"cd '{target_dir}' && {command}" - original_cmd_for_msg = command # For logging - context_msg = f"(ctf:{target_dir})" +def _run_ctf(ctf, command, stdout=False, timeout=100, stream=False, call_id=None): try: - output = ctf.get_shell(full_command, timeout=timeout) + # Ensure the command is executed in a shell that supports command + # chaining + output = ctf.get_shell(command, timeout=timeout) + # exploit_logger.log_ok() + if stdout: - print(f"\033[32m{context_msg} $ {original_cmd_for_msg}\n{output}\033[0m") # noqa E501 - return output + print("\033[32m" + output + "\033[0m") + return output # output if output else result.stder except Exception as e: # pylint: disable=broad-except - error_msg = f"Error executing CTF command '{original_cmd_for_msg}' in '{target_dir}': {e}" # noqa E501 - print(color(error_msg, fg="red")) - return error_msg - -def _run_ssh(command, stdout=False, timeout=100, workspace_dir=None): - """Runs command via SSH. Assumes SSH agent or passwordless setup unless sshpass is used externally.""" # noqa E501 - ssh_user = os.environ.get('SSH_USER') - ssh_host = os.environ.get('SSH_HOST') - ssh_pass = os.environ.get('SSH_PASS') - remote_command = command - original_cmd_for_msg = command - context_msg = f"({ssh_user}@{ssh_host})" - - # Construct base SSH command list - if ssh_pass: - ssh_cmd_list = ["sshpass", "-p", ssh_pass, "ssh", f"{ssh_user}@{ssh_host}"] # noqa E501 - else: - ssh_cmd_list = ["ssh", f"{ssh_user}@{ssh_host}"] - ssh_cmd_list.append(remote_command) - - try: - # Use subprocess.run with list of args for better security than shell=True - result = subprocess.run( - ssh_cmd_list, - capture_output=True, - text=True, - check=False, # Don't raise exception on non-zero exit code - timeout=timeout - ) - output = result.stdout if result.stdout else result.stderr - if stdout: - print(f"\033[32m{context_msg} $ {original_cmd_for_msg}\n{output}\033[0m") # noqa E501 - # Return combined output, potentially including errors - return output.strip() - except subprocess.TimeoutExpired as e: - error_output = e.stdout if e.stdout else str(e) - timeout_msg = f"Timeout executing SSH command: {error_output}" - if stdout: - print(f"\033[33m{context_msg} $ {original_cmd_for_msg}\nTIMEOUT\n{error_output}\033[0m") # noqa E501 - return timeout_msg - except FileNotFoundError: - # Handle case where ssh or sshpass isn't installed - error_msg = f"'sshpass' or 'ssh' command not found. Ensure they are installed and in PATH." # noqa E501 - print(color(error_msg, fg="red")) - return error_msg - except Exception as e: # pylint: disable=broad-except - error_msg = f"Error executing SSH command '{original_cmd_for_msg}' on {ssh_host}: {e}" # noqa E501 - print(color(error_msg, fg="red")) - return error_msg + print(color(f"Error executing CTF command: {e}", fg="red")) + # exploit_logger.log_error(str(e)) + return f"Error executing CTF command: {str(e)}" -def _run_local(command, stdout=False, timeout=100, stream=False, call_id=None, tool_name=None, workspace_dir=None): - """Runs command locally in the specified workspace_dir.""" +def _run_local(command, stdout=False, timeout=100, stream=False, call_id=None, tool_name=None): # If streaming is enabled and we have a call_id if stream and call_id: - return _run_local_streamed(command, call_id, timeout, tool_name, workspace_dir) + return _run_local_streamed(command, call_id, timeout, tool_name) - target_dir = workspace_dir or _get_workspace_dir() - original_cmd_for_msg = command # For logging - context_msg = f"(local:{target_dir})" try: + # nosec B602 - shell=True is required for command chaining result = subprocess.run( command, shell=True, # nosec B602 capture_output=True, text=True, - check=False, - timeout=timeout, - cwd=target_dir - ) + check=False, + timeout=timeout) output = result.stdout if result.stdout else result.stderr if stdout: - print(f"\033[32m{context_msg} $ {original_cmd_for_msg}\n{output}\033[0m") # noqa E501 + print("\033[32m" + output + "\033[0m") # Skip passing output to cli_print_tool_output when CAI_STREAM=true # This prevents duplicate output