decodingtrust-agent-sdk 0.2.2__tar.gz → 0.2.4__tar.gz

{decodingtrust_agent_sdk-0.2.2/decodingtrust_agent_sdk.egg-info → decodingtrust_agent_sdk-0.2.4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: decodingtrust-agent-sdk
-Version: 0.2.2
+Version: 0.2.4
 Summary: DecodingTrust Agent Platform (DTap) — A controllable and interactive red-teaming platform for AI agents
 Author-email: DTap Team <zhaorun@uchicago.edu>
 License:                                  Apache License
@@ -253,10 +253,13 @@ Requires-Dist: claude-agent-sdk>=0.1.18; extra == "claude"
 Requires-Dist: claude-code-sdk>=0.0.25; extra == "claude"
 Provides-Extra: google
 Requires-Dist: google-adk>=1.27.5; extra == "google"
+Requires-Dist: a2a-sdk<0.4.0,>=0.3.22; extra == "google"
 Requires-Dist: google-generativeai>=0.3.0; extra == "google"
 Requires-Dist: google-genai>=1.0.0; extra == "google"
 Requires-Dist: google-api-core>=2.28.0; extra == "google"
 Requires-Dist: google-api-python-client>=2.100.0; extra == "google"
+Provides-Extra: strands
+Requires-Dist: strands-agents>=1.40.0; extra == "strands"
 Provides-Extra: langchain
 Requires-Dist: langchain>=0.3.0; extra == "langchain"
 Requires-Dist: langchain-core>=0.3.0; extra == "langchain"
@@ -265,7 +268,7 @@ Requires-Dist: langchain-anthropic>=0.2.0; extra == "langchain"
 Provides-Extra: pocketflow
 Requires-Dist: pocketflow==0.0.3; extra == "pocketflow"
 Provides-Extra: all
-Requires-Dist: decodingtrust-agent-sdk[claude,google,langchain,openai,pocketflow]; extra == "all"
+Requires-Dist: decodingtrust-agent-sdk[claude,google,langchain,openai,pocketflow,strands]; extra == "all"
 Provides-Extra: dev
 Requires-Dist: pytest>=7.0; extra == "dev"
 Requires-Dist: pytest-asyncio>=0.21; extra == "dev"
@@ -276,7 +279,7 @@ Dynamic: license-file
 
 <div align="center">
 
-<img src="dt-logo-new.png" alt="DecodingTrust-Agent" width="180"/>
+<img src="https://raw.githubusercontent.com/AI-secure/DecodingTrust-Agent/main/dt-logo-new.png" alt="DecodingTrust-Agent" width="180"/>
 
 # DecodingTrust-Agent Platform (DTAP)
 
@@ -285,7 +288,7 @@ Dynamic: license-file
 [![GitHub Repo stars](https://img.shields.io/github/stars/AI-secure/DecodingTrust-Agent?style=social)](https://github.com/AI-secure/DecodingTrust-Agent/stargazers)
 [![GitHub last commit](https://img.shields.io/github/last-commit/AI-secure/DecodingTrust-Agent)](https://github.com/AI-secure/DecodingTrust-Agent/commits/main)
 [![GitHub contributors](https://img.shields.io/github/contributors/AI-secure/DecodingTrust-Agent?color=orange)](https://github.com/AI-secure/DecodingTrust-Agent/graphs/contributors)
-[![License](https://img.shields.io/github/license/AI-secure/DecodingTrust-Agent)](LICENSE)
+[![License](https://img.shields.io/github/license/AI-secure/DecodingTrust-Agent)](https://github.com/AI-secure/DecodingTrust-Agent/blob/main/LICENSE)
 [![Website](https://img.shields.io/badge/website-decodingtrust--agent.com-blue)](https://decodingtrust-agent.com)
 [![Paper](https://img.shields.io/badge/paper-DTAP-b31b1b)](https://decodingtrust-agent.com)
 

