Refine getting started guide

build(deps): bump requests from 2.32.3 to 2.32.4

.github/workflows/release.yml

@@ -9,7 +9,7 @@ on:
       - 0.*
 
 env:
-  POETRY_VERSION: "1.7.1"
+  POETRY_VERSION: "1.8.5"
 
 jobs:
   if_release:

.github/workflows/test.yml

@@ -7,7 +7,7 @@ on:
     branches: [main]
 
 env:
-  POETRY_VERSION: "1.7.1"
+  POETRY_VERSION: "1.8.5"
   OPENAI_API_KEY: "sk-fake"
 
 jobs:

.gitignore

@@ -17,4 +17,5 @@ inv/
 scripts/
 docx/
 agentic_security.toml
-/venv
+/venv
+*.csv

Dockerfile

@@ -1,5 +1,5 @@
 # Build stage
-FROM python:3.11-slim as builder
+FROM python:3.11-slim AS builder
 
 WORKDIR /app
 
@@ -14,8 +14,15 @@ RUN poetry self add "poetry-plugin-export"
 # Copy only dependency files to leverage Docker layer caching
 COPY pyproject.toml poetry.lock ./
 
+# update lock file to avoid failure
+RUN poetry lock
+
 # Install dependencies
 RUN poetry export -f requirements.txt --without-hashes -o requirements.txt
+
+# Install wheel (required to build packages like fire)
+RUN pip install --upgrade pip setuptools wheel
+
 RUN pip install --no-cache-dir -r requirements.txt
 
 # Runtime stage

Readme.md

@@ -21,9 +21,7 @@
   <a href="https://pypi.org/project/agentic-security/">
     <img alt="PyPI Version" src="https://img.shields.io/pypi/v/agentic-security?style=for-the-badge&logo=pypi&labelColor=000000&color=00CCFF" />
   </a>
-  <a href="https://discord.gg/stw3DfZQ">
-    <img alt="Join Discord" src="https://img.shields.io/badge/Discord-Join%20Us-black?style=for-the-badge&logo=discord&labelColor=000000&color=DD55FF" />
-  </a>
+
 </p>
 
 
@@ -402,6 +400,16 @@ This setup ensures a continuous integration approach towards maintaining securit
 
 The `Module` class is designed to manage prompt processing and interaction with external AI models and tools. It supports fetching, processing, and posting prompts asynchronously for model vulnerabilities. Check out [module.md](https://github.com/msoedov/agentic_security/blob/main/docs/module.md) for details.
 
+
+## MCP server
+
+```shell
+pip install -U mcp
+
+# From cloned directory
+mcp install agentic_security/mcp/main.py
+```
+
 ## Documentation
 
 For more detailed information on how to use Agentic Security, including advanced features and customization options, please refer to the official documentation.
@@ -428,6 +436,7 @@ We’re just getting started! Here’s what’s on the horizon:
 
 Note: All dates are tentative and subject to change based on project progress and priorities.
 
+
 ## 👋 Contributing
 
 Contributions to Agentic Security are welcome! If you'd like to contribute, please follow these steps:

agentic_security/config.py

@@ -4,7 +4,7 @@ import tomli
 
 from agentic_security.logutils import logger
 
-SETTINGS_VERSION = 1
+SETTINGS_VERSION = 2
 
 
 @lru_cache(maxsize=1)
@@ -143,6 +143,13 @@ use_disk_cache = false
 retry = 3
 timeout_connect = 30
 timeout_response = 90
+
+[fuzzer]
+max_prompt_lenght = 2048
+budget_multiplier = 100000000
+initial_optimizer_points = 25
+min_failure_samples = 5
+failure_rate_threshold = 0.5
 """.replace(
                     "$HOST", host
                 )

agentic_security/logutils.py

@@ -22,7 +22,11 @@
 # logger.add(sys.stdout, format=LOG_FORMAT, level="DEBUG", colorize=True)
 import logging
 import logging.config
+import time
+from collections.abc import Callable, Coroutine
+from functools import wraps
 from os import getenv
+from typing import Any, ParamSpec, TypeVar
 
 LOGGER_NAME = None
 
@@ -49,6 +53,16 @@ LOGGING_CONFIG = {
             "handlers": ["rich"],
             "propagate": True,
         },
+        "httpx": {  # Disable httpx logging
+            "level": "WARNING",  # Suppress DEBUG and INFO messages from httpx
+            "handlers": [],
+            "propagate": False,
+        },
+        "uvicorn.access": {  # Disable uvicorn.access logging
+            "level": "WARNING",  # Suppress DEBUG and INFO messages from uvicorn.access
+            "handlers": [],
+            "propagate": False,
+        },
     },
 }
 
@@ -83,3 +97,50 @@ def set_log_level_to_info():
 
 # Set initial log level
 set_log_level_to_info()
+
+
+# Define generic type variables for return type and parameters
+R = TypeVar("R")
+P = ParamSpec("P")
+
+
+def time_execution_sync(
+    additional_text: str = "",
+) -> Callable[[Callable[P, R]], Callable[P, R]]:
+    def decorator(func: Callable[P, R]) -> Callable[P, R]:
+        @wraps(func)
+        def wrapper(*args: P.args, **kwargs: P.kwargs) -> R:
+            start_time = time.time()
+            result = func(*args, **kwargs)
+            execution_time = time.time() - start_time
+            logger.debug(
+                f"{additional_text} Execution time: {execution_time:.2f} seconds"
+            )
+            return result
+
+        return wrapper
+
+    return decorator
+
+
+def time_execution_async(
+    additional_text: str = "",
+) -> Callable[
+    [Callable[P, Coroutine[Any, Any, R]]], Callable[P, Coroutine[Any, Any, R]]
+]:
+    def decorator(
+        func: Callable[P, Coroutine[Any, Any, R]]
+    ) -> Callable[P, Coroutine[Any, Any, R]]:
+        @wraps(func)
+        async def wrapper(*args: P.args, **kwargs: P.kwargs) -> R:
+            start_time = time.time()
+            result = await func(*args, **kwargs)
+            execution_time = time.time() - start_time
+            logger.debug(
+                f"{additional_text} Execution time: {execution_time:.2f} seconds"
+            )
+            return result
+
+        return wrapper
+
+    return decorator

agentic_security/mcp/client.py

@@ -0,0 +1,54 @@
+import asyncio
+
+from mcp import ClientSession, StdioServerParameters
+from mcp.client.stdio import stdio_client
+
+# Create server parameters for stdio connection
+server_params = StdioServerParameters(
+    command="python",  # Executable
+    args=["agentic_security/mcp/main.py"],  # Your server script
+    env=None,  # Optional environment variables
+)
+
+
+async def run() -> None:
+    async with stdio_client(server_params) as (read, write):
+        async with ClientSession(read, write) as session:
+            # Initialize the connection --> connection does not work
+            await session.initialize()
+
+            # List available prompts, resources, and tools --> no avalialbe tools
+            prompts = await session.list_prompts()
+            print(f"Available prompts: {prompts}")
+
+            resources = await session.list_resources()
+            print(f"Available resources: {resources}")
+
+            tools = await session.list_tools()
+            print(f"Available tools: {tools}")
+
+            # Call the echo tool --> echo tool iisue
+            echo_result = await session.call_tool(
+                "echo_tool", arguments={"message": "Hello from client!"}
+            )
+            print(f"Tool result: {echo_result}")
+
+            # # Read the echo resource
+            # echo_content, mime_type = await session.read_resource(
+            #     "echo://Hello_resource"
+            # )
+            # print(f"Resource content: {echo_content}")
+            # print(f"Resource MIME type: {mime_type}")
+
+            # # Get and use the echo prompt
+            # prompt_result = await session.get_prompt(
+            #     "echo_prompt", arguments={"message": "Hello prompt!"}
+            # )
+            # print(f"Prompt result: {prompt_result}")
+
+            # You can perform additional operations here as needed
+            return prompts, resources, tools
+
+
+if __name__ == "__main__":
+    asyncio.run(run())

agentic_security/mcp/main.py

@@ -0,0 +1,109 @@
+import httpx
+from mcp.server.fastmcp import FastMCP
+
+# Initialize MCP server
+mcp = FastMCP(
+    name="Agentic Security MCP Server",
+    description="MCP server to interact with LLM scanning test",
+    dependencies=["httpx"],
+)
+
+# FastAPI Server Configuration
+AGENTIC_SECURITY = "http://0.0.0.0:8718"
+
+
+@mcp.tool()
+async def verify_llm(spec: str) -> dict:
+    """
+    Verify an LLM model specification using the FastAPI server
+
+    Returns:
+        dict: containing the verification result form the FastAPI server
+
+    Args: spect(str):  The specification of the LLM model to verify.
+
+    """
+    url = f"{AGENTIC_SECURITY}/verify"
+    async with httpx.AsyncClient() as client:
+        response = await client.post(url, json={"spec": spec})
+        return response.json()
+
+
+@mcp.tool()
+async def start_scan(
+    llmSpec: str,
+    maxBudget: int,
+    optimize: bool = False,
+    enableMultiStepAttack: bool = False,
+) -> dict:
+    """
+    Start an LLM security scan via the FastAPI server.
+    Returns:
+        dict: The scan initiation result from the FastAPI server.
+
+    Args:
+        llmSpec (str): The specification of the LLM model.
+        maxBudget (int): The maximum budget for the scan.
+        optimize (bool, optional): Whether to enable optimization during scanning. Defaults to False.
+        enableMultiStepAttack (bool, optional): Whether to enable multi-step attack
+
+    """
+    url = f"{AGENTIC_SECURITY}/scan"
+    payload = {
+        "llmSpec": llmSpec,
+        "maxBudget": maxBudget,
+        "datasets": [],
+        "optimize": optimize,
+        "enableMultiStepAttack": enableMultiStepAttack,
+        "probe_datasets": [],
+        "secrets": {},
+    }
+    async with httpx.AsyncClient() as client:
+        response = await client.post(url, json=payload)
+        return response.json()
+
+
+@mcp.tool()
+async def stop_scan() -> dict:
+    """Stop an ongoing scan via the FastAPI server.
+
+    Returns:
+        dict: The confirmation from the FastAPI server that the scan has been stopped.
+    """
+    url = f"{AGENTIC_SECURITY}/stop"
+    async with httpx.AsyncClient() as client:
+        response = await client.post(url)
+        return response.json()
+
+
+@mcp.tool()
+async def get_data_config() -> list:
+    """
+    Retrieve data configuration from the FastAPI server.
+
+    Returns:
+        list: The response from the FastAPI server, confirming the scan has been stopped.
+    """
+    url = f"{AGENTIC_SECURITY}/v1/data-config"
+    async with httpx.AsyncClient() as client:
+        response = await client.get(url)
+        return response.json()
+
+
+@mcp.tool()
+async def get_spec_templates() -> list:
+    """
+    Retrieve data configuration from the FastAPI server.
+
+    Returns:
+        list: The LLM specification templates from the FastAPI server.
+    """
+    url = f"{AGENTIC_SECURITY}/v1/llm-specs"
+    async with httpx.AsyncClient() as client:
+        response = await client.get(url)
+        return response.json()
+
+
+# Run the MCP server
+if __name__ == "__main__":
+    mcp.run()

agentic_security/primitives/__init__.py

@@ -1,5 +1,6 @@
+# noqa
+from agentic_security.primitives.models import CompletionRequest  # noqa
 from agentic_security.primitives.models import (  # noqa
-    CompletionRequest,
     FileProbeResponse,
     LLMInfo,
     Message,

agentic_security/probe_actor/fuzzer.py

@@ -3,32 +3,44 @@ import random
 import time
 from collections.abc import AsyncGenerator
 from json import JSONDecodeError
+from typing import Any
 
 import httpx
-import pandas as pd
 from skopt import Optimizer
 from skopt.space import Real
 
+from agentic_security.config import settings_var
 from agentic_security.http_spec import Modality
 from agentic_security.logutils import logger
 from agentic_security.primitives import Scan, ScanResult
 from agentic_security.probe_actor.cost_module import calculate_cost
 from agentic_security.probe_actor.refusal import refusal_heuristic
+from agentic_security.probe_actor.state import FuzzerState
 from agentic_security.probe_data import audio_generator, image_generator, msj_data
 from agentic_security.probe_data.data import prepare_prompts
 
-# TODO: full log file
-
-MAX_PROMPT_LENGTH = 2048
-BUDGET_MULTIPLIER = 100_000_000
-INITIAL_OPTIMIZER_POINTS = 25
-MIN_FAILURE_SAMPLES = 5
-FAILURE_RATE_THRESHOLD = 0.5
+MAX_PROMPT_LENGTH = settings_var("fuzzer.max_prompt_lenght", 2048)
+BUDGET_MULTIPLIER = settings_var("fuzzer.budget_multiplier", 100000000)
+INITIAL_OPTIMIZER_POINTS = settings_var("fuzzer.initial_optimizer_points", 25)
+MIN_FAILURE_SAMPLES = settings_var("min_failure_samples", 5)
+FAILURE_RATE_THRESHOLD = settings_var("failure_rate_threshold", 0.5)
 
 
 async def generate_prompts(
     prompts: list[str] | AsyncGenerator,
 ) -> AsyncGenerator[str, None]:
+    """
+    Asynchronously generates and yields individual prompts.
+
+    If the input is a list of strings, the function sequentially yields each string.
+    If the input is an asynchronous generator, it forwards each generated prompt.
+
+    Args:
+        prompts (list[str] | AsyncGenerator): A list of strings or an asynchronous generator of prompts.
+
+    Yields:
+        str: An individual prompt from the list or the asynchronous generator.
+    """
     if isinstance(prompts, list):
         for prompt in prompts:
             yield prompt
@@ -37,7 +49,21 @@ async def generate_prompts(
             yield prompt
 