in streaming mode @@ -460,21 +261,20 @@ def _run_local(command, stdout=False, timeout=100, stream=False, call_id=None, t # Optional: Add cli_print_tool_output call here if needed for non-streaming pass - return output.strip() + return output except subprocess.TimeoutExpired as e: - error_output = e.stdout if e.stdout else str(e) + error_output = e.stdout.decode() if e.stdout else str(e) if stdout: print("\033[32m" + error_output + "\033[0m") - return error_output + return error_output except Exception as e: # pylint: disable=broad-except - error_msg = f"Error executing local command: {e}" - print(color(error_msg, fg="red")) - return error_msg + error_msg = f"Error executing local command: {e}" + print(color(error_msg, fg="red")) + return error_msg -def _run_local_streamed(command, call_id, timeout=100, tool_name=None, workspace_dir=None): - """Run a local command with streaming output to the Tool output panel.""" - target_dir = workspace_dir or _get_workspace_dir() +def _run_local_streamed(command, call_id, timeout=100, tool_name=None): + """Run a local command with streaming output to the Tool output panel""" try: # Try to import Rich for nice display try: @@ -498,8 +298,7 @@ def _run_local_streamed(command, call_id, timeout=100, tool_name=None, workspace stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True, - bufsize=1, - cwd=target_dir # Set CWD for local process + bufsize=1 ) # If tool_name is not provided, derive it from the command @@ -529,7 +328,7 @@ def _run_local_streamed(command, call_id, timeout=100, tool_name=None, workspace header.append(")", style="yellow") tool_time = 0 start_time = time.time() - total_time = time.time() - START_TIME + total_time = time.time() - START_TIME timing_info = [] if total_time: timing_info.append(f"Total: {format_time(total_time)}") @@ -706,8 +505,7 @@ def run_command(command, ctf=None, stdout=False, # pylint: disable=too-many-arg async_mode=False, session_id=None, timeout=100, stream=False, call_id=None, tool_name=None): """ - Run command in the appropriate environment (Docker, CTF, SSH, Local) - and workspace. + Run command either in CTF container or on the local attacker machine Args: command: The command to execute @@ -722,168 +520,34 @@ def run_command(command, ctf=None, stdout=False, # pylint: disable=too-many-arg If None, the tool name will be derived from the command. Returns: - str: Command output, status message, or session ID. + str: Command output, status message, or session ID """ # If session_id is provided, send command to that session if session_id: if session_id not in ACTIVE_SESSIONS: return f"Session {session_id} not found" - session = ACTIVE_SESSIONS[session_id] - result = session.send_input(command) # Send the raw command string + + result = send_to_session(session_id, command) if stdout: output = get_session_output(session_id, clear=False) - env_type = "Local" - if session.container_id: - env_type = f"Container({session.container_id[:12]})" - elif session.ctf: - env_type = "CTF" - print(f"\033[32m(Session {session_id} in {env_type}:{session.workspace_dir}) >> {command}\n{output}\033[0m") # noqa E501 - return result # Return the result of sending input ("Input sent..." or error) + print("\033[32m" + output + "\033[0m") + return result + + # If async_mode, create a new session + if async_mode: + session_id = create_shell_session(command, ctf) + if stdout: + # Wait a moment for initial output + time.sleep(0.5) + output = get_session_output(session_id, clear=False) + print("\033[32m" + output + "\033[0m") + return f"Created session {session_id}. Use this ID to interact with the session." # Generate a call_id if we're streaming and one wasn't provided if stream and not call_id: call_id = str(uuid.uuid4())[:8] - - # 2. Determine Execution Environment (Container > CTF > SSH > Local) - active_container = os.getenv("CAI_ACTIVE_CONTAINER", "") - is_ssh_env = all(os.getenv(var) for var in ['SSH_USER', 'SSH_HOST']) - - # --- Docker Container Execution --- - if active_container and not ctf and not is_ssh_env: - container_id = active_container - container_workspace = _get_container_workspace_path() - context_msg = f"(docker:{container_id[:12]}:{container_workspace})" - - # Handle Async Session Creation in Container - if async_mode: - # Create a session specifically for the container environment - new_session_id = create_shell_session(command, container_id=container_id) # noqa E501 - if "Failed" in new_session_id: # Check if session creation failed - return new_session_id - if stdout: - # Wait a moment for initial output - time.sleep(0.2) - output = get_session_output(new_session_id, clear=False) - print(f"\033[32m(Started Session {new_session_id} in {context_msg})\n{output}\033[0m") # noqa E501 - return f"Started async session {new_session_id} in container {container_id[:12]}. Use this ID to interact." # noqa E501 - - # Handle Streaming Container Execution - not yet implemented for containers - if stream: - # For now, display that streaming isn't supported for containers - from cai.util import cli_print_tool_output - if call_id and tool_name: - tool_args = {"command": command, "container": container_id[:12]} - cli_print_tool_output( - tool_name, - tool_args, - "Streaming not yet supported for container execution. Running normally...", - call_id=call_id - ) - - # Handle Synchronous Execution in Container - try: - # Ensure container workspace exists (best effort) - # Consider moving this to workspace set/container activation - mkdir_cmd = ["docker", "exec", container_id, "mkdir", "-p", container_workspace] # noqa E501 - subprocess.run(mkdir_cmd, capture_output=True, text=True, check=False, timeout=10) # noqa E501 - - # Construct the docker exec command with workspace context - cmd_list = [ - "docker", "exec", - "-w", container_workspace, # Set working directory - container_id, - "sh", "-c", command # Execute command via shell - ] - result = subprocess.run( - cmd_list, - capture_output=True, - text=True, - check=False, # Don't raise exception on non-zero exit - timeout=timeout - ) - - output = result.stdout if result.stdout else result.stderr - output = output.strip() # Clean trailing newline - - if stdout: - print(f"\033[32m{context_msg} $ {command}\n{output}\033[0m") # noqa E501 - - # Check if command failed specifically because container isn't running - if result.returncode != 0 and "is not running" in result.stderr: - print(color(f"{context_msg} Container is not running. Attempting execution on host instead.", fg="yellow")) # noqa E501 - # Fallback to local execution, preserving workspace context - return _run_local(command, stdout, timeout, stream, call_id, tool_name, _get_workspace_dir()) # noqa E501 - - return output # Return combined stdout/stderr - - except subprocess.TimeoutExpired: - timeout_msg = "Timeout executing command in container." - if stdout: - print(f"\033[33m{context_msg} $ {command}\nTIMEOUT\033[0m") # noqa E501 - print(color("Attempting execution on host instead.", fg="yellow")) - # Fallback to local execution on timeout - return _run_local(command, stdout, timeout, stream, call_id, tool_name, _get_workspace_dir()) # noqa E501 - except Exception as e: # pylint: disable=broad-except - error_msg = f"Error executing command in container: {str(e)}" - print(color(f"{context_msg} {error_msg}", fg="red")) - print(color("Attempting execution on host instead.", fg="yellow")) - # Fallback to local execution on other errors - return _run_local(command, stdout, timeout, stream, call_id, tool_name, _get_workspace_dir()) # noqa E501 - - # --- CTF Execution --- + + # Otherwise, run command normally if ctf: - # Handling streaming for CTF - not fully implemented yet - if stream: - from cai.util import cli_print_tool_output - if call_id and tool_name: - tool_args = {"command": command, "ctf": True} - cli_print_tool_output( - tool_name, - tool_args, - "Streaming not yet supported for CTF execution. Running normally...", - call_id=call_id - ) - - # _run_ctf handles workspace internally using _get_workspace_dir() default - return _run_ctf(ctf, command, stdout, timeout) # Pass None for workspace_dir - - # --- SSH