{decodingtrust_agent_sdk-0.2.2 → decodingtrust_agent_sdk-0.2.4}/README.md

@@ -1,6 +1,6 @@
 <div align="center">
 
-<img src="dt-logo-new.png" alt="DecodingTrust-Agent" width="180"/>
+<img src="https://raw.githubusercontent.com/AI-secure/DecodingTrust-Agent/main/dt-logo-new.png" alt="DecodingTrust-Agent" width="180"/>
 
 # DecodingTrust-Agent Platform (DTAP)
 
@@ -9,7 +9,7 @@
 [![GitHub Repo stars](https://img.shields.io/github/stars/AI-secure/DecodingTrust-Agent?style=social)](https://github.com/AI-secure/DecodingTrust-Agent/stargazers)
 [![GitHub last commit](https://img.shields.io/github/last-commit/AI-secure/DecodingTrust-Agent)](https://github.com/AI-secure/DecodingTrust-Agent/commits/main)
 [![GitHub contributors](https://img.shields.io/github/contributors/AI-secure/DecodingTrust-Agent?color=orange)](https://github.com/AI-secure/DecodingTrust-Agent/graphs/contributors)
-[![License](https://img.shields.io/github/license/AI-secure/DecodingTrust-Agent)](LICENSE)
+[![License](https://img.shields.io/github/license/AI-secure/DecodingTrust-Agent)](https://github.com/AI-secure/DecodingTrust-Agent/blob/main/LICENSE)
 [![Website](https://img.shields.io/badge/website-decodingtrust--agent.com-blue)](https://decodingtrust-agent.com)
 [![Paper](https://img.shields.io/badge/paper-DTAP-b31b1b)](https://decodingtrust-agent.com)
 

decodingtrust_agent_sdk-0.2.4/agent/__init__.py

@@ -0,0 +1,50 @@
+"""Agent architecture registry.
+
+Each adapter is imported defensively: a missing optional dependency for one
+framework (e.g. ``pocketflow``, ``strands``) must not break importing the whole
+``agent`` package or any other adapter. Unavailable adapters are simply omitted
+from ``AGENT_REGISTRY`` (mirrors the lazy registry in ``utils.agent_helpers``).
+"""
+
+# (registry_key, module_path, class_name)
+_ADAPTERS = [
+    ("pocketflow", "agent.pocketflow.src.agent", "MCPReactAgent"),
+    ("openaisdk", "agent.openaisdk.src.agent", "OpenAISDKAgent"),
+    ("googleadk", "agent.googleadk.src.agent", "GoogleADKAgent"),
+    ("langchain", "agent.langchain.src.agent", "LangChainAgent"),
+    ("claudesdk", "agent.claudesdk.src.agent", "ClaudeSDKAgent"),
+    ("openclaw", "agent.openclaw.src.agent", "OpenClawAgent"),
+    ("strands", "agent.strands.src.agent", "StrandsSDKAgent"),
+    ("hermes", "agent.hermes.src.agent", "HermesAgent"),
+]
+
+# Agent architecture registry (only successfully-imported adapters)
+AGENT_REGISTRY = {}
+
+import importlib as _importlib
+
+for _key, _module_path, _class_name in _ADAPTERS:
+    try:
+        _mod = _importlib.import_module(_module_path)
+        _cls = getattr(_mod, _class_name)
+    except Exception as _e:  # ImportError or any load-time error from optional deps
+        globals()[_class_name] = None
+        print(f"[WARNING] Agent '{_key}' not available: {_e}")
+    else:
+        globals()[_class_name] = _cls
+        AGENT_REGISTRY[_key] = _cls
+
+AVAILABLE_ARCHITECTURES = list(AGENT_REGISTRY.keys())
+
+__all__ = [
+    "MCPReactAgent",
+    "OpenAISDKAgent",
+    "GoogleADKAgent",
+    "LangChainAgent",
+    "ClaudeSDKAgent",
+    "OpenClawAgent",
+    "StrandsSDKAgent",
+    "HermesAgent",
+    "AGENT_REGISTRY",
+    "AVAILABLE_ARCHITECTURES",
+]

{decodingtrust_agent_sdk-0.2.2 → decodingtrust_agent_sdk-0.2.4}/agent/claudesdk/src/agent.py