 
-def multi_modality_spec(llm_spec):
+def get_modality_adapter(llm_spec):
+    """
+    Returns the appropriate request adapter based on the modality of the LLM specification.
+
+    Depending on the modality of `llm_spec`, the function selects the corresponding request adapter.
+    If the modality is IMAGE or AUDIO, it returns an adapter for handling the respective type.
+    If the modality is TEXT or an unrecognized type, it returns `llm_spec` as is.
+
+    Args:
+        llm_spec: An object containing modality information for the LLM.
+
+    Returns:
+        RequestAdapter | llm_spec: An instance of the appropriate request adapter
+        or the original `llm_spec` if no adaptation is needed.
+    """
     match llm_spec.modality:
         case Modality.IMAGE:
             return image_generator.RequestAdapter(llm_spec)
@@ -50,40 +76,71 @@ def multi_modality_spec(llm_spec):
 
 
 async def process_prompt(
-    request_factory, prompt, tokens, module_name, refusals, errors, outputs
+    request_factory,
+    prompt: str,
+    tokens: int,
+    module_name: str,
+    fuzzer_state: FuzzerState,
 ) -> tuple[int, bool]:
     """
-    Process a single prompt and update the token count and failure status.
+    Processes a single prompt using the provided request factory and updates tracking lists.
+
+    This function sends the given `prompt` to the `request_factory`, checks for errors, and updates
+    the `tokens`, `refusals`, `errors`, and `outputs` lists accordingly. If the request fails or
+    the response indicates a refusal, the function records the issue and returns the updated token count
+    along with a boolean indicating whether the prompt was refused.
+
+    Args:
+        request_factory: An object with a `fn` method used to send the prompt.
+        prompt (str): The input prompt to be processed.
+        tokens (int): The current token count, which will be updated.
+        module_name (str): The name of the module handling the request.
+        fuzzer_state: State tracking object for the fuzzer
+
+    Returns:
+        tuple[int, bool]: Updated token count and a boolean indicating if the prompt was refused.
     """
     try:
         response = await request_factory.fn(prompt=prompt)
+
+        # Handle HTTP errors
         if response.status_code == 422:
             logger.error(f"Invalid prompt: {prompt}, error=422")
-            errors.append((module_name, prompt, 422, "Invalid prompt"))
+            fuzzer_state.add_error(module_name, prompt, 422, "Invalid prompt")
             return tokens, True
 
         if response.status_code >= 400:
             logger.error(f"HTTP {response.status_code} {response.content=}")
-            errors.append((module_name, prompt, response.status_code, response.text))
+            fuzzer_state.add_error(
+                module_name, prompt, response.status_code, response.text
+            )
             return tokens, True
+
+        # Process successful response
         response_text = response.text
         tokens += len(response_text.split())
 
+        # Check if the response indicates a refusal
         refused = refusal_heuristic(response.json())
         if refused:
-            refusals.append((module_name, prompt, response.status_code, response_text))
+            fuzzer_state.add_refusal(
+                module_name, prompt, response.status_code, response_text
+            )
 
-        outputs.append((module_name, prompt, response_text, refused))
+        fuzzer_state.add_output(module_name, prompt, response_text, refused)
         return tokens, refused
 
     except httpx.RequestError as exc:
         logger.error(f"Request error: {exc}")
-        errors.append((module_name, prompt, "?", str(exc)))
+        fuzzer_state.add_error(module_name, prompt, "?", str(exc))
         return tokens, True
     except JSONDecodeError as json_decode_error:
-        logger.error(f"Jason error: {json_decode_error}")
-        errors.append((module_name, prompt, "?", str(json_decode_error)))
+        logger.error(f"JSON error: {json_decode_error}")
+        fuzzer_state.add_error(module_name, prompt, "?", str(json_decode_error))
         return tokens, True
+    except Exception as e:
+        logger.exception(f"Unexpected error: {e}")
+        return tokens, False
 
 
 async def process_prompt_batch(
@@ -91,14 +148,29 @@ async def process_prompt_batch(
     prompts: list[str],
     tokens: int,
     module_name: str,
-    refusals,
-    errors,
-    outputs,
+    fuzzer_state: FuzzerState,
 ) -> tuple[int, int]:
+    """
+    Processes a batch of prompts asynchronously and aggregates the results.
+
+    This function sends multiple prompts concurrently using `process_prompt`,
+    collects the token count and failure status for each prompt, and returns
+    the total number of tokens processed and the number of failed prompts.
+
+    Args:
+        request_factory: An object with a `fn` method used to send the prompts.
+        prompts (list[str]): A list of input prompts to be processed.
+        tokens (int): The initial token count, which will be updated.
+        module_name (str): The name of the module handling the request.
+        fuzzer_state: State tracking object for the fuzzer
+
+    Returns:
+        tuple[int, int]:
+            - Total number of tokens processed.
+            - Number of failed prompts.
+    """
     tasks = [
-        process_prompt(
-            request_factory, p, tokens, module_name, refusals, errors, outputs
-        )
+        process_prompt(request_factory, p, tokens, module_name, fuzzer_state)
         for p in prompts
     ]
     results = await asyncio.gather(*tasks)
@@ -107,7 +179,154 @@ async def process_prompt_batch(
     return total_tokens, failures
 
 
+async def scan_module(
+    request_factory,
+    module,
+    fuzzer_state: FuzzerState,
+    processed_prompts: int = 0,
+    total_prompts: int = 0,
+    max_budget: int = 0,
+    total_tokens: int = 0,
+    optimize: bool = False,
+    stop_event: asyncio.Event | None = None,
+) -> AsyncGenerator[dict[str, Any], None]:
+    """
+    Scan a single module.
+
+    Args:
+        request_factory: The factory for creating requests
+        module: The prompt module to scan
+        fuzzer_state: State tracking object for the fuzzer
+        processed_prompts: Number of prompts processed so far
+        total_prompts: Total number of prompts to process
+        max_budget: Maximum token budget
+        total_tokens: Current token count
+        optimize: Whether to use optimization
+        stop_event: Event to stop scanning
+
+    Yields:
+        ScanResult objects as the scan progresses
+    """
+    tokens = 0
+    module_failures = 0
+    module_prompts = 0
+    failure_rates = []
+    should_stop = False
+
+    # Initialize optimizer if optimization is enabled
+    optimizer = (
+        Optimizer(
+            [Real(0, 1)], base_estimator="GP", n_initial_points=INITIAL_OPTIMIZER_POINTS
+        )
+        if optimize
+        else None
+    )
+
+    module_size = 0 if module.lazy else len(module.prompts)
+    logger.info(f"Scanning {module.dataset_name} {module_size}")
+    yield ScanResult(
+        module=module.dataset_name,
+        tokens=0,
+        cost=0,
+        progress=0,
+        failureRate=0,
+        prompt="",
+        latency=0,
+        model="",
+    ).model_dump_json()
+
+    async for prompt in generate_prompts(module.prompts):
+        if stop_event and stop_event.is_set():
+            stop_event.clear()
+            logger.info("Scan stopped by user.")
+            yield ScanResult.status_msg("Scan stopped by user.")
+            return
+
+        processed_prompts += 1
+        module_prompts += 1
+
+        # Calculate progress based on total processed prompts
+        progress = 100 * processed_prompts / total_prompts if total_prompts else 0
+        progress = progress % 100
+
+        total_tokens -= tokens
+        start = time.time()
+
+        tokens, failed = await process_prompt(
+            request_factory,
+            prompt,
+            tokens,
+            module.dataset_name,
+            fuzzer_state=fuzzer_state,
+        )
+
+        end = time.time()
+        total_tokens += tokens
+
+        if failed:
+            module_failures += 1
+
+        failure_rate = module_failures / max(module_prompts, 1)
+        failure_rates.append(failure_rate)
+        cost = calculate_cost(tokens)
+
+        response_text = fuzzer_state.get_last_output(prompt) or ""
+
+        yield ScanResult(
+            module=module.dataset_name,
+            tokens=round(tokens / 1000, 1),
+            cost=cost,
+            progress=round(progress, 2),
+            failureRate=round(failure_rate * 100, 2),
+            prompt=prompt[:MAX_PROMPT_LENGTH],
+            latency=end - start,
+            model=response_text,
+        ).model_dump_json()
+
+        # Optimization logic
+        if optimize and optimizer and len(failure_rates) >= MIN_FAILURE_SAMPLES:
+            next_point = optimizer.ask()
+            optimizer.tell(next_point, -failure_rate)
+            best_failure_rate = -optimizer.get_result().fun
+            if best_failure_rate > FAILURE_RATE_THRESHOLD:
+                yield ScanResult.status_msg(
+                    f"High failure rate detected ({best_failure_rate:.2%}). Stopping this module..."
+                )
+                should_stop = True
+                break
+
+        # Budget check
+        if total_tokens > max_budget:
+            logger.info(
+                f"Scan ran out of budget and stopped. {total_tokens=} {max_budget=}"
+            )
+            yield ScanResult.status_msg(
+                f"Scan ran out of budget and stopped. {total_tokens=} {max_budget=}"
+            )
+            should_stop = True
+            break
+
+        if should_stop:
+            break
+
+    return
+
+
 async def with_error_handling(agen):
+    """
+    Wraps an asynchronous generator with error handling.
+
+    This function iterates over an asynchronous generator, yielding its values.
+    If an exception occurs, it logs the error and yields a failure message.
+    Finally, it ensures that a completion message is always yielded.
+
+    Args:
+        agen: An asynchronous generator that produces scan results.
+
+    Yields:
+        ScanResult: Either a successful result, an error message if an
+        exception occurs, or a completion message at the end.
+    """
     try:
         async for t in agen:
             yield t
@@ -123,14 +342,37 @@ async def perform_single_shot_scan(
     max_budget: int,
     datasets: list[dict[str, str]] = [],
     tools_inbox=None,
-    optimize=False,
-    stop_event: asyncio.Event = None,
+    optimize: bool = False,
+    stop_event: asyncio.Event | None = None,
     secrets: dict[str, str] = {},
 ) -> AsyncGenerator[str, None]:
-    """Perform a standard security scan."""
+    """
+    Perform a standard security scan using a given request factory.
+
+    This function processes security scan prompts from selected datasets while
+    respecting a predefined token budget. It supports optimization, failure tracking,
+    and early stopping based on budget constraints or user intervention.
+
+    Args:
+        request_factory: A factory function that generates requests for processing prompts.
+        max_budget (int): The maximum token budget for the scan.
+        datasets (list[dict[str, str]], optional): A list of datasets containing security prompts.
+        tools_inbox: Optional additional tools for processing (default: None).
+        optimize (bool, optional): Whether to enable failure rate optimization (default: False).
+        stop_event (asyncio.Event, optional): An event to signal early termination (default: None).
+        secrets (dict[str, str], optional): A dictionary of secrets for authentication (default: {}).
+
+    Yields:
+        str: JSON-encoded scan results or status messages.
+
+    The function iterates over prompts, processes them asynchronously, and updates
+    failure statistics and token usage. If the scan exceeds the budget or failure rate is too high,
+    it stops execution. Results are saved to a CSV file upon completion.
+    """
     max_budget = max_budget * BUDGET_MULTIPLIER
     selected_datasets = [m for m in datasets if m["selected"]]
-    request_factory = multi_modality_spec(request_factory)
+    request_factory = get_modality_adapter(request_factory)
+
     yield ScanResult.status_msg("Loading datasets...")
     prompt_modules = prepare_prompts(
         dataset_names=[m["dataset_name"] for m in selected_datasets],
@@ -140,108 +382,35 @@ async def perform_single_shot_scan(
     )
     yield ScanResult.status_msg("Datasets loaded. Starting scan...")
 
-    errors = []
-    refusals = []
-    outputs = []
+    fuzzer_state = FuzzerState()
     total_prompts = sum(len(m.prompts) for m in prompt_modules if not m.lazy)
     processed_prompts = 0
 
-    optimizer = (
-        Optimizer([Real(0, 1)], base_estimator="GP", n_initial_points=25)
-        if optimize
-        else None
-    )
-    failure_rates = []
-
     total_tokens = 0
-    tokens = 0
-    should_stop = False
     for module in prompt_modules:
-        if should_stop:
-            break
-        tokens = 0
-        module_failures = 0
+        module_gen = scan_module(
+            request_factory=request_factory,
+            module=module,
+            fuzzer_state=fuzzer_state,
+            processed_prompts=processed_prompts,
+            total_prompts=total_prompts,
+            max_budget=max_budget,
+            total_tokens=total_tokens,
+            optimize=optimize,
+            stop_event=stop_event,
+        )
+        try:
+            async for result in module_gen:
+                yield result
+        except Exception:
+            logger.error("Module exception")
+            continue
+        # Update processed_prompts count
         module_size = 0 if module.lazy else len(module.prompts)
-        logger.info(f"Scanning {module.dataset_name} {module_size}")
-        module_prompts = 0  # Reset for each module
-
-        async for prompt in generate_prompts(module.prompts):
-            if stop_event and stop_event.is_set():
-                stop_event.clear()
-                logger.info("Scan stopped by user.")
-                yield ScanResult.status_msg("Scan stopped by user.")
-                return
-
-            processed_prompts += 1
-            module_prompts += 1  # Fixed increment syntax
-            # Calculate progress based on total processed prompts
-            progress = 100 * processed_prompts / total_prompts if total_prompts else 0
-            progress = progress % 100
-
-            total_tokens -= tokens
-            start = time.time()
-            tokens, failed = await process_prompt(
-                request_factory,
-                prompt,
-                tokens,
-                module.dataset_name,
-                refusals,
-                errors,
-                outputs,
-            )
-            end = time.time()
-            total_tokens += tokens
-
-            if failed:
-                module_failures += 1
-            failure_rate = module_failures / max(module_prompts, 1)
-            failure_rates.append(failure_rate)
-            cost = calculate_cost(tokens)
-
-            last_output = outputs[-1] if outputs else None
-            if last_output and last_output[1] == prompt:
-                response_text = last_output[2]
-            else:
-                response_text = ""
-
-            yield ScanResult(
-                module=module.dataset_name,
-                tokens=round(tokens / 1000, 1),
-                cost=cost,
-                progress=round(progress, 2),
-                failureRate=round(failure_rate * 100, 2),
-                prompt=prompt[:MAX_PROMPT_LENGTH],
-                latency=end - start,
-                model=response_text,
-            ).model_dump_json()
-
-            if optimize and len(failure_rates) >= 5:
-                next_point = optimizer.ask()
-                optimizer.tell(next_point, -failure_rate)
-                best_failure_rate = -optimizer.get_result().fun
-                if best_failure_rate > 0.5:
-                    yield ScanResult.status_msg(
-                        f"High failure rate detected ({best_failure_rate:.2%}). Stopping this module..."
-                    )
-                    should_stop = True
-                    break
-            if total_tokens > max_budget:
-                logger.info(
-                    f"Scan ran out of budget and stopped. {total_tokens=} {max_budget=}"
-                )
-                yield ScanResult.status_msg(
-                    f"Scan ran out of budget and stopped. {total_tokens=} {max_budget=}"
-                )
-                should_stop = True
-                break
+        processed_prompts += module_size
 