Execution --- - if is_ssh_env: - # Async for SSH would require session management via SSH client features - if async_mode: - return "Async mode not fully supported for SSH environment via this function yet." - - # Handling streaming for SSH - not fully implemented yet - if stream: - from cai.util import cli_print_tool_output - if call_id and tool_name: - tool_args = {"command": command, "ssh": True} - cli_print_tool_output( - tool_name, - tool_args, - "Streaming not yet supported for SSH execution. Running normally...", - call_id=call_id - ) - - # _run_ssh handles command execution, workspace is relative to remote home - return _run_ssh(command, stdout, timeout) # Workspace dir less relevant here - - # --- Local Execution (Default Fallback) --- - # Let _run_local handle determining the host workspace - # Handle Async Session Creation Locally - if async_mode: - # create_shell_session uses _get_workspace_dir() when container_id is None - new_session_id = create_shell_session(command) - if isinstance(new_session_id, str) and "Failed" in new_session_id: # Check failure - return new_session_id - # Retrieve the actual workspace dir the session is using - session = ACTIVE_SESSIONS.get(new_session_id) - actual_workspace = session.workspace_dir if session else "unknown" - if stdout: - time.sleep(0.2) # Allow session buffer to populate - output = get_session_output(new_session_id, clear=False) - print(f"\033[32m(Started Session {new_session_id} in local:{actual_workspace})\n{output}\033[0m") - return f"Started async session {new_session_id} locally. Use this ID to interact." - - # Handle Synchronous Execution Locally using _run_local default with streaming support - return _run_local(command, stdout, timeout, stream, call_id, tool_name, None) + return _run_ctf(ctf, command, stdout, timeout, stream, call_id) + return _run_local(command, stdout, timeout, stream, call_id, tool_name) From 23e1e65052ee3eab34ef89af6bc23b987df957d2 Mon Sep 17 00:00:00 2001 From: luijait Date: Tue, 6 May 2025 17:18:29 +0200 Subject: [PATCH 07/10] Delete unnecesary lines on run.py --- src/cai/sdk/agents/run.py | 33 --------------------------------- 1 file changed, 33 deletions(-) diff --git a/src/cai/sdk/agents/run.py b/src/cai/sdk/agents/run.py index 30d82d2f..e2379560 100644 --- a/src/cai/sdk/agents/run.py +++ b/src/cai/sdk/agents/run.py @@ -779,22 +779,6 @@ class Runner: run_config: RunConfig, tool_use_tracker: AgentToolUseTracker, ) -> SingleStepResult: - # Log the raw model response output, focusing on tool calls - logger.debug(f"[_get_single_step_result_from_response] Raw new_response.id: {new_response.referenceable_id}") - if new_response.output: - for i, item in enumerate(new_response.output): - item_type = type(item).__name__ - logger.debug(f"[_get_single_step_result_from_response] Raw output item [{i}] type: {item_type}") - if hasattr(item, "name") and hasattr(item, "arguments"): # For ResponseFunctionToolCall - logger.debug( - f"[_get_single_step_result_from_response] Raw ResponseFunctionToolCall item [{i}]: " - f"Name='{getattr(item, 'name', 'N/A')}', " - f"Args='{getattr(item, 'arguments', 'N/A')}', " - f"CallID='{getattr(item, 'call_id', 'N/A')}'" - ) - elif hasattr(item, "text"): # For ResponseOutputText - logger.debug(f"[_get_single_step_result_from_response] Raw ResponseOutputText item [{i}]: Text='{getattr(item, 'text', '')[:100]}...'") - processed_response = RunImpl.process_model_response( agent=agent, @@ -803,26 +787,9 @@ class Runner: output_schema=output_schema, handoffs=handoffs, ) - - # Log the processed response, focusing on tools_used - logger.debug(f"[_get_single_step_result_from_response] Processed response type: {type(processed_response).