@@ -1,11 +1,14 @@
 import os
+import shutil
+import tempfile
 import uuid
 from datetime import datetime
 from typing import Dict, Any, Optional, List, Union
 
-from claude_code_sdk import (
+from claude_agent_sdk import (
+    AgentDefinition,
+    ClaudeAgentOptions,
     ClaudeSDKClient,
-    ClaudeCodeOptions,
     ClaudeSDKError,
     Message,
     AssistantMessage,
@@ -67,9 +70,12 @@ class ClaudeSDKAgent(Agent):
         # Skill injection temp directory
         self._skill_temp_dir: Optional[str] = None
 
+        # Per-instance Claude Code home. 
+        self._claude_home: str = tempfile.mkdtemp(prefix="claude-home-")
+
         # Claude SDK client for multi-turn conversations
         self._client: Optional[ClaudeSDKClient] = None
-        self._options: Optional[ClaudeCodeOptions] = None
+        self._options: Optional[ClaudeAgentOptions] = None
 
         # Multi-turn conversation state
         self._trace_id: Optional[str] = None
@@ -122,6 +128,10 @@ class ClaudeSDKAgent(Agent):
             "permission_mode": "bypassPermissions",
             "max_turns": self.runtime_config.max_turns,
             "model": self.runtime_config.model,
+            # Isolate Claude Code state per instance so parallel victims don't
+            # share sessions/locks and interrupt each other. Merge onto the
+            # current environment so API keys and proxy settings are preserved.
+            "env": {**os.environ, "CLAUDE_CONFIG_DIR": self._claude_home},
         }
 
         if self._skill_temp_dir:
@@ -144,10 +154,14 @@ class ClaudeSDKAgent(Agent):
         for server_config, server_name in zip(self.mcp_servers, self.mcp_server_names):
             self.mcp_server_configs[server_name] = server_config
 
+        # Create inline subagents (non-A2A) — also registers their MCP servers
+        agents = self._create_sub_agents(self.mcp_server_configs)
+
         mcp = self.mcp_server_configs if self.mcp_server_configs else None
         options_kwargs = self._build_options_kwargs(mcp_servers=mcp)
+        options_kwargs["agents"] = agents
 
-        self._options = ClaudeCodeOptions(**options_kwargs)
+        self._options = ClaudeAgentOptions(**options_kwargs)
 
         # Create and connect the client for multi-turn conversations
         self._client = ClaudeSDKClient(options=self._options)
@@ -204,11 +218,15 @@ class ClaudeSDKAgent(Agent):
         # Store server names for reference
         self.mcp_server_names = list(sdk_mcp_servers.keys())
 
+        # Create inline subagents (non-A2A) — also registers their MCP servers
+        agents = self._create_sub_agents(sdk_mcp_servers)
+
         # Build Claude SDK options with SDK MCP servers
         mcp = sdk_mcp_servers if sdk_mcp_servers else None
         options_kwargs = self._build_options_kwargs(mcp_servers=mcp)
+        options_kwargs["agents"] = agents
 
-        self._options = ClaudeCodeOptions(**options_kwargs)
+        self._options = ClaudeAgentOptions(**options_kwargs)
 
         # Create and connect the client
         self._client = ClaudeSDKClient(options=self._options)
@@ -247,6 +265,52 @@ class ClaudeSDKAgent(Agent):
         else:
             raise ValueError(f"Unsupported transport type: {server_config.transport}")
 
+    def _create_sub_agents(self, mcp_servers_dict: Dict[str, Any]) -> Optional[Dict[str, AgentDefinition]]:
+        """
+        Create subagent definitions from config and register their MCP servers.
+
+        For each non-A2A subagent, registers its MCP servers into the provided
+        mcp_servers_dict (top-level) so the subagent can access them, and builds
+        an AgentDefinition.
+
+        Args:
+            mcp_servers_dict: Top-level MCP servers dict to merge subagent servers into
+
+        Returns:
+            Dict of agent definitions, or None if no subagents configured
+        """
+        if not self.config.sub_agents:
+            return None
+
+        agents: Dict[str, AgentDefinition] = {}
+
+        for sub_config in self.config.sub_agents:
+            if sub_config.a2a:
+                # A2A subagents are handled separately
+                continue
+
+            # Register subagent's MCP servers at the top level
+            sub_server_names = []
+            for server_config in sub_config.mcp_servers:
+                if server_config.enabled:
+                    mcp_servers_dict[server_config.name] = self._create_mcp_server(server_config)
+                    sub_server_names.append(server_config.name)
+
+            description = sub_config.description or f"Subagent: {sub_config.name}"
+
+            # Build tools list: MCP tool glob patterns for this subagent's servers
+            tools = [f"mcp__{name}__*" for name in sub_server_names]
+
+            print(f"[SUBAGENT] Creating AgentDefinition '{sub_config.name}' with {len(sub_server_names)} MCP servers")
+
+            agents[sub_config.name] = AgentDefinition(
+                description=description,
+                prompt=sub_config.system_prompt,
+                tools=tools if tools else None,
+            )
+
+        return agents if agents else None
+
     async def run(
         self,
         user_input: Union[str, List[str]],
@@ -398,3 +462,7 @@ class ClaudeSDKAgent(Agent):
         if self._skill_temp_dir:
             cleanup_temp_directory(self._skill_temp_dir)
             self._skill_temp_dir = None
+
+        # Clean up per-instance Claude Code home
+        if self._claude_home and os.path.isdir(self._claude_home):
+            shutil.rmtree(self._claude_home, ignore_errors=True)

{decodingtrust_agent_sdk-0.2.2 → decodingtrust_agent_sdk-0.2.4}/agent/claudesdk/src/utils.py

@@ -5,7 +5,7 @@ from datetime import datetime, timezone
 from typing import List, Dict, Any, Optional
 from collections import defaultdict
 
-from claude_code_sdk import (
+from claude_agent_sdk import (
     Message,
     UserMessage,
     AssistantMessage,

{decodingtrust_agent_sdk-0.2.2 → decodingtrust_agent_sdk-0.2.4}/agent/googleadk/src/agent.py

@@ -6,6 +6,8 @@ from typing import Dict, Any, Optional, List, Union
 
 from google.genai import types
 from google.adk.agents.llm_agent import LlmAgent
+from google.adk.agents.remote_a2a_agent import RemoteA2aAgent, AGENT_CARD_WELL_KNOWN_PATH
+from google.adk.models.lite_llm import LiteLlm
 from google.adk.runners import Runner
 from google.adk.sessions import InMemorySessionService
 from google.adk.artifacts.in_memory_artifact_service import InMemoryArtifactService
@@ -15,7 +17,13 @@ from google.adk.tools.mcp_tool.mcp_session_manager import (
     StdioServerParameters,
 )
 
-from dt_arena.src.types.agent import Agent, AgentConfig, RuntimeConfig, MCPServerConfig, AgentResult
+from dt_arena.src.types.agent import (
+    Agent,
+    AgentConfig,
+    RuntimeConfig,
+    MCPServerConfig,
+    AgentResult,
+)
 from dt_arena.src.types.trajectory import Trajectory
 
 from .mcp_wrapper import McpToolset
@@ -66,6 +74,7 @@ class GoogleADKAgent(Agent):
 
         # Google ADK specific components
         self.toolsets: List[McpToolset] = []
+        self.subagent_toolsets: List[McpToolset] = []  # Track subagent MCP toolsets for cleanup
         self.session_service = None
         self.artifact_service = None
         self.runner = None
@@ -106,6 +115,31 @@ class GoogleADKAgent(Agent):
             return None
         return load_skills_as_toolset(self._skill_temp_dir, skill_subpath="skills")
 
+    def _get_model(self):
+        """
+        Get the appropriate model object for LlmAgent.
+
+        For Gemini models, returns the model string directly.
+        For non-Gemini models (OpenAI, Anthropic), wraps in LiteLlm with provider prefix.
+
+        Returns:
+            str or LiteLlm: Model string for Gemini, LiteLlm wrapper for others
+        """
+        model = self.runtime_config.model
+
+        # Check if provider prefix already present
+        if "/" in model:
+            return LiteLlm(model=model)
+
+        # Simple matching for non-Gemini models
+        if "gpt" in model.lower():
+            return LiteLlm(model=f"openai/{model}")
+        elif "claude" in model.lower():
+            return LiteLlm(model=f"anthropic/{model}")
+
+        # Default: assume Gemini model, pass directly
+        return model
+
     async def initialize(self) -> None:
         """Initialize agent and connect to MCP servers"""
         if not self.config:
@@ -129,12 +163,16 @@ class GoogleADKAgent(Agent):
         if skill_toolset:
             agent_tools.append(skill_toolset)
 
+        # Create subagents if configured
+        sub_agents = await self._create_sub_agents()
+
         # Create Google ADK agent instance
         self.agent = LlmAgent(
-            model=self.runtime_config.model,
+            model=self._get_model(),
             name=self.config.name,
             instruction=self.config.system_prompt,
             tools=agent_tools,
+            sub_agents=sub_agents or [],
         )
 
         # Create runner
@@ -152,6 +190,97 @@ class GoogleADKAgent(Agent):
             user_id='user_default'
         )
 
+    async def _create_sub_agents(self) -> List[Union[LlmAgent, RemoteA2aAgent]]:
+        """
+        Create subagents from config.
+
+        Returns:
+            List of LlmAgent (non-A2A) or RemoteA2aAgent (A2A) instances
+        """
+        sub_agents = []
+
+        for sub_config in self.config.sub_agents:
+            if sub_config.a2a:
+                # A2A remote subagent
+                sub_agent = await self._create_a2a_subagent(sub_config)
+            else:
+                # Inline (non-A2A) subagent
+                sub_agent = await self._create_inline_subagent(sub_config)
+
+            if sub_agent:
+                sub_agents.append(sub_agent)
+
+        return sub_agents
+
+    async def _create_a2a_subagent(self, sub_config: AgentConfig) -> RemoteA2aAgent:
+        """
+        Create an A2A remote subagent.
+
+        Args:
+            sub_config: Subagent configuration with a2a=True
+
+        Returns:
+            RemoteA2aAgent instance
+        """
+        port = sub_config.port or 9001
+
+        # Build agent card URL
+        agent_card_url = f"http://localhost:{port}{AGENT_CARD_WELL_KNOWN_PATH}"
+
+        # Use description from config or generate default
+        description = sub_config.description or f"Remote A2A agent: {sub_config.name}"
+
+        # Apply A2A injection if configured
+        if self.runtime_config.a2a_injection:
+            injection = self.runtime_config.a2a_injection.get(sub_config.name)
+            if injection:
+                if injection.type == "suffix":
+                    description = f"{description}\n{injection.content}"
+                    print(f"[A2A INJECTION] Applied suffix injection to '{sub_config.name}'")
+                elif injection.type == "override":
+                    description = injection.content
+                    print(f"[A2A INJECTION] Applied override injection to '{sub_config.name}'")
+
+        print(f"[A2A] Creating RemoteA2aAgent '{sub_config.name}' at {agent_card_url}")
+
+        return RemoteA2aAgent(
+            name=sub_config.name,
+            description=description,
+            agent_card=agent_card_url,
+        )
+
+    async def _create_inline_subagent(self, sub_config: AgentConfig) -> LlmAgent:
+        """
+        Create an inline (non-A2A) subagent.
+
+        Args:
+            sub_config: Subagent configuration with a2a=False
+
+        Returns:
+            LlmAgent instance
+        """
+        # Load MCP servers for this subagent
+        sub_toolsets = []
+        for server_config in sub_config.mcp_servers:
+            if server_config.enabled:
+                toolset = self._create_mcp_server(server_config)
+                sub_toolsets.append(toolset)
+                # Track for cleanup
+                self.subagent_toolsets.append(toolset)
+
+        # Use description from config
+        description = sub_config.description or f"Subagent: {sub_config.name}"
+
+        print(f"[SUBAGENT] Creating inline LlmAgent '{sub_config.name}' with {len(sub_toolsets)} MCP servers")
+
+        return LlmAgent(
+            model=self._get_model(),
+            name=sub_config.name,
+            instruction=sub_config.system_prompt,
+            description=description,
+            tools=sub_toolsets if sub_toolsets else None,
+        )
+
     def _create_mcp_server(self, server_config: MCPServerConfig) -> Any:
         """
         Create Google ADK-specific MCP toolset instance.
@@ -384,7 +513,7 @@ class GoogleADKAgent(Agent):
         # Reset conversation state
         self.reset_conversation()
 
-        # Close all MCP toolset connections
+        # Close all MCP toolset connections (main agent)
         for toolset in self.toolsets:
             try:
                 await toolset.close()
@@ -399,3 +528,13 @@ class GoogleADKAgent(Agent):
         if self._skill_temp_dir:
             cleanup_temp_directory(self._skill_temp_dir)
             self._skill_temp_dir = None
+
+        # Close all subagent MCP toolset connections
+        for toolset in self.subagent_toolsets:
+            try:
+                await toolset.close()
+            except asyncio.CancelledError:
+                pass
+            except Exception as e:
+                print(f"[WARNING] Failed to cleanup subagent toolset: {e}")
+        self.subagent_toolsets.clear()

decodingtrust_agent_sdk-0.2.4/agent/hermes/__init__.py

@@ -0,0 +1,10 @@
+from .src.agent import HermesAgent
+from .src.mcp_proxy import HermesMCPProxyServer, HermesMCPProxyManager
+from .src.utils import HermesTrajectoryConverter
+
+__all__ = [
+    "HermesAgent",
+    "HermesMCPProxyServer",
+    "HermesMCPProxyManager",
+    "HermesTrajectoryConverter",
+]

decodingtrust_agent_sdk-0.2.4/agent/hermes/example.py

@@ -0,0 +1,143 @@
+#!/usr/bin/env python3
+"""
+Hermes Agent Example - CLI Interface
+
+Run the Hermes (NousResearch/hermes-agent) adapter against a DT-Arena config,
+with support for attack evaluation via tool-description injection and prompt
+injection.
+
+Usage:
+    python agent/hermes/example.py --config path/to/config.yaml
+    python agent/hermes/example.py --config path/to/config.yaml --model openai/gpt-4o
+    python agent/hermes/example.py --config path/to/config.yaml --debug
+
+Prerequisites:
+    - hermes-agent installed in this environment (pip install hermes-agent)
+    - Provider API key in env (OPENAI_API_KEY / ANTHROPIC_API_KEY / OPENROUTER_API_KEY / ...)
+
+Environment Variables:
+    HERMES_TIMEOUT_SECONDS - Full-run subprocess timeout (default: 1000)
+"""
+
+import os
+import sys
+import asyncio
+import argparse
+
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), "../.."))
+
+from dt_arena.src.types.agent import AgentConfig, RuntimeConfig
+from dt_arena.src.types.task import TaskConfig, AttackConfig
+
+from agent.hermes import HermesAgent
+from utils.injection_helpers import (
+    build_tool_injections_from_config,
+    build_skill_injections_from_config,
+    apply_prompt_injections,
+)
+from utils.task_helpers import extract_dataset_path
+
+
+async def main():
+    parser = argparse.ArgumentParser(description="Run Hermes Agent with a DT-Arena config file")
+    parser.add_argument("--config", type=str, required=True,
+                        help="Path to YAML configuration file (Task, Agent, Attack sections)")
+    parser.add_argument("--model", type=str, default="anthropic/claude-sonnet-4.6",
+                        help="Model to use (default: anthropic/claude-sonnet-4.6)")
+    parser.add_argument("--temperature", type=float, default=0.1, help="Sampling temperature")
+    parser.add_argument("--max-turns", type=int, default=30, help="Maximum tool-calling iterations")
+    parser.add_argument("--output-dir", type=str, default=None, help="Output directory")
+    parser.add_argument("--debug", action="store_true", help="Enable debug logging")
+    args = parser.parse_args()
+
+    config_path = args.config
+    if not os.path.exists(config_path):
+        print(f"[ERROR] Configuration file not found: {config_path}")
+        sys.exit(1)
+
+    dataset_path = extract_dataset_path(config_path)
+    base_output_dir = args.output_dir or os.path.join(os.getcwd(), "results")
+    output_dir = os.path.join(base_output_dir, "hermes", dataset_path)
+    os.makedirs(output_dir, exist_ok=True)
+
+    print(f"[INFO] Config: {config_path}")
+    print(f"[INFO] Model: {args.model}")
+    print(f"[INFO] Output: {output_dir}")
+    print("-" * 80)
+
+    try:
+        task_config = TaskConfig.from_yaml(config_path)
+        agent_config = AgentConfig.from_yaml(config_path)
+        attack_config = AttackConfig.from_yaml(config_path)
+
+        mcp_injection = build_tool_injections_from_config(attack_config)
+        skill_injection = build_skill_injections_from_config(attack_config)
+
+        runtime_config = RuntimeConfig(
+            model=args.model,
+            temperature=args.temperature,
+            max_turns=args.max_turns,
+            output_dir=output_dir,
+            mcp_injection=mcp_injection,
+            skill_injection=skill_injection,
+            debug=args.debug,
+        )
+
+        agent = HermesAgent(agent_config=agent_config, runtime_config=runtime_config)
+
+        print(f"\n[TASK] Task ID: {task_config.task_id}")
+        print(f"[TASK] Domain: {task_config.domain}")
+        if attack_config:
+            print(f"[ATTACK] Risk Category: {attack_config.risk_category}")
+            print(f"[ATTACK] Threat Model: {attack_config.threat_model}")
+
+        # Resolve instruction (with prompt injections)
+        is_direct_attack = attack_config and attack_config.threat_model == "direct"
+        if is_direct_attack:
+            user_instruction = apply_prompt_injections(None, attack_config) \
+                or task_config.original_instruction
+        else:
+            user_instruction = apply_prompt_injections(task_config.original_instruction, attack_config)
+
+        if isinstance(user_instruction, str):
+            user_instruction = user_instruction.strip()
+        else:
+            user_instruction = [s.strip() for s in user_instruction]
+
+        print("\n" + "=" * 80)
+        print("[INSTRUCTION]")
+        print(user_instruction)
+        print("=" * 80)
+
+        async with agent:
+            print("\n[INFO] Running Hermes agent...")
+            metadata = {
+                "task_id": task_config.task_id,
+                "domain": task_config.domain,
+                "category": attack_config.risk_category if attack_config else None,
+                "malicious_goal": attack_config.malicious_goal if attack_config else None,
+            }
+            result = await agent.run(user_instruction, metadata=metadata)
+
+            print("\n" + "=" * 80)
+            print("[AGENT RESPONSE]")
+            print(result.final_output)
+            print("=" * 80)
+            print(f"[INFO] Turns: {result.turn_count}")
+            print(f"[INFO] Trace ID: {result.trace_id}")
+            if result.trajectory:
+                print(f"[INFO] Trajectory steps: {len(result.trajectory.data['trajectory'])}")
+            return result
+
+    except Exception as e:
+        print(f"\n[ERROR] {e}")
+        import traceback
+        traceback.print_exc()
+        sys.exit(1)
+
+
+if __name__ == "__main__":
+    try:
+        asyncio.run(main())
+    except KeyboardInterrupt:
+        print("\n[INFO] Interrupted by user")

decodingtrust_agent_sdk-0.2.4/agent/hermes/src/__init__.py

@@ -0,0 +1,14 @@
+"""
+Hermes Agent Implementation - Source Module
+"""
+
+from .agent import HermesAgent
+from .mcp_proxy import HermesMCPProxyServer, HermesMCPProxyManager
+from .utils import HermesTrajectoryConverter
+
+__all__ = [
+    "HermesAgent",
+    "HermesMCPProxyServer",
+    "HermesMCPProxyManager",
+    "HermesTrajectoryConverter",
+]