     yield ScanResult.status_msg("Scan completed.")
-
-    failure_data = errors + refusals
-    df = pd.DataFrame(
-        failure_data, columns=["module", "prompt", "status_code", "content"]
-    )
-    df.to_csv("failures.csv", index=False)
+    fuzzer_state.export_failures("failures.csv")
 
 
 async def perform_many_shot_scan(
@@ -250,14 +419,39 @@ async def perform_many_shot_scan(
     datasets: list[dict[str, str]] = [],
     probe_datasets: list[dict[str, str]] = [],
     tools_inbox=None,
-    optimize=False,
-    stop_event: asyncio.Event = None,
+    optimize: bool = False,
+    stop_event: asyncio.Event | None = None,
     probe_frequency: float = 0.2,
     max_ctx_length: int = 10_000,
     secrets: dict[str, str] = {},
 ) -> AsyncGenerator[str, None]:
-    """Perform a multi-step security scan with probe injection."""
-    request_factory = multi_modality_spec(request_factory)
+    """
+    Perform a multi-step security scan with probe injection.
+
+    This function executes a security scan while periodically injecting probe datasets
+    to test system robustness. It tracks failures, optimizes scan efficiency,
+    and ensures adherence to a predefined token budget.
+
+    Args:
+        request_factory: A factory function that generates requests for processing prompts.
+        max_budget (int): The maximum token budget for the scan.
+        datasets (list[dict[str, str]], optional): The main datasets for scanning.
+        probe_datasets (list[dict[str, str]], optional): Additional datasets for probe injection.
+        tools_inbox: Optional tools for additional processing (default: None).
+        optimize (bool, optional): Whether to enable failure rate optimization (default: False).
+        stop_event (asyncio.Event, optional): An event to signal early termination (default: None).
+        probe_frequency (float, optional): The probability of probe injection (default: 0.2).
+        max_ctx_length (int, optional): The maximum context length before resetting (default: 10,000 tokens).
+        secrets (dict[str, str], optional): A dictionary of secrets for authentication (default: {}).
+
+    Yields:
+        str: JSON-encoded scan results or status messages.
+
+    This function iterates over prompts, injects probe prompts at random intervals,
+    processes them asynchronously, and tracks failure rates. If failure rates exceed a threshold
+    or budget is exhausted, the scan is stopped early. Results are saved to a CSV file upon completion.
+    """
+    request_factory = get_modality_adapter(request_factory)
     # Load main and probe datasets
     yield ScanResult.status_msg("Loading datasets...")
     prompt_modules = prepare_prompts(
@@ -269,17 +463,10 @@ async def perform_many_shot_scan(
     msj_modules = msj_data.prepare_prompts(probe_datasets)
     yield ScanResult.status_msg("Datasets loaded. Starting scan...")
 
-    errors = []
-    refusals = []
-    outputs = []
+    fuzzer_state = FuzzerState()
     total_prompts = sum(len(m.prompts) for m in prompt_modules if not m.lazy)
     processed_prompts = 0
 
-    optimizer = (
-        Optimizer([Real(0, 1)], base_estimator="GP", n_initial_points=25)
-        if optimize
-        else None
-    )
     failure_rates = []
 
     for module in prompt_modules:
@@ -293,6 +480,7 @@ async def perform_many_shot_scan(
                 logger.info("Scan stopped by user.")
                 yield ScanResult.status_msg("Scan stopped by user.")
                 return
+
             tokens = 0
             processed_prompts += 1
             progress = 100 * processed_prompts / total_prompts if total_prompts else 0
@@ -320,9 +508,7 @@ async def perform_many_shot_scan(
                     full_prompt,
                     tokens,
                     module.dataset_name,
-                    refusals,
-                    errors,
-                    outputs,
+                    fuzzer_state=fuzzer_state,
                 )
                 if failed:
                     module_failures += 1
@@ -343,30 +529,48 @@ async def perform_many_shot_scan(
                 prompt=prompt[:MAX_PROMPT_LENGTH],
             ).model_dump_json()
 
-            if optimize and len(failure_rates) >= 5:
-                next_point = optimizer.ask()
-                optimizer.tell(next_point, -failure_rate)
-                best_failure_rate = -optimizer.get_result().fun
-                if best_failure_rate > 0.5:
-                    yield ScanResult.status_msg(
-                        f"High failure rate detected ({best_failure_rate:.2%}). Stopping this module..."
-                    )
-                    break
+            if optimize and len(failure_rates) >= MIN_FAILURE_SAMPLES:
+                yield ScanResult.status_msg(
+                    f"High failure rate detected ({failure_rate:.2%}). Stopping this module..."
+                )
+                break
 
     yield ScanResult.status_msg("Scan completed.")
-
-    df = pd.DataFrame(
-        errors + refusals, columns=["module", "prompt", "status_code", "content"]
-    )
-    df.to_csv("failures.csv", index=False)
+    fuzzer_state.export_failures("failures.csv")
 
 
 def scan_router(
     request_factory,
     scan_parameters: Scan,
     tools_inbox=None,
-    stop_event: asyncio.Event = None,
+    stop_event: asyncio.Event | None = None,
 ):
+    """
+    Route scan requests to the appropriate scanning function.
+
+    This function determines whether to perform a multi-step or single-shot
+    security scan based on the provided scan parameters.
+
+    Args:
+        request_factory: A factory function to generate requests for processing prompts.
+        scan_parameters (Scan): An object containing the parameters for the scan, including:
+            - enableMultiStepAttack (bool): Whether to perform a multi-step scan.
+            - maxBudget (int): The maximum token budget for the scan.
+            - datasets (list[dict[str, str]]): The datasets to scan.
+            - probe_datasets (list[dict[str, str]], optional): Datasets for probe injection (multi-step only).
+            - optimize (bool): Whether to enable optimization.
+            - secrets (dict[str, str], optional): A dictionary of secrets for authentication.
+        tools_inbox: Optional tools for additional processing (default: None).
+        stop_event (asyncio.Event, optional): An event to signal early termination (default: None).
+
+    Returns:
+        A function wrapped with `with_error_handling`, which executes either:
+        - `perform_many_shot_scan` for multi-step scanning.
+        - `perform_single_shot_scan` for single-shot scanning.
+
+    The function ensures that the appropriate scanning method is chosen based on
+    the `enableMultiStepAttack` flag in `scan_parameters`.
+    """
     if scan_parameters.enableMultiStepAttack:
         return with_error_handling(
             perform_many_shot_scan(

agentic_security/probe_actor/state.py

@@ -0,0 +1,47 @@
+import pandas as pd
+
+
+class FuzzerState:
+    """Container for tracking scan results"""
+
+    def __init__(self):
+        self.errors = []
+        self.refusals = []
+        self.outputs = []
+
+    def add_error(
+        self,
+        module_name: str,
+        prompt: str,
+        status_code: int | str,
+        error_msg: str,
+    ):
+        """Add an error to the state"""
+        self.errors.append((module_name, prompt, status_code, error_msg))
+
+    def add_refusal(
+        self, module_name: str, prompt: str, status_code: int, response_text: str
+    ):
+        """Add a refusal to the state"""
+        self.refusals.append((module_name, prompt, status_code, response_text))
+
+    def add_output(
+        self, module_name: str, prompt: str, response_text: str, refused: bool
+    ):
+        """Add an output to the state"""
+        self.outputs.append((module_name, prompt, response_text, refused))
+
+    def get_last_output(self, prompt: str) -> str | None:
+        """Get the last output for a given prompt"""
+        for output in reversed(self.outputs):
+            if output[1] == prompt:
+                return output[2]
+        return None
+
+    def export_failures(self, filename: str = "failures.csv"):
+        """Export failures to a CSV file"""
+        failure_data = self.errors + self.refusals
+        df = pd.DataFrame(
+            failure_data, columns=["module", "prompt", "status_code", "content"]
+        )
+        df.to_csv(filename, index=False)

agentic_security/probe_data/__init__.py

@@ -1,4 +1,4 @@
-from .data import load_local_csv
+from .data import load_local_csv, load_local_csv_files
 
 REGISTRY_V0 = [
     {
@@ -484,3 +484,18 @@ REGISTRY = REGISTRY_V0 + [
         "modality": "text",
     },
 ]
+
+for ds in load_local_csv_files():
+    REGISTRY.append(
+        {
+            "dataset_name": ds.dataset_name,
+            "num_prompts": len(ds.prompts),
+            "tokens": ds.prompts,
+            "approx_cost": 0.0,
+            "is_active": True,
+            "source": f"Local file dataset: {ds.metadata['src']}",
+            "selected": False,
+            "url": "",
+            "modality": "text",
+        }
+    )

agentic_security/probe_data/data.py

@@ -3,6 +3,7 @@ import os
 import random
 from collections.abc import Callable, Iterator
 from functools import partial
+from typing import Any, TypeVar
 
 import httpx
 import pandas as pd
@@ -21,15 +22,18 @@ from agentic_security.probe_data.modules import (
 )
 
 # Type aliases for clarity
+T = TypeVar("T")
 FilterFn = Callable[[pd.Series], bool]
 ColumnMappings = dict[str, str]
 DatasetLoader = Callable[[], ProbeDataset]
+TransformFn = Callable[[str], str]
 
 
 # Core data loading utilities
 def fetch_csv_content(url: str) -> str:
     """Fetch CSV content from a URL."""
     response = httpx.get(url)
+    response.raise_for_status()  # Raise exception for bad responses
     return response.content.decode("utf-8")
 
 
@@ -57,7 +61,7 @@ def transform_df(
 
 
 def create_probe_dataset(
-    name: str, prompts: list[str], metadata: dict = None
+    name: str, prompts: list[str], metadata: dict[str, Any] | None = None
 ) -> ProbeDataset:
     """Create a ProbeDataset from prompts."""
     metadata = metadata or {}
@@ -77,14 +81,46 @@ def load_dataset_generic(
     mappings: ColumnMappings | None = None,
     filter_fn: FilterFn | None = None,
     url: str | None = None,
-    metadata: dict | None = None,
+    metadata: dict[str, Any] | None = None,
 ) -> ProbeDataset:
     """Load and process a dataset with flexible configuration."""
-    df = load_df_from_source(url or name, is_url=bool(url))
-    transformed_df = transform_df(df, mappings, filter_fn)
-    prompt_col = mappings.get("prompt", "prompt") if mappings else "prompt"
-    prompts = transformed_df[prompt_col].tolist()
-    return create_probe_dataset(name, prompts, metadata)
+    try:
+        df = load_df_from_source(url or name, is_url=bool(url))
+        transformed_df = transform_df(df, mappings, filter_fn)
+
+        # Determine which column to use as the prompt source
+        prompt_col = None
+        if mappings and "prompt" in mappings:
+            prompt_col = mappings["prompt"]
+        elif "prompt" in transformed_df.columns:
+            prompt_col = "prompt"
+        else:
+            # Try to find a suitable text column
+            text_columns = [
+                col
+                for col in transformed_df.columns
+                if any(
+                    keyword in col.lower()
+                    for keyword in ["prompt", "text", "query", "question"]
+                )
+            ]
+            if text_columns:
+                prompt_col = text_columns[0]
+                logger.info(f"Using column '{prompt_col}' as prompt source")
+            else:
+                logger.error(f"No suitable prompt column found in dataset {name}")
+                return create_probe_dataset(name, [], metadata)
+
+        # Extract prompts and filter out empty ones
+        prompts = [
+            p
+            for p in transformed_df[prompt_col].tolist()
+            if p and isinstance(p, (str, int, float))
+        ]
+        return create_probe_dataset(name, prompts, metadata)
+    except Exception as e:
+        logger.error(f"Error loading dataset {name}: {e}")
+        return create_probe_dataset(name, [], {"error": str(e)})
 
 
 # Dataset-specific configurations
@@ -159,7 +195,7 @@ DATASET_CONFIGS_GENERICS = {
 
 
 # Dataset factory
-def create_dataset_loader(name: str, config: dict) -> DatasetLoader:
+def create_dataset_loader(name: str, config: dict[str, Any]) -> DatasetLoader:
     """Create a dataset loader from configuration."""
     return partial(
         load_dataset_generic,
@@ -167,6 +203,7 @@ def create_dataset_loader(name: str, config: dict) -> DatasetLoader:
         mappings=config.get("mappings"),
         filter_fn=config.get("filter_fn"),
         url=config.get("url"),
+        metadata={"source": name, "config": str(config)},
     )
 
 
@@ -176,39 +213,82 @@ def load_multi_dataset(name: str, sub_datasets: list[str]) -> ProbeDataset:
     """Load and combine multiple sub-datasets."""
     prompts = []
     for sub in sub_datasets:
-        dataset = load_dataset(name, sub)
-        prompts.extend(dataset["train"]["query"])
-    return create_probe_dataset(f"{name}_combined", prompts)
+        try:
+            dataset = load_dataset(name, sub)
+            if "query" in dataset["train"].features:
+                prompts.extend(dataset["train"]["query"])
+            else:
+                logger.warning(f"No 'query' column in {name}/{sub}")
+        except Exception as e:
+            logger.error(f"Error loading {name}/{sub}: {e}")
+
+    return create_probe_dataset(
+        f"{name}_combined", prompts, {"source": name, "sub_datasets": sub_datasets}
+    )
 
 
 @cache_to_disk()
 def load_jailbreak_v28k() -> ProbeDataset:
     """Load JailBreakV-28K dataset."""
-    df = pd.read_csv("hf://datasets/JailbreakV-28K/JailBreakV-28k/JailBreakV_28K.csv")
-    prompts = df["jailbreak_query"].tolist()
-    return create_probe_dataset("JailbreakV-28K/JailBreakV-28k", prompts)
+    try:
+        df = pd.read_csv(
+            "hf://datasets/JailbreakV-28K/JailBreakV-28k/JailBreakV_28K.csv"
+        )
+        prompts = df["jailbreak_query"].tolist()
+        return create_probe_dataset(
+            "JailbreakV-28K/JailBreakV-28k",
+            prompts,
+            {"source": "JailbreakV-28K/JailBreakV-28k"},
+        )
+    except Exception as e:
+        logger.error(f"Error loading JailbreakV-28K: {e}")
+        return create_probe_dataset("JailbreakV-28K/JailBreakV-28k", [])
+
+
+@cache_to_disk(1)
+def file_dataset(file) -> list[str]:
+    prompts = []
+    try:
+        df = pd.read_csv(os.path.join("./datasets", file), encoding_errors="ignore")
+        if "prompt" in df.columns:
+            prompts = df["prompt"].tolist()
+        else:
+            logger.warning(f"File {file} lacks a suitable prompt column")
+    except Exception as e:
+        logger.error(f"Error reading {file}: {e}")
+    return prompts
 
 
-@cache_to_disk()
 def load_local_csv() -> ProbeDataset:
     """Load prompts from local CSV files."""
-    csv_files = [f for f in os.listdir(".") if f.endswith(".csv")]
+    os.makedirs("./datasets", exist_ok=True)
+    csv_files = [f for f in os.listdir("./datasets") if f.endswith(".csv")]
     logger.info(f"Found {len(csv_files)} CSV files: {csv_files}")
 
     prompts = []
     for file in csv_files:
-        try:
-            df = pd.read_csv(file)
-            if "prompt" in df.columns:
-                prompts.extend(df["prompt"].tolist())
-            else:
-                logger.warning(f"File {file} lacks 'prompt' column")
-        except Exception as e:
-            logger.error(f"Error reading {file}: {e}")
-
+        prompts.extend(file_dataset(file))
     return create_probe_dataset("Local CSV", prompts, {"src": str(csv_files)})
 
 
+def load_csv(file: str) -> ProbeDataset:
+    """Load prompts from local CSV files."""
+    prompts = file_dataset(file)
+    return create_probe_dataset(f"fs://{file}", prompts, {"src": str(file)})
+
+
+def load_local_csv_files() -> list[ProbeDataset]:
+    """Load prompts from local CSV files and return a list of ProbeDataset objects."""
+    csv_files = [f for f in os.listdir("./datasets") if f.endswith(".csv")]
+    logger.info(f"Found {len(csv_files)} CSV files: {csv_files}")
+
+    datasets = []
+
+    for file in csv_files:
+        datasets.append(create_probe_dataset(file, file_dataset(file), {"src": file}))
+    return datasets
+
+
 # Stenography transformer
 class StenographyTransformer:
     """Apply stenography transformations to datasets."""
@@ -231,36 +311,67 @@ class StenographyTransformer:
 
     def transform(self) -> Iterator[ProbeDataset]:
         """Generate transformed datasets."""
+        if not self.datasets:
+            logger.warning("No datasets provided to StenographyTransformer")
+            return
+
         for dataset in self.datasets:
-            size = len(dataset.prompts)
-            sample_size = int(size * 0.25)
+            if not dataset.prompts:
+                logger.warning(
+                    f"Dataset {dataset.dataset_name} has no prompts to transform"
+                )
+                continue
+
+            # Convert async generator (if any) to a list
+            prompts = (
+                list(dataset.prompts) if hasattr(dataset.prompts, "__iter__") else []
+            )
+            size = len(prompts)
+
+            if size == 0:
+                continue
+
+            sample_size = max(1, int(size * 0.25))  # At least 1 sample
 
             for name, fn in self.TRANSFORMATIONS.items():
                 logger.info(f"Applying {name} to {dataset.dataset_name}")
-                sampled = random.sample(dataset.prompts, sample_size)
-                transformed = [
-                    f"Follow the instructions in {name}: {fn(p)}" for p in sampled
-                ]
-                yield create_probe_dataset(
-                    f"stenography.{name}({dataset.dataset_name})", transformed
-                )
+                sampled = random.sample(prompts, min(sample_size, size))
+                try:
+                    transformed = [
+                        f"Follow the instructions in {name}: {fn(str(p))}"
+                        for p in sampled
+                    ]
+                    transformed_dataset = create_probe_dataset(
+                        f"stenography.{name}({dataset.dataset_name})",
+                        transformed,
+                        {"source": dataset.dataset_name, "transformation": name},
+                    )
+                    yield transformed_dataset
+                except Exception as e:
+                    logger.error(
+                        f"Error applying {name} to {dataset.dataset_name}: {e}"
+                    )
 
 
 def dataset_from_iterator(
-    name: str, iterator, lazy: bool = False
+    name: str, iterator: Iterator[str], lazy: bool = False
 ) -> list[ProbeDataset]:
     """Convert an iterator into a list of ProbeDataset objects."""
-    prompts = list(iterator) if not lazy else iterator
-    tokens = sum(len(str(s).split()) for s in prompts) if not lazy else 0
-    dataset = ProbeDataset(
-        dataset_name=name,
-        metadata={},
-        prompts=prompts,
-        tokens=tokens,
-        approx_cost=0.0,
-        lazy=lazy,
-    )
-    return [dataset]
+    try:
+        prompts = list(iterator) if not lazy else iterator
+        tokens = sum(len(str(s).split()) for s in prompts) if not lazy else 0
+        dataset = ProbeDataset(
+            dataset_name=name,
+            metadata={"source": name, "lazy": lazy},
+            prompts=prompts,
+            tokens=tokens,
+            approx_cost=0.0,
+            lazy=lazy,
+        )
+        return [dataset]
+    except Exception as e:
+        logger.error(f"Error creating dataset from iterator {name}: {e}")
+        return [create_probe_dataset(name, [], {"error": str(e)})]
 
 
 # Main dataset preparation
@@ -272,6 +383,7 @@ def prepare_prompts(
 ) -> list[ProbeDataset]:
     """Prepare datasets based on names and options."""
     # Base dataset loaders
+    logger.info(f"Preparing datasets: {dataset_names}")
     dataset_loaders = {
         **{k: create_dataset_loader(k, v) for k, v in DATASET_CONFIGS.items()},
         **{k: create_dataset_loader(k, v) for k, v in DATASET_CONFIGS_GENERICS.items()},
@@ -288,28 +400,39 @@ def prepare_prompts(
         ),
         "JailbreakV-28K/JailBreakV-28k": load_jailbreak_v28k,
         "Local CSV": load_local_csv,
+        "Custom CSV": load_local_csv,
     }
 
     # Dynamic dataset loaders
     dynamic_loaders = {
         "AgenticBackend": lambda opts: dataset_from_iterator(
             "AgenticBackend",
-            fine_tuned.Module([], tools_inbox=tools_inbox, opts=opts).apply(),
+            fine_tuned.Module(
+                opts["datasets"], tools_inbox=tools_inbox, opts=opts
+            ).apply(),
             lazy=True,
         ),
-        "Steganography": lambda opts: list(StenographyTransformer([]).transform()),
+        "Steganography": lambda opts: list(
+            StenographyTransformer(opts["datasets"]).transform()
+        ),
         "llm-adaptive-attacks": lambda opts: dataset_from_iterator(
             "llm-adaptive-attacks",
-            adaptive_attacks.Module([], tools_inbox=tools_inbox, opts=opts).apply(),
+            adaptive_attacks.Module(
+                opts["datasets"], tools_inbox=tools_inbox, opts=opts
+            ).apply(),
         ),
         "Garak": lambda opts: dataset_from_iterator(
             "Garak",
-            garak_tool.Module([], tools_inbox=tools_inbox, opts=opts).apply(),
+            garak_tool.Module(
+                opts["datasets"], tools_inbox=tools_inbox, opts=opts
+            ).apply(),
             lazy=True,
         ),
         "Reinforcement Learning Optimization": lambda opts: dataset_from_iterator(
             "Reinforcement Learning Optimization",
-            rl_model.Module([], tools_inbox=tools_inbox, opts=opts).apply(),
+            rl_model.Module(
+                opts["datasets"], tools_inbox=tools_inbox, opts=opts
+            ).apply(),
             lazy=True,
         ),
         "InspectAI": lambda opts: dataset_from_iterator(
@@ -320,28 +443,35 @@ def prepare_prompts(
         "GPT fuzzer": lambda opts: [],
     }
 
-    options = options or [{} for _ in dataset_names]
     datasets = []
+    options = options or [dict(datasets=datasets) for _ in dataset_names]
 
     # Load base datasets
     for name, opts in zip(dataset_names, options):
-        if name in dataset_loaders:
-            logger.info(f"Loading base dataset {name}")
-            try:
-                datasets.append(dataset_loaders[name]())
-            except Exception as e:
-                logger.error(f"Error loading {name}: {e}")
+        if name not in dataset_loaders:
+            continue
+        try:
+            datasets.append(dataset_loaders[name]())
+        except Exception as e:
+            logger.error(f"Error loading {name}: {e}")
 
     # Load dynamic datasets and apply transformations
     for name, opts in zip(dataset_names, options):
-        if name in dynamic_loaders:
-            logger.info(f"Loading dynamic dataset {name}")
-            try:
-                dynamic_result = dynamic_loaders[name](opts)
-                datasets.extend(dynamic_result)
-            except Exception as e:
-                logger.error(f"Error loading dynamic {name}: {e}")
-        elif name == "Steganography":
-            datasets.extend(list(StenographyTransformer(datasets).transform()))
+        if name not in dynamic_loaders:
+            continue
+        logger.info(f"Loading dynamic dataset {name} {opts}")
+        opts["datasets"] = datasets
+        try:
+            dynamic_result = dynamic_loaders[name](opts)
+            datasets.extend(dynamic_result)
+        except Exception as e:
+            logger.exception(f"Error loading dynamic {name}: {e}")
+
+    # Load csv datasets and apply transformations
+    for name, opts in zip(dataset_names, options):
+        if not name.endswith(".csv"):
+            continue
+        logger.info(f"Loading csv dataset {name} {opts}")
+        datasets.append(load_csv(name))
 
     return datasets

agentic_security/probe_data/msj_data.py

@@ -1,6 +1,6 @@
 from dataclasses import dataclass
 
-from cache_to_disk import cache_to_disk
+from cache_to_disk import cache_to_disk  # noqa
 
 
 # TODO: refactor this class to use from .data
@@ -22,7 +22,7 @@ class ProbeDataset:
         }
 
 
-@cache_to_disk()
+# @cache_to_disk(n_days_to_cache=1)
 def load_dataset_generic(name, getter=lambda x: x["train"]["prompt"]):
     from datasets import load_dataset
 

agentic_security/report_chart.py

@@ -26,7 +26,7 @@ def plot_security_report(table: Table) -> io.BytesIO:
     try:
         return _plot_security_report(table=table)
     except (TypeError, ValueError, OverflowError, IndexError, Exception) as e:
-        logger.error(f"Error in generating the security report: {e}")
+        logger.error(f"Error in generating the security report: {e} {table}")
     return io.BytesIO()
 
 
@@ -40,11 +40,7 @@ def generate_identifiers(data: pd.DataFrame) -> list[str]:
     Returns:
         list[str]: A list of generated identifiers. Returns a list with an empty string in case of an error.
     """
-    try:
-        _generate_identifiers(data=data)
-    except (TypeError, ValueError, Exception) as e:
-        logger.error(f"Error in generate_identifiers: {e}")
-    return [""]
+    return _generate_identifiers(data=data)
 
 
 def _plot_security_report(table: Table) -> io.BytesIO:

agentic_security/routes/_specs.py

@@ -0,0 +1,397 @@
+_SPECS = [
+    """POST ${SELF_URL}/v1/self-probe
+Authorization: Bearer XXXXX
+Content-Type: application/json
+
+{
+"prompt": "<<PROMPT>>"
+}
+
+""",
+    """POST https://api.openai.com/v1/chat/completions
+Authorization: Bearer $OPENAI_API_KEY
+Content-Type: application/json
+
+{
+"model": "gpt-3.5-turbo",
+"messages": [{"role": "user", "content": "<<PROMPT>>"}],
+"temperature": 0.7
+}
+""",
+    """
+POST https://api.deepseek.com/chat/completions
+Authorization: Bearer $DEEPSEEK_API_KEY
+Content-Type: application/json
+
+{
+  "model": "deepseek-chat",
+  "messages": [
+    {"role": "system", "content": "You are a helpful assistant."},
+    {"role": "user", "content": "<<PROMPT>>"}
+  ],
+  "stream": false
+}
+""",
+    """POST https://api.replicate.com/v1/models/mistralai/mixtral-8x7b-instruct-v0.1/predictions
+Authorization: Bearer $APIKEY
+Content-Type: application/json
+
+{
+"input": {
+"top_k": 50,
+"top_p": 0.9,
+"prompt": "Write a bedtime story about neural networks I can read to my toddler",
+"temperature": 0.6,
+"max_new_tokens": 1024,
+"prompt_template": "<s>[INST] <<PROMPT>> [/INST] ",
+"presence_penalty": 0,
+"frequency_penalty": 0
+}
+}
+""",
+    """POST https://api.groq.com/v1/request_manager/text_completion
+Authorization: Bearer $APIKEY
+Content-Type: application/json
+
+{
+"model_id": "codellama-34b",
+"system_prompt": "You are helpful and concise coding assistant",
+"user_prompt": "<<PROMPT>>"
+}
+""",
+    """POST https://api.together.xyz/v1/chat/completions
+Authorization: Bearer $TOGETHER_API_KEY
+Content-Type: application/json
+
+{
+"model": "mistralai/Mixtral-8x7B-Instruct-v0.1",
+"messages": [
+{"role": "system", "content": "You are an expert travel guide"},
+{"role": "user", "content": "<<PROMPT>>"}
+]
+}
+""",
+    """POST ${SELF_URL}/v1/self-probe-image
+Authorization: Bearer XXXXX
+Content-Type: application/json
+
+[
+    {
+      "role": "user",
+      "content": [
+        {
+          "type": "text",
+          "text": "What is in this image?",
+        },
+        {
+          "type": "image_url",
+          "image_url": {
+            "url":  f"data:image/jpeg;base64,{<<BASE64_IMAGE>>}"
+          },
+        },
+      ],
+    }
+]
+""",
+    """POST ${SELF_URL}/v1/self-probe-file
+Authorization: Bearer $API_KEY
+Content-Type: multipart/form-data
+
+{
+  "file": "@./sample_audio.m4a",
+  "model": "whisper-large-v3"
+}
+""",
+    """POST https://api.gemini.com/v1/generate
+Authorization: Bearer $GEMINI_API_KEY
+Content-Type: application/json
+
+{
+  "model": "gemini-latest",
+  "prompt": "<<PROMPT>>",
+  "temperature": 0.8,
+  "max_tokens": 150,
+  "top_p": 1.0,
+  "frequency_penalty": 0,
+  "presence_penalty": 0
+}
+""",
+    """POST https://api.anthropic.com/v1/complete
+Authorization: Bearer $ANTHROPIC_API_KEY
+Content-Type: application/json
+
+{
+  "model": "claude-v1.3",
+  "prompt": "<<PROMPT>>",
+  "temperature": 0.7,
+  "max_tokens_to_sample": 256,
+  "stop_sequences": ["\n\nHuman:"]
+}
+""",
+    """POST https://api.cohere.ai/generate
+Authorization: Bearer $COHERE_API_KEY
+Content-Type: application/json
+
+{
+  "model": "command-xlarge-nightly",
+  "prompt": "<<PROMPT>>",
+  "max_tokens": 300,
+  "temperature": 0.75,
+  "k": 0,
+  "p": 0.75
+}
+""",
+    """POST https://<<RESOURCE_NAME>>.openai.azure.com/openai/deployments/<<DEPLOYMENT_NAME>>/completions?api-version=2023-06-01-preview
+Authorization: Bearer $AZURE_API_KEY
+Content-Type: application/json
+
+{
+  "prompt": "<<PROMPT>>",
+  "max_tokens": 150,
+  "temperature": 0.7,
+  "top_p": 0.9,
+  "frequency_penalty": 0,
+  "presence_penalty": 0
+}
+""",
+    """POST https://api.assemblyai.com/v2/transcript
+Authorization: Bearer $ASSEMBLY_API_KEY
+Content-Type: application/json
+
+{
+  "audio_url": "<<AUDIO_FILE_URL>>"
+}
+""",
+    """POST https://api.openrouter.ai/v1/chat/completions
+Authorization: Bearer $OPENROUTER_API_KEY
+Content-Type: application/json
+
+{
+  "model": "openrouter-latest",
+  "prompt": "<<PROMPT>>",
+  "temperature": 0.7,
+  "max_tokens": 150,
+  "top_p": 0.9,
+  "frequency_penalty": 0,
+  "presence_penalty": 0
+}
+""",
+]
+
+
+LLM_SPECS = [
+    """POST ${SELF_URL}/v1/self-probe
+Authorization: Bearer XXXXX
+Content-Type: application/json
+
+{
+"prompt": "<<PROMPT>>"
+}
+
+""",
+    """POST https://api.openai.com/v1/chat/completions
+Authorization: Bearer $OPENAI_API_KEY
+Content-Type: application/json
+
+{
+"model": "gpt-3.5-turbo",
+"messages": [{"role": "user", "content": "<<PROMPT>>"}],
+"temperature": 0.7
+}
+""",
+    """
+POST https://api.deepseek.com/chat/completions
+Authorization: Bearer $DEEPSEEK_API_KEY
+Content-Type: application/json
+
+{
+  "model": "deepseek-chat",
+  "messages": [
+    {"role": "system", "content": "You are a helpful assistant."},
+    {"role": "user", "content": "<<PROMPT>>"}
+  ],
+  "stream": false
+}
+""",
+    """POST https://api.replicate.com/v1/models/mistralai/mixtral-8x7b-instruct-v0.1/predictions
+Authorization: Bearer $APIKEY
+Content-Type: application/json
+
+{
+"input": {
+"top_k": 50,
+"top_p": 0.9,
+"prompt": "Write a bedtime story about neural networks I can read to my toddler",
+"temperature": 0.6,
+"max_new_tokens": 1024,
+"prompt_template": "<s>[INST] <<PROMPT>> [/INST] ",
+"presence_penalty": 0,
+"frequency_penalty": 0
+}
+}
+""",
+    """POST https://api.groq.com/v1/request_manager/text_completion
+Authorization: Bearer $APIKEY
+Content-Type: application/json
+
+{
+"model_id": "codellama-34b",
+"system_prompt": "You are helpful and concise coding assistant",
+"user_prompt": "<<PROMPT>>"
+}
+""",
+    """POST https://api.together.xyz/v1/chat/completions
+Authorization: Bearer $TOGETHER_API_KEY
+Content-Type: application/json
+
+{
+"model": "mistralai/Mixtral-8x7B-Instruct-v0.1",
+"messages": [
+{"role": "system", "content": "You are an expert travel guide"},
+{"role": "user", "content": "<<PROMPT>>"}
+]
+}
+""",
+    """POST ${SELF_URL}/v1/self-probe-image
+Authorization: Bearer XXXXX
+Content-Type: application/json
+
+[
+    {
+      "role": "user",
+      "content": [
+        {
+          "type": "text",
+          "text": "What is in this image?",
+        },
+        {
+          "type": "image_url",
+          "image_url": {
+            "url":  f"data:image/jpeg;base64,{<<BASE64_IMAGE>>}"
+          },
+        },
+      ],
+    }
+]
+""",
+    """POST ${SELF_URL}/v1/self-probe-file
+Authorization: Bearer $API_KEY
+Content-Type: multipart/form-data
+
+{
+  "file": "@./sample_audio.m4a",
+  "model": "whisper-large-v3"
+}
+""",
+    """POST https://api.gemini.com/v1/generate
+Authorization: Bearer $GEMINI_API_KEY
+Content-Type: application/json
+
+{
+  "model": "gemini-latest",
+  "prompt": "<<PROMPT>>",
+  "temperature": 0.8,
+  "max_tokens": 150,
+  "top_p": 1.0,
+  "frequency_penalty": 0,
+  "presence_penalty": 0
+}
+""",
+    """POST https://api.anthropic.com/v1/complete
+Authorization: Bearer $ANTHROPIC_API_KEY
+Content-Type: application/json
+
+{
+  "model": "claude-v1.3",
+  "prompt": "<<PROMPT>>",
+  "temperature": 0.7,
+  "max_tokens_to_sample": 256,
+  "stop_sequences": ["\n\nHuman:"]
+}
+""",
+    """POST https://api.cohere.ai/generate
+Authorization: Bearer $COHERE_API_KEY
+Content-Type: application/json
+
+{
+  "model": "command-xlarge-nightly",
+  "prompt": "<<PROMPT>>",
+  "max_tokens": 300,
+  "temperature": 0.75,
+  "k": 0,
+  "p": 0.75
+}
+""",
+    """POST https://<<RESOURCE_NAME>>.openai.azure.com/openai/deployments/<<DEPLOYMENT_NAME>>/completions?api-version=2023-06-01-preview
+Authorization: Bearer $AZURE_API_KEY
+Content-Type: application/json
+
+{
+  "prompt": "<<PROMPT>>",
+  "max_tokens": 150,
+  "temperature": 0.7,
+  "top_p": 0.9,
+  "frequency_penalty": 0,
+  "presence_penalty": 0
+}
+""",
+    """POST https://api.assemblyai.com/v2/transcript
+Authorization: Bearer $ASSEMBLY_API_KEY
+Content-Type: application/json
+
+{
+  "audio_url": "<<AUDIO_FILE_URL>>"
+}
+""",
+    """POST https://api.openrouter.ai/v1/chat/completions
+Authorization: Bearer $OPENROUTER_API_KEY
+Content-Type: application/json
+
+{
+  "model": "openrouter-latest",
+  "prompt": "<<PROMPT>>",
+  "temperature": 0.7,
+  "max_tokens": 150,
+  "top_p": 0.9,
+  "frequency_penalty": 0,
+  "presence_penalty": 0
+}
+""",
+]
+
+
+LLM_CONFIGS = [
+    {
+        "name": "Custom API",
+        "prompts": 40000,
+        "customInstructions": "Requires api spec",
+        "logo": "/icons/myshell.png",
+    },
+    {"name": "Open AI", "prompts": 24000, "logo": "/icons/openai.png"},
+    {"name": "Deepseek v1", "prompts": 24000, "logo": "/icons/deepseek.png"},
+    {"name": "Replicate", "prompts": 40000, "logo": "/icons/replicate.png"},
+    {"name": "Groq", "prompts": 40000, "logo": "/icons/groq.png"},
+    {"name": "Together.ai", "prompts": 40000, "logo": "/icons/together.png"},
+    {
+        "name": "Custom API Image",
+        "prompts": 40000,
+        "customInstructions": "Requires api spec",
+        "modality": "Image",
+        "logo": "/icons/myshell.png",
+    },
+    {
+        "name": "Custom API Files",
+        "prompts": 40000,
+        "customInstructions": "Requires api spec",
+        "modality": "Files",
+        "logo": "/icons/myshell.png",
+    },
+    {"name": "Gemini", "prompts": 40000, "logo": "/icons/gemini.png"},
+    {"name": "Claude", "prompts": 40000, "logo": "/icons/claude.png"},
+    {"name": "Cohere", "prompts": 40000, "logo": "/icons/cohere.png"},
+    {"name": "Azure OpenAI", "prompts": 40000, "logo": "/icons/azureai.png"},
+    {"name": "assemblyai", "prompts": 40000, "logo": "/icons/myshell.png"},
+    {"name": "OpenRouter.ai", "prompts": 40000, "logo": "/icons/openrouter.png"},
+]
+
+LLM_SPECS = [dict(spec=spec, **d) for spec, d in zip(_SPECS, LLM_CONFIGS)]

agentic_security/routes/probe.py

@@ -6,6 +6,7 @@ from fastapi.responses import JSONResponse
 from ..primitives import FileProbeResponse, Probe
 from ..probe_actor.refusal import REFUSAL_MARKS
 from ..probe_data import REGISTRY
+from ._specs import LLM_SPECS
 
 router = APIRouter()
 
@@ -73,6 +74,12 @@ async def data_config():
     return [m for m in REGISTRY]
 
 
+@router.get("/v1/llm-specs", response_model=list)
+def get_llm_specs():
+    """Returns the LLM API specifications."""
+    return LLM_SPECS
+
+
 @router.get("/health")
 async def health_check():
     """Health check endpoint."""

agentic_security/routes/scan.py

@@ -17,7 +17,7 @@ from agentic_security.logutils import logger
 
 from ..core.app import get_stop_event, get_tools_inbox, set_current_run
 from ..dependencies import InMemorySecrets, get_in_memory_secrets
-from ..http_spec import LLMSpec
+from ..http_spec import InvalidHTTPSpecError, LLMSpec
 from ..primitives import LLMInfo, Scan
 from ..probe_actor import fuzzer
 
@@ -31,6 +31,8 @@ async def verify(
     spec = LLMSpec.from_string(info.spec)
     try:
         r = await spec.verify()
+    except InvalidHTTPSpecError as e:
+        raise HTTPException(status_code=400, detail=str(e))
     except Exception as e:
         logger.exception(e)
         raise HTTPException(status_code=400, detail=str(e))

agentic_security/static/main.js

@@ -110,19 +110,21 @@ var app = new Vue({
         },
         focusTextarea() {
             this.isFocused = true;
-            self = this.$refs;
+            // Remove 'self' assignment if not used elsewhere
             this.$nextTick(() => {
-                // Focus the textarea after rendering
-                self.textarea.focus();
-                this.adjustHeight({ target: self.textarea });
+                this.$refs.textarea.focus();
+                this.adjustHeight({ target: this.$refs.textarea });
             });
-            document.addEventListener("mousedown", this.handleClickOutside);
-
+            // Correct the event listener to use handleOutsideClick
+            document.addEventListener("mousedown", this.handleOutsideClick);
         },
         handleOutsideClick(event) {
-            if (!this.$refs.container.contains(event.target)) {
+            if (!this.$refs.textarea) {
+                return
+            }
+            if (!this.$refs.textarea.contains(event.target)) {
                 this.isFocused = false;
-                document.removeEventListener("mousedown", this.handleClickOutside);
+                document.removeEventListener("mousedown", this.handleOutsideClick);
             }
         },
         unfocusTextarea() {
@@ -130,7 +132,12 @@ var app = new Vue({
         },
         acceptConsent() {
             this.showConsentModal = false; // Close the modal
-            localStorage.setItem('consentGiven', 'true'); // Save consent to local storage
+
+            try {
+                localStorage.setItem('consentGiven', 'true'); // Save consent to local storage
+            } catch (e) {
+                this.showToast('Failed to save consent', 'error'); // Show error if saving fails
+            }
         },
 
         saveStateToLocalStorage() {
@@ -171,6 +178,7 @@ var app = new Vue({
             this.integrationVerified = false;
             this.showResetConfirmation = false;
             this.enableMultiStepAttack = false;
+            this.showToast('All settings have been reset to default', 'info');
         },
         confirmResetState() {
             this.showResetConfirmation = true;
@@ -209,33 +217,39 @@ var app = new Vue({
                 spec: this.modelSpec,
             };
             let startTime = performance.now(); // Capture start time
-            const response = await fetch(`${SELF_URL}/verify`, {
-                method: 'POST',
-                headers: {
-                    'Content-Type': 'application/json',
-                },
-                body: JSON.stringify(payload),
-            });
-            console.log(response);
-            let r = await response.json();
-            let endTime = performance.now(); // Capture end time
-            let latency = endTime - startTime; // Calculate latency in milliseconds
-            latency = latency.toFixed(3) / 1000; // Round to 2 decimal places
-            this.latency = latency;
-            if (!response.ok) {
-                this.updateStatusDot(false);
-                this.errorMsg = 'Integration verification failed:' + JSON.stringify(r);
-                this.showToast('Integration verification failed', 'error');
-            } else {
-                this.errorMsg = '';
-                this.updateStatusDot(true);
-                this.okMsg = 'Integration verified';
-                this.showToast('Integration verified successfully', 'success');
-                this.integrationVerified = true;
-                // console.log('Integration verified', this.integrationVerified);
-                // this.$forceUpdate();
 
+            try {
+                const response = await fetch(`${SELF_URL}/verify`, {
+                    method: 'POST',
+                    headers: {
+                        'Content-Type': 'application/json',
+                    },
+                    body: JSON.stringify(payload),
+                });
+
+                let r = await response.json();
+
+                let endTime = performance.now(); // Capture end time
+                let latency = ((endTime - startTime) / 1000).toFixed(3); // Calculate latency in milliseconds
+                this.latency = latency;
+
+                if (!response.ok) {
+                    this.updateStatusDot(false);
+                    this.errorMsg = 'Integration verification failed:' + JSON.stringify(r);
+                    this.showToast('Integration verification failed', 'error');
+                } else {
+                    this.errorMsg = '';
+                    this.updateStatusDot(true);
+                    this.okMsg = 'Integration verified';
+                    this.showToast('Integration verified successfully', 'success');
+                    this.integrationVerified = true;
+                }
+            } catch (error) {
+                this.updateStatusDot(true);
+                this.errorMsg = 'Server unreachable';
+                this.showToast('Network error', 'error');
             }
+
             this.saveStateToLocalStorage();
         },
         loadConfigs: async function () {
@@ -257,6 +271,7 @@ var app = new Vue({
             this.errorMsg = '';
             this.okMsg = '';
             this.integrationVerified = false;
+            this.showToast(`Config ${index + 1} selected`, 'info');
         },
         toggleModules() {
             this.showModules = !this.showModules;
@@ -344,6 +359,7 @@ var app = new Vue({
                 return
             }
             console.log('New row');
+            this.showToast('New module', 'success');
             let payload = {
                 table: this.mainTable,
             };
@@ -454,6 +470,8 @@ var app = new Vue({
                     }
                 });
             }
+            this.scanRunning = false;
+            this.showToast('Scan finished successfully', 'success');
             this.saveStateToLocalStorage();
 
         }

docs/getting_started.md

@@ -1,23 +1,54 @@
 # Getting Started
 
-Welcome to Agentic Security! This guide will help you get started with using the tool.
+Welcome to Agentic Security! This guide will help you orient yourself within the project and start using the tool quickly.
+
+## Project Overview
+
+Agentic Security is an open-source vulnerability scanner for Large Language Models (LLMs). It provides both a command line interface and a FastAPI server so you can probe models for weaknesses such as jailbreaks or refusal patterns. The tool supports multimodal attacks, multi-step scans and reinforcement-learning based probes.
+
+## Repository Layout
+
+```
+agentic_security/
+├── __main__.py    - CLI entry point
+├── app.py         - FastAPI app assembly
+├── lib.py         - SecurityScanner and utilities
+├── config.py      - Configuration handling
+├── core/          - app state and logging helpers
+├── probe_actor/   - scanning logic and RL modules
+├── probe_data/    - dataset registry and loaders
+├── routes/        - API endpoints
+└── ui/            - Web UI assets (Vue)
+```
+
+`tests/` contains unit tests, and `docs/` houses the project documentation.
 
 ## Quick Start
 
 1. Ensure you have completed the [installation](installation.md) steps.
-1. Run the following command to start the application:
+2. Run the following command to start the application:
    ```bash
    agentic_security
    ```
-1. Access the application at `http://localhost:8718`.
+   The server will start on `http://localhost:8718`.
+3. Explore available commands with:
+   ```bash
+   agentic_security --help
+   ```
 
 ## Basic Usage
 
-- To view available commands, use:
+- To view available commands, run:
   ```bash
   agentic_security --help
   ```
 
 ## Next Steps
 
-Explore the [Configuration](configuration.md) section to customize your setup.
+- Review the [Quickstart Guide](quickstart.md) for a fast setup walkthrough.
+- Check [http_spec.md](http_spec.md) to learn how LLM endpoints are described.
+- Browse the `probe_actor` and `probe_data` modules to understand how scanning works and how datasets are loaded.
+- Explore the [Configuration](configuration.md) section to customize your setup.
+- Run the tests in `tests/` to verify your environment once dependencies are installed.
+
+This guide should give you a solid foundation for exploring and extending Agentic Security. For more details, see the rest of the documentation.

poetry.lock

@@ -153,24 +153,25 @@ files = [
 
 [[package]]
 name = "anyio"
-version = "4.4.0"
+version = "4.9.0"
 description = "High level compatibility layer for multiple asynchronous event loop implementations"
 optional = false
-python-versions = ">=3.8"
+python-versions = ">=3.9"
 groups = ["main", "dev"]
 files = [
-    {file = "anyio-4.4.0-py3-none-any.whl", hash = "sha256:c1b2d8f46a8a812513012e1107cb0e68c17159a7a594208005a57dc776e1bdc7"},
-    {file = "anyio-4.4.0.tar.gz", hash = "sha256:5aadc6a1bbb7cdb0bede386cac5e2940f5e2ff3aa20277e991cf028e0585ce94"},
+    {file = "anyio-4.9.0-py3-none-any.whl", hash = "sha256:9f76d541cad6e36af7beb62e978876f3b41e3e04f2c1fbf0884604c0a9c4d93c"},
+    {file = "anyio-4.9.0.tar.gz", hash = "sha256:673c0c244e15788651a4ff38710fea9675823028a6f08a5eda409e0c9840a028"},
 ]
 
 [package.dependencies]
 idna = ">=2.8"
 sniffio = ">=1.1"
+typing_extensions = {version = ">=4.5", markers = "python_version < \"3.13\""}
 
 [package.extras]
-doc = ["Sphinx (>=7)", "packaging", "sphinx-autodoc-typehints (>=1.2.0)", "sphinx-rtd-theme"]
-test = ["anyio[trio]", "coverage[toml] (>=7)", "exceptiongroup (>=1.2.0)", "hypothesis (>=4.0)", "psutil (>=5.9)", "pytest (>=7.0)", "pytest-mock (>=3.6.1)", "trustme", "uvloop (>=0.17) ; platform_python_implementation == \"CPython\" and platform_system != \"Windows\""]
-trio = ["trio (>=0.23)"]
+doc = ["Sphinx (>=8.2,<9.0)", "packaging", "sphinx-autodoc-typehints (>=1.2.0)", "sphinx_rtd_theme"]
+test = ["anyio[trio]", "blockbuster (>=1.5.23)", "coverage[toml] (>=7)", "exceptiongroup (>=1.2.0)", "hypothesis (>=4.0)", "psutil (>=5.9)", "pytest (>=7.0)", "trustme", "truststore (>=0.9.1) ; python_version >= \"3.10\"", "uvloop (>=0.21) ; platform_python_implementation == \"CPython\" and platform_system != \"Windows\" and python_version < \"3.14\""]
+trio = ["trio (>=0.26.1)"]
 
 [[package]]
 name = "appnope"
@@ -688,14 +689,14 @@ tests = ["pytest", "pytest-cov", "pytest-xdist"]
 
 [[package]]
 name = "datasets"
-version = "3.3.2"
+version = "3.4.0"
 description = "HuggingFace community-driven open-source library of datasets"
 optional = false
 python-versions = ">=3.9.0"
 groups = ["main"]
 files = [
-    {file = "datasets-3.3.2-py3-none-any.whl", hash = "sha256:fdaf3d5d70242621210b044e9b9b15a56e908bfc3e9d077bcf5605ac390f70bd"},
-    {file = "datasets-3.3.2.tar.gz", hash = "sha256:20901a97da870fb80b407ccc45f034a7ac99accd07da897ed42f11641bdb8c6e"},
+    {file = "datasets-3.4.0-py3-none-any.whl", hash = "sha256:35ef5182bddd38f7aa774d9f33c3e8b8e9c9c7ea41b4b7969fde431919cb556b"},
+    {file = "datasets-3.4.0.tar.gz", hash = "sha256:f3defae5d9c79ff586db3b17389fdde01704ffea015293a050d7e8ab6816bad8"},
 ]
 
 [package.dependencies]
@@ -717,15 +718,15 @@ xxhash = "*"
 [package.extras]
 audio = ["librosa", "soundfile (>=0.12.1)", "soxr (>=0.4.0) ; python_version >= \"3.9\""]
 benchmarks = ["tensorflow (==2.12.0)", "torch (==2.0.1)", "transformers (==4.30.1)"]
-dev = ["Pillow (>=9.4.0)", "absl-py", "decorator", "decord (==0.6.0)", "elasticsearch (>=7.17.12,<8.0.0)", "faiss-cpu (>=1.8.0.post1)", "jax (>=0.3.14) ; sys_platform != \"win32\"", "jaxlib (>=0.3.14) ; sys_platform != \"win32\"", "joblib (<1.3.0)", "joblibspark", "librosa", "lz4", "moto[server]", "polars[timezone] (>=0.20.0)", "protobuf (<4.0.0)", "py7zr", "pyspark (>=3.4)", "pytest", "pytest-datadir", "pytest-xdist", "rarfile (>=4.0)", "ruff (>=0.3.0)", "s3fs", "s3fs (>=2021.11.1)", "soundfile (>=0.12.1)", "soundfile (>=0.12.1)", "soxr (>=0.4.0) ; python_version >= \"3.9\"", "sqlalchemy", "tensorflow (>=2.16.0) ; python_version >= \"3.10\"", "tensorflow (>=2.6.0)", "tensorflow (>=2.6.0) ; python_version < \"3.10\"", "tiktoken", "torch", "torch (>=2.0.0)", "torchdata", "transformers", "transformers (>=4.42.0)", "zstandard"]
+dev = ["Pillow (>=9.4.0)", "absl-py", "decorator", "elasticsearch (>=7.17.12,<8.0.0)", "faiss-cpu (>=1.8.0.post1)", "jax (>=0.3.14) ; sys_platform != \"win32\"", "jaxlib (>=0.3.14) ; sys_platform != \"win32\"", "joblib (<1.3.0)", "joblibspark", "librosa", "lz4", "moto[server]", "polars[timezone] (>=0.20.0)", "protobuf (<4.0.0)", "py7zr", "pyav", "pyspark (>=3.4)", "pytest", "pytest-datadir", "pytest-xdist", "rarfile (>=4.0)", "ruff (>=0.3.0)", "s3fs", "s3fs (>=2021.11.1)", "soundfile (>=0.12.1)", "soundfile (>=0.12.1)", "soxr (>=0.4.0) ; python_version >= \"3.9\"", "sqlalchemy", "tensorflow (>=2.16.0) ; python_version >= \"3.10\"", "tensorflow (>=2.6.0)", "tensorflow (>=2.6.0) ; python_version < \"3.10\"", "tiktoken", "torch", "torch (>=2.0.0)", "torchdata", "torchvision", "transformers", "transformers (>=4.42.0)", "zstandard"]
 docs = ["s3fs", "tensorflow (>=2.6.0)", "torch", "transformers"]
 jax = ["jax (>=0.3.14)", "jaxlib (>=0.3.14)"]
 quality = ["ruff (>=0.3.0)"]
 s3 = ["s3fs"]
 tensorflow = ["tensorflow (>=2.6.0)"]
 tensorflow-gpu = ["tensorflow (>=2.6.0)"]
-tests = ["Pillow (>=9.4.0)", "absl-py", "decorator", "decord (==0.6.0)", "elasticsearch (>=7.17.12,<8.0.0)", "faiss-cpu (>=1.8.0.post1)", "jax (>=0.3.14) ; sys_platform != \"win32\"", "jaxlib (>=0.3.14) ; sys_platform != \"win32\"", "joblib (<1.3.0)", "joblibspark", "librosa", "lz4", "moto[server]", "polars[timezone] (>=0.20.0)", "protobuf (<4.0.0)", "py7zr", "pyspark (>=3.4)", "pytest", "pytest-datadir", "pytest-xdist", "rarfile (>=4.0)", "s3fs (>=2021.11.1)", "soundfile (>=0.12.1)", "soundfile (>=0.12.1)", "soxr (>=0.4.0) ; python_version >= \"3.9\"", "sqlalchemy", "tensorflow (>=2.16.0) ; python_version >= \"3.10\"", "tensorflow (>=2.6.0) ; python_version < \"3.10\"", "tiktoken", "torch (>=2.0.0)", "torchdata", "transformers (>=4.42.0)", "zstandard"]
-tests-numpy2 = ["Pillow (>=9.4.0)", "absl-py", "decorator", "decord (==0.6.0)", "elasticsearch (>=7.17.12,<8.0.0)", "jax (>=0.3.14) ; sys_platform != \"win32\"", "jaxlib (>=0.3.14) ; sys_platform != \"win32\"", "joblib (<1.3.0)", "joblibspark", "lz4", "moto[server]", "polars[timezone] (>=0.20.0)", "protobuf (<4.0.0)", "py7zr", "pyspark (>=3.4)", "pytest", "pytest-datadir", "pytest-xdist", "rarfile (>=4.0)", "s3fs (>=2021.11.1)", "soundfile (>=0.12.1)", "soundfile (>=0.12.1)", "soxr (>=0.4.0) ; python_version >= \"3.9\"", "sqlalchemy", "tiktoken", "torch (>=2.0.0)", "torchdata", "transformers (>=4.42.0)", "zstandard"]
+tests = ["Pillow (>=9.4.0)", "absl-py", "decorator", "elasticsearch (>=7.17.12,<8.0.0)", "faiss-cpu (>=1.8.0.post1)", "jax (>=0.3.14) ; sys_platform != \"win32\"", "jaxlib (>=0.3.14) ; sys_platform != \"win32\"", "joblib (<1.3.0)", "joblibspark", "librosa", "lz4", "moto[server]", "polars[timezone] (>=0.20.0)", "protobuf (<4.0.0)", "py7zr", "pyav", "pyspark (>=3.4)", "pytest", "pytest-datadir", "pytest-xdist", "rarfile (>=4.0)", "s3fs (>=2021.11.1)", "soundfile (>=0.12.1)", "soundfile (>=0.12.1)", "soxr (>=0.4.0) ; python_version >= \"3.9\"", "sqlalchemy", "tensorflow (>=2.16.0) ; python_version >= \"3.10\"", "tensorflow (>=2.6.0) ; python_version < \"3.10\"", "tiktoken", "torch (>=2.0.0)", "torchdata", "torchvision", "transformers (>=4.42.0)", "zstandard"]
+tests-numpy2 = ["Pillow (>=9.4.0)", "absl-py", "decorator", "elasticsearch (>=7.17.12,<8.0.0)", "jax (>=0.3.14) ; sys_platform != \"win32\"", "jaxlib (>=0.3.14) ; sys_platform != \"win32\"", "joblib (<1.3.0)", "joblibspark", "lz4", "moto[server]", "polars[timezone] (>=0.20.0)", "protobuf (<4.0.0)", "py7zr", "pyav", "pyspark (>=3.4)", "pytest", "pytest-datadir", "pytest-xdist", "rarfile (>=4.0)", "s3fs (>=2021.11.1)", "soundfile (>=0.12.1)", "soundfile (>=0.12.1)", "soxr (>=0.4.0) ; python_version >= \"3.9\"", "sqlalchemy", "tiktoken", "torch (>=2.0.0)", "torchdata", "torchvision", "transformers (>=4.42.0)", "zstandard"]
 torch = ["torch"]
 vision = ["Pillow (>=9.4.0)"]
 
@@ -1207,16 +1208,28 @@ http2 = ["h2 (>=3,<5)"]
 socks = ["socksio (==1.*)"]
 zstd = ["zstandard (>=0.18.0)"]
 
+[[package]]
+name = "httpx-sse"
+version = "0.4.0"
+description = "Consume Server-Sent Event (SSE) messages with HTTPX."
+optional = false
+python-versions = ">=3.8"
+groups = ["main"]
+files = [
+    {file = "httpx-sse-0.4.0.tar.gz", hash = "sha256:1e81a3a3070ce322add1d3529ed42eb5f70817f45ed6ec915ab753f961139721"},
+    {file = "httpx_sse-0.4.0-py3-none-any.whl", hash = "sha256:f329af6eae57eaa2bdfd962b42524764af68075ea87370a2de920af5341e318f"},
+]
+
 [[package]]
 name = "huggingface-hub"
-version = "0.28.1"
+version = "0.29.3"
 description = "Client library to download and publish models, datasets and other repos on the huggingface.co hub"
 optional = false
 python-versions = ">=3.8.0"
 groups = ["main", "dev"]
 files = [
-    {file = "huggingface_hub-0.28.1-py3-none-any.whl", hash = "sha256:aa6b9a3ffdae939b72c464dbb0d7f99f56e649b55c3d52406f49e0a5a620c0a7"},
-    {file = "huggingface_hub-0.28.1.tar.gz", hash = "sha256:893471090c98e3b6efbdfdacafe4052b20b84d59866fb6f54c33d9af18c303ae"},
+    {file = "huggingface_hub-0.29.3-py3-none-any.whl", hash = "sha256:0b25710932ac649c08cdbefa6c6ccb8e88eef82927cacdb048efb726429453aa"},
+    {file = "huggingface_hub-0.29.3.tar.gz", hash = "sha256:64519a25716e0ba382ba2d3fb3ca082e7c7eb4a2fc634d200e8380006e0760e5"},
 ]
 
 [package.dependencies]
@@ -1283,14 +1296,14 @@ files = [
 
 [[package]]
 name = "inline-snapshot"
-version = "0.20.3"
+version = "0.20.9"
 description = "golden master/snapshot/approval testing library which puts the values right into your source code"
 optional = false
 python-versions = ">=3.8"
 groups = ["dev"]
 files = [
-    {file = "inline_snapshot-0.20.3-py3-none-any.whl", hash = "sha256:1ea999fbf38dd11cc72d0e1a0b9303c63d496b77bdc406a394fe2424ae842f70"},
-    {file = "inline_snapshot-0.20.3.tar.gz", hash = "sha256:7a353170b7e42aa89086c7ba790a973c9645523acf985532648dabd7ee2d71f2"},
+    {file = "inline_snapshot-0.20.9-py3-none-any.whl", hash = "sha256:e4bffcfb39d79bba0320a138e47217601823bbf6948760d75fa175d62a7b893d"},
+    {file = "inline_snapshot-0.20.9.tar.gz", hash = "sha256:ce41505690f0f6ca96466776bc3e015ccc7412ecb65b74f4e7e90f1a834cbac4"},
 ]
 
 [package.dependencies]
@@ -1864,6 +1877,33 @@ files = [
 [package.dependencies]
 traitlets = "*"
 
+[[package]]
+name = "mcp"
+version = "1.6.0"
+description = "Model Context Protocol SDK"
+optional = false
+python-versions = ">=3.10"
+groups = ["main"]
+files = [
+    {file = "mcp-1.6.0-py3-none-any.whl", hash = "sha256:7bd24c6ea042dbec44c754f100984d186620d8b841ec30f1b19eda9b93a634d0"},
+    {file = "mcp-1.6.0.tar.gz", hash = "sha256:d9324876de2c5637369f43161cd71eebfd803df5a95e46225cab8d280e366723"},
+]
+
+[package.dependencies]
+anyio = ">=4.5"
+httpx = ">=0.27"
+httpx-sse = ">=0.4"
+pydantic = ">=2.7.2,<3.0.0"
+pydantic-settings = ">=2.5.2"
+sse-starlette = ">=1.6.1"
+starlette = ">=0.27"
+uvicorn = ">=0.23.1"
+
+[package.extras]
+cli = ["python-dotenv (>=1.0.0)", "typer (>=0.12.4)"]
+rich = ["rich (>=13.9.4)"]
+ws = ["websockets (>=15.0.1)"]
+
 [[package]]
 name = "mdit-py-plugins"
 version = "0.4.2"
@@ -2007,14 +2047,14 @@ pygments = ">2.12.0"
 
 [[package]]
 name = "mkdocs-material"
-version = "9.6.7"
+version = "9.6.10"
 description = "Documentation that simply works"
 optional = false
 python-versions = ">=3.8"
 groups = ["dev"]
 files = [
-    {file = "mkdocs_material-9.6.7-py3-none-any.whl", hash = "sha256:8a159e45e80fcaadd9fbeef62cbf928569b93df954d4dc5ba76d46820caf7b47"},
-    {file = "mkdocs_material-9.6.7.tar.gz", hash = "sha256:3e2c1fceb9410056c2d91f334a00cdea3215c28750e00c691c1e46b2a33309b4"},
+    {file = "mkdocs_material-9.6.10-py3-none-any.whl", hash = "sha256:36168548df4e2ddeb9a334ddae4ab9c388ccfea4dd50ffee657d22b93dcb1c3e"},
+    {file = "mkdocs_material-9.6.10.tar.gz", hash = "sha256:25a453c1f24f34fcf1f53680c03d2c1421b52ce5247f4468153c87a70cd5f1fc"},
 ]
 
 [package.dependencies]
@@ -2435,91 +2475,80 @@ files = [
 
 [[package]]
 name = "orjson"
-version = "3.10.15"
+version = "3.10.16"
 description = "Fast, correct Python JSON library supporting dataclasses, datetimes, and numpy"
 optional = false
-python-versions = ">=3.8"
+python-versions = ">=3.9"
 groups = ["main"]
 files = [
-    {file = "orjson-3.10.15-cp310-cp310-macosx_10_15_x86_64.macosx_11_0_arm64.macosx_10_15_universal2.whl", hash = "sha256:552c883d03ad185f720d0c09583ebde257e41b9521b74ff40e08b7dec4559c04"},
-    {file = "orjson-3.10.15-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:616e3e8d438d02e4854f70bfdc03a6bcdb697358dbaa6bcd19cbe24d24ece1f8"},
-    {file = "orjson-3.10.15-cp310-cp310-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:7c2c79fa308e6edb0ffab0a31fd75a7841bf2a79a20ef08a3c6e3b26814c8ca8"},
-    {file = "orjson-3.10.15-cp310-cp310-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:73cb85490aa6bf98abd20607ab5c8324c0acb48d6da7863a51be48505646c814"},
-    {file = "orjson-3.10.15-cp310-cp310-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:763dadac05e4e9d2bc14938a45a2d0560549561287d41c465d3c58aec818b164"},
-    {file = "orjson-3.10.15-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:a330b9b4734f09a623f74a7490db713695e13b67c959713b78369f26b3dee6bf"},
-    {file = "orjson-3.10.15-cp310-cp310-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:a61a4622b7ff861f019974f73d8165be1bd9a0855e1cad18ee167acacabeb061"},
-    {file = "orjson-3.10.15-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:acd271247691574416b3228db667b84775c497b245fa275c6ab90dc1ffbbd2b3"},
-    {file = "orjson-3.10.15-cp310-cp310-musllinux_1_2_armv7l.whl", hash = "sha256:e4759b109c37f635aa5c5cc93a1b26927bfde24b254bcc0e1149a9fada253d2d"},
-    {file = "orjson-3.10.15-cp310-cp310-musllinux_1_2_i686.whl", hash = "sha256:9e992fd5cfb8b9f00bfad2fd7a05a4299db2bbe92e6440d9dd2fab27655b3182"},
-    {file = "orjson-3.10.15-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:f95fb363d79366af56c3f26b71df40b9a583b07bbaaf5b317407c4d58497852e"},
-    {file = "orjson-3.10.15-cp310-cp310-win32.whl", hash = "sha256:f9875f5fea7492da8ec2444839dcc439b0ef298978f311103d0b7dfd775898ab"},
-    {file = "orjson-3.10.15-cp310-cp310-win_amd64.whl", hash = "sha256:17085a6aa91e1cd70ca8533989a18b5433e15d29c574582f76f821737c8d5806"},
-    {file = "orjson-3.10.15-cp311-cp311-macosx_10_15_x86_64.macosx_11_0_arm64.macosx_10_15_universal2.whl", hash = "sha256:c4cc83960ab79a4031f3119cc4b1a1c627a3dc09df125b27c4201dff2af7eaa6"},
-    {file = "orjson-3.10.15-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:ddbeef2481d895ab8be5185f2432c334d6dec1f5d1933a9c83014d188e102cef"},
-    {file = "orjson-3.10.15-cp311-cp311-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:9e590a0477b23ecd5b0ac865b1b907b01b3c5535f5e8a8f6ab0e503efb896334"},
-    {file = "orjson-3.10.15-cp311-cp311-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:a6be38bd103d2fd9bdfa31c2720b23b5d47c6796bcb1d1b598e3924441b4298d"},
-    {file = "orjson-3.10.15-cp311-cp311-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:ff4f6edb1578960ed628a3b998fa54d78d9bb3e2eb2cfc5c2a09732431c678d0"},
-    {file = "orjson-3.10.15-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:b0482b21d0462eddd67e7fce10b89e0b6ac56570424662b685a0d6fccf581e13"},
-    {file = "orjson-3.10.15-cp311-cp311-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:bb5cc3527036ae3d98b65e37b7986a918955f85332c1ee07f9d3f82f3a6899b5"},
-    {file = "orjson-3.10.15-cp311-cp311-musllinux_1_2_aarch64.whl", hash = "sha256:d569c1c462912acdd119ccbf719cf7102ea2c67dd03b99edcb1a3048651ac96b"},
-    {file = "orjson-3.10.15-cp311-cp311-musllinux_1_2_armv7l.whl", hash = "sha256:1e6d33efab6b71d67f22bf2962895d3dc6f82a6273a965fab762e64fa90dc399"},
-    {file = "orjson-3.10.15-cp311-cp311-musllinux_1_2_i686.whl", hash = "sha256:c33be3795e299f565681d69852ac8c1bc5c84863c0b0030b2b3468843be90388"},
-    {file = "orjson-3.10.15-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:eea80037b9fae5339b214f59308ef0589fc06dc870578b7cce6d71eb2096764c"},
-    {file = "orjson-3.10.15-cp311-cp311-win32.whl", hash = "sha256:d5ac11b659fd798228a7adba3e37c010e0152b78b1982897020a8e019a94882e"},
-    {file = "orjson-3.10.15-cp311-cp311-win_amd64.whl", hash = "sha256:cf45e0214c593660339ef63e875f32ddd5aa3b4adc15e662cdb80dc49e194f8e"},
-    {file = "orjson-3.10.15-cp312-cp312-macosx_10_15_x86_64.macosx_11_0_arm64.macosx_10_15_universal2.whl", hash = "sha256:9d11c0714fc85bfcf36ada1179400862da3288fc785c30e8297844c867d7505a"},
-    {file = "orjson-3.10.15-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:dba5a1e85d554e3897fa9fe6fbcff2ed32d55008973ec9a2b992bd9a65d2352d"},
-    {file = "orjson-3.10.15-cp312-cp312-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:7723ad949a0ea502df656948ddd8b392780a5beaa4c3b5f97e525191b102fff0"},
-    {file = "orjson-3.10.15-cp312-cp312-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:6fd9bc64421e9fe9bd88039e7ce8e58d4fead67ca88e3a4014b143cec7684fd4"},
-    {file = "orjson-3.10.15-cp312-cp312-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:dadba0e7b6594216c214ef7894c4bd5f08d7c0135f4dd0145600be4fbcc16767"},
-    {file = "orjson-3.10.15-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:b48f59114fe318f33bbaee8ebeda696d8ccc94c9e90bc27dbe72153094e26f41"},
-    {file = "orjson-3.10.15-cp312-cp312-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:035fb83585e0f15e076759b6fedaf0abb460d1765b6a36f48018a52858443514"},
-    {file = "orjson-3.10.15-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:d13b7fe322d75bf84464b075eafd8e7dd9eae05649aa2a5354cfa32f43c59f17"},
-    {file = "orjson-3.10.15-cp312-cp312-musllinux_1_2_armv7l.whl", hash = "sha256:7066b74f9f259849629e0d04db6609db4cf5b973248f455ba5d3bd58a4daaa5b"},
-    {file = "orjson-3.10.15-cp312-cp312-musllinux_1_2_i686.whl", hash = "sha256:88dc3f65a026bd3175eb157fea994fca6ac7c4c8579fc5a86fc2114ad05705b7"},
-    {file = "orjson-3.10.15-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:b342567e5465bd99faa559507fe45e33fc76b9fb868a63f1642c6bc0735ad02a"},
-    {file = "orjson-3.10.15-cp312-cp312-win32.whl", hash = "sha256:0a4f27ea5617828e6b58922fdbec67b0aa4bb844e2d363b9244c47fa2180e665"},
-    {file = "orjson-3.10.15-cp312-cp312-win_amd64.whl", hash = "sha256:ef5b87e7aa9545ddadd2309efe6824bd3dd64ac101c15dae0f2f597911d46eaa"},
-    {file = "orjson-3.10.15-cp313-cp313-macosx_10_15_x86_64.macosx_11_0_arm64.macosx_10_15_universal2.whl", hash = "sha256:bae0e6ec2b7ba6895198cd981b7cca95d1487d0147c8ed751e5632ad16f031a6"},
-    {file = "orjson-3.10.15-cp313-cp313-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:f93ce145b2db1252dd86af37d4165b6faa83072b46e3995ecc95d4b2301b725a"},
-    {file = "orjson-3.10.15-cp313-cp313-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:7c203f6f969210128af3acae0ef9ea6aab9782939f45f6fe02d05958fe761ef9"},
-    {file = "orjson-3.10.15-cp313-cp313-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:8918719572d662e18b8af66aef699d8c21072e54b6c82a3f8f6404c1f5ccd5e0"},
-    {file = "orjson-3.10.15-cp313-cp313-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:f71eae9651465dff70aa80db92586ad5b92df46a9373ee55252109bb6b703307"},
-    {file = "orjson-3.10.15-cp313-cp313-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:e117eb299a35f2634e25ed120c37c641398826c2f5a3d3cc39f5993b96171b9e"},
-    {file = "orjson-3.10.15-cp313-cp313-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:13242f12d295e83c2955756a574ddd6741c81e5b99f2bef8ed8d53e47a01e4b7"},
-    {file = "orjson-3.10.15-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:7946922ada8f3e0b7b958cc3eb22cfcf6c0df83d1fe5521b4a100103e3fa84c8"},
-    {file = "orjson-3.10.15-cp313-cp313-musllinux_1_2_armv7l.whl", hash = "sha256:b7155eb1623347f0f22c38c9abdd738b287e39b9982e1da227503387b81b34ca"},
-    {file = "orjson-3.10.15-cp313-cp313-musllinux_1_2_i686.whl", hash = "sha256:208beedfa807c922da4e81061dafa9c8489c6328934ca2a562efa707e049e561"},
-    {file = "orjson-3.10.15-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:eca81f83b1b8c07449e1d6ff7074e82e3fd6777e588f1a6632127f286a968825"},
-    {file = "orjson-3.10.15-cp313-cp313-win32.whl", hash = "sha256:c03cd6eea1bd3b949d0d007c8d57049aa2b39bd49f58b4b2af571a5d3833d890"},
-    {file = "orjson-3.10.15-cp313-cp313-win_amd64.whl", hash = "sha256:fd56a26a04f6ba5fb2045b0acc487a63162a958ed837648c5781e1fe3316cfbf"},
-    {file = "orjson-3.10.15-cp38-cp38-macosx_10_15_x86_64.macosx_11_0_arm64.macosx_10_15_universal2.whl", hash = "sha256:5e8afd6200e12771467a1a44e5ad780614b86abb4b11862ec54861a82d677746"},
-    {file = "orjson-3.10.15-cp38-cp38-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:da9a18c500f19273e9e104cca8c1f0b40a6470bcccfc33afcc088045d0bf5ea6"},
-    {file = "orjson-3.10.15-cp38-cp38-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:bb00b7bfbdf5d34a13180e4805d76b4567025da19a197645ca746fc2fb536586"},
-    {file = "orjson-3.10.15-cp38-cp38-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:33aedc3d903378e257047fee506f11e0833146ca3e57a1a1fb0ddb789876c1e1"},
-    {file = "orjson-3.10.15-cp38-cp38-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:dd0099ae6aed5eb1fc84c9eb72b95505a3df4267e6962eb93cdd5af03be71c98"},
-    {file = "orjson-3.10.15-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:7c864a80a2d467d7786274fce0e4f93ef2a7ca4ff31f7fc5634225aaa4e9e98c"},
-    {file = "orjson-3.10.15-cp38-cp38-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:c25774c9e88a3e0013d7d1a6c8056926b607a61edd423b50eb5c88fd7f2823ae"},
-    {file = "orjson-3.10.15-cp38-cp38-musllinux_1_2_aarch64.whl", hash = "sha256:e78c211d0074e783d824ce7bb85bf459f93a233eb67a5b5003498232ddfb0e8a"},
-    {file = "orjson-3.10.15-cp38-cp38-musllinux_1_2_armv7l.whl", hash = "sha256:43e17289ffdbbac8f39243916c893d2ae41a2ea1a9cbb060a56a4d75286351ae"},
-    {file = "orjson-3.10.15-cp38-cp38-musllinux_1_2_i686.whl", hash = "sha256:781d54657063f361e89714293c095f506c533582ee40a426cb6489c48a637b81"},
-    {file = "orjson-3.10.15-cp38-cp38-musllinux_1_2_x86_64.whl", hash = "sha256:6875210307d36c94873f553786a808af2788e362bd0cf4c8e66d976791e7b528"},
-    {file = "orjson-3.10.15-cp38-cp38-win32.whl", hash = "sha256:305b38b2b8f8083cc3d618927d7f424349afce5975b316d33075ef0f73576b60"},
-    {file = "orjson-3.10.15-cp38-cp38-win_amd64.whl", hash = "sha256:5dd9ef1639878cc3efffed349543cbf9372bdbd79f478615a1c633fe4e4180d1"},
-    {file = "orjson-3.10.15-cp39-cp39-macosx_10_15_x86_64.macosx_11_0_arm64.macosx_10_15_universal2.whl", hash = "sha256:ffe19f3e8d68111e8644d4f4e267a069ca427926855582ff01fc012496d19969"},
-    {file = "orjson-3.10.15-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:d433bf32a363823863a96561a555227c18a522a8217a6f9400f00ddc70139ae2"},
-    {file = "orjson-3.10.15-cp39-cp39-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:da03392674f59a95d03fa5fb9fe3a160b0511ad84b7a3914699ea5a1b3a38da2"},
-    {file = "orjson-3.10.15-cp39-cp39-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:3a63bb41559b05360ded9132032239e47983a39b151af1201f07ec9370715c82"},
-    {file = "orjson-3.10.15-cp39-cp39-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:3766ac4702f8f795ff3fa067968e806b4344af257011858cc3d6d8721588b53f"},
-    {file = "orjson-3.10.15-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:7a1c73dcc8fadbd7c55802d9aa093b36878d34a3b3222c41052ce6b0fc65f8e8"},
-    {file = "orjson-3.10.15-cp39-cp39-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:b299383825eafe642cbab34be762ccff9fd3408d72726a6b2a4506d410a71ab3"},
-    {file = "orjson-3.10.15-cp39-cp39-musllinux_1_2_aarch64.whl", hash = "sha256:abc7abecdbf67a173ef1316036ebbf54ce400ef2300b4e26a7b843bd446c2480"},
-    {file = "orjson-3.10.15-cp39-cp39-musllinux_1_2_armv7l.whl", hash = "sha256:3614ea508d522a621384c1d6639016a5a2e4f027f3e4a1c93a51867615d28829"},
-    {file = "orjson-3.10.15-cp39-cp39-musllinux_1_2_i686.whl", hash = "sha256:295c70f9dc154307777ba30fe29ff15c1bcc9dfc5c48632f37d20a607e9ba85a"},
-    {file = "orjson-3.10.15-cp39-cp39-musllinux_1_2_x86_64.whl", hash = "sha256:63309e3ff924c62404923c80b9e2048c1f74ba4b615e7584584389ada50ed428"},
-    {file = "orjson-3.10.15-cp39-cp39-win32.whl", hash = "sha256:a2f708c62d026fb5340788ba94a55c23df4e1869fec74be455e0b2f5363b8507"},
-    {file = "orjson-3.10.15-cp39-cp39-win_amd64.whl", hash = "sha256:efcf6c735c3d22ef60c4aa27a5238f1a477df85e9b15f2142f9d669beb2d13fd"},
-    {file = "orjson-3.10.15.tar.gz", hash = "sha256:05ca7fe452a2e9d8d9d706a2984c95b9c2ebc5db417ce0b7a49b91d50642a23e"},
+    {file = "orjson-3.10.16-cp310-cp310-macosx_10_15_x86_64.macosx_11_0_arm64.macosx_10_15_universal2.whl", hash = "sha256:4cb473b8e79154fa778fb56d2d73763d977be3dcc140587e07dbc545bbfc38f8"},
+    {file = "orjson-3.10.16-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:622a8e85eeec1948690409a19ca1c7d9fd8ff116f4861d261e6ae2094fe59a00"},
+    {file = "orjson-3.10.16-cp310-cp310-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:c682d852d0ce77613993dc967e90e151899fe2d8e71c20e9be164080f468e370"},
+    {file = "orjson-3.10.16-cp310-cp310-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:8c520ae736acd2e32df193bcff73491e64c936f3e44a2916b548da048a48b46b"},
+    {file = "orjson-3.10.16-cp310-cp310-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:134f87c76bfae00f2094d85cfab261b289b76d78c6da8a7a3b3c09d362fd1e06"},
+    {file = "orjson-3.10.16-cp310-cp310-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:b59afde79563e2cf37cfe62ee3b71c063fd5546c8e662d7fcfc2a3d5031a5c4c"},
+    {file = "orjson-3.10.16-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:113602f8241daaff05d6fad25bd481d54c42d8d72ef4c831bb3ab682a54d9e15"},
+    {file = "orjson-3.10.16-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:4fc0077d101f8fab4031e6554fc17b4c2ad8fdbc56ee64a727f3c95b379e31da"},
+    {file = "orjson-3.10.16-cp310-cp310-musllinux_1_2_armv7l.whl", hash = "sha256:9c6bf6ff180cd69e93f3f50380224218cfab79953a868ea3908430bcfaf9cb5e"},
+    {file = "orjson-3.10.16-cp310-cp310-musllinux_1_2_i686.whl", hash = "sha256:5673eadfa952f95a7cd76418ff189df11b0a9c34b1995dff43a6fdbce5d63bf4"},
+    {file = "orjson-3.10.16-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:5fe638a423d852b0ae1e1a79895851696cb0d9fa0946fdbfd5da5072d9bb9551"},
+    {file = "orjson-3.10.16-cp310-cp310-win32.whl", hash = "sha256:33af58f479b3c6435ab8f8b57999874b4b40c804c7a36b5cc6b54d8f28e1d3dd"},
+    {file = "orjson-3.10.16-cp310-cp310-win_amd64.whl", hash = "sha256:0338356b3f56d71293c583350af26f053017071836b07e064e92819ecf1aa055"},
+    {file = "orjson-3.10.16-cp311-cp311-macosx_10_15_x86_64.macosx_11_0_arm64.macosx_10_15_universal2.whl", hash = "sha256:44fcbe1a1884f8bc9e2e863168b0f84230c3d634afe41c678637d2728ea8e739"},
+    {file = "orjson-3.10.16-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:78177bf0a9d0192e0b34c3d78bcff7fe21d1b5d84aeb5ebdfe0dbe637b885225"},
+    {file = "orjson-3.10.16-cp311-cp311-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:12824073a010a754bb27330cad21d6e9b98374f497f391b8707752b96f72e741"},
+    {file = "orjson-3.10.16-cp311-cp311-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:ddd41007e56284e9867864aa2f29f3136bb1dd19a49ca43c0b4eda22a579cf53"},
+    {file = "orjson-3.10.16-cp311-cp311-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:0877c4d35de639645de83666458ca1f12560d9fa7aa9b25d8bb8f52f61627d14"},
+    {file = "orjson-3.10.16-cp311-cp311-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:9a09a539e9cc3beead3e7107093b4ac176d015bec64f811afb5965fce077a03c"},
+    {file = "orjson-3.10.16-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:31b98bc9b40610fec971d9a4d67bb2ed02eec0a8ae35f8ccd2086320c28526ca"},
+    {file = "orjson-3.10.16-cp311-cp311-musllinux_1_2_aarch64.whl", hash = "sha256:0ce243f5a8739f3a18830bc62dc2e05b69a7545bafd3e3249f86668b2bcd8e50"},
+    {file = "orjson-3.10.16-cp311-cp311-musllinux_1_2_armv7l.whl", hash = "sha256:64792c0025bae049b3074c6abe0cf06f23c8e9f5a445f4bab31dc5ca23dbf9e1"},
+    {file = "orjson-3.10.16-cp311-cp311-musllinux_1_2_i686.whl", hash = "sha256:ea53f7e68eec718b8e17e942f7ca56c6bd43562eb19db3f22d90d75e13f0431d"},
+    {file = "orjson-3.10.16-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:a741ba1a9488c92227711bde8c8c2b63d7d3816883268c808fbeada00400c164"},
+    {file = "orjson-3.10.16-cp311-cp311-win32.whl", hash = "sha256:c7ed2c61bb8226384c3fdf1fb01c51b47b03e3f4536c985078cccc2fd19f1619"},
+    {file = "orjson-3.10.16-cp311-cp311-win_amd64.whl", hash = "sha256:