__name__}") - if hasattr(processed_response, 'is_final_output'): - logger.debug(f"[_get_single_step_result_from_response] Processed response: is_final_output={processed_response.is_final_output}") - else: - logger.debug(f"[_get_single_step_result_from_response] Processed response does not have is_final_output attribute") - - if hasattr(processed_response, 'final_output_from_llm') and processed_response.final_output_from_llm is not None: - logger.debug(f"[_get_single_step_result_from_response] Processed response: final_output_from_llm='{str(processed_response.final_output_from_llm)[:100]}...'") # Log tools used with robust type checking if hasattr(processed_response, 'tools_used') and processed_response.tools_used: - # Log summarizing number of tools used first - logger.debug(f"[_get_single_step_result_from_response] Found {len(processed_response.tools_used)} tools used") - - # Add spacing between blocks in terminal output - if os.environ.get('CAI_STREAM', 'false').lower() == 'true': - print("") # Visual separator only when streaming is enabled - for i, tool_call in enumerate(processed_response.tools_used): try: # Safely extract tool name with multiple fallbacks From d67f944b5e6af17cb764de906b51afb0474ac2c0 Mon Sep 17 00:00:00 2001 From: luijait Date: Tue, 6 May 2025 17:19:32 +0200 Subject: [PATCH 08/10] Delete unnecesary lines on run.py --- src/cai/sdk/agents/run.py | 1 - 1 file changed, 1 deletion(-) diff --git a/src/cai/sdk/agents/run.py b/src/cai/sdk/agents/run.py index e2379560..103661e0 100644 --- a/src/cai/sdk/agents/run.py +++ b/src/cai/sdk/agents/run.py @@ -4,7 +4,6 @@ import asyncio import copy from dataclasses import dataclass, field from typing import Any, cast -import os from openai.types.responses import ResponseCompletedEvent From 3fb20717217db71610713c5c77fd6328114388f3 Mon Sep 17 00:00:00 2001 From: luijait Date: Tue, 6 May 2025 17:22:02 +0200 Subject: [PATCH 09/10] Delete trash on openaichatcompletions --- src/cai/sdk/agents/models/openai_chatcompletions.py | 11 ----------- 1 file changed, 11 deletions(-) diff --git a/src/cai/sdk/agents/models/openai_chatcompletions.py b/src/cai/sdk/agents/models/openai_chatcompletions.py index 21c5ef81..bf35c06c 100644 --- a/src/cai/sdk/agents/models/openai_chatcompletions.py +++ b/src/cai/sdk/agents/models/openai_chatcompletions.py @@ -1089,10 +1089,6 @@ class OpenAIChatCompletionsModel(Model): direct_stats = final_stats.copy() direct_stats["interaction_cost"] = float(interaction_cost) direct_stats["total_cost"] = float(total_cost) - - # Add a small delay to avoid overlapping outputs - await asyncio.sleep(0.3) - # Use the direct copy with guaranteed float costs finish_agent_streaming(streaming_context, direct_stats) @@ -1103,12 +1099,6 @@ class OpenAIChatCompletionsModel(Model): # Find the assistant message to print for item in final_response.output: if isinstance(item, ResponseOutputMessage) and item.role == 'assistant': - # Add a small delay to avoid overlapping outputs - await asyncio.sleep(0.3) - - # Print clear visual separator before message - print("\n") - cli_print_agent_messages( agent_name=getattr(self, 'agent_name', 'Agent'), message=item, @@ -1612,7 +1602,6 @@ class OpenAIChatCompletionsModel(Model): } }] - # Handle special Ollama generic_linux_command format if isinstance(delta, dict) and 'content' in delta: content = delta['content'] # Try to detect if the content is a JSON string with function call format From e6c9b5bfe8ab22b09c011b42eebd506545eeb3c1 Mon Sep 17 00:00:00 2001 From: luijait Date: Tue, 6 May 2025 17:24:33 +0200 Subject: [PATCH 10/10] Delete trash on openaichatcompletions --- src/cai/sdk/agents/models/openai_chatcompletions.py | 4 ---- 1 file changed, 4 deletions(-) diff --git a/src/cai/sdk/agents/models/openai_chatcompletions.py b/src/cai/sdk/agents/models/openai_chatcompletions.py index bf35c06c..18134f83 100644 --- a/src/cai/sdk/agents/models/openai_chatcompletions.py +++ b/src/cai/sdk/agents/models/openai_chatcompletions.py @@ -550,10 +550,6 @@ class OpenAIChatCompletionsModel(Model): model_str = str(self.model).lower() is_ollama = self.is_ollama or "ollama" in model_str or ":" in model_str or "qwen" in model_str - - # Add a small delay to make sure any previous tool outputs are fully rendered - # This helps prevent overlapping of panels in the terminal - await asyncio.sleep(0.2) # Add visual separation before agent output if streaming_context and should_show_rich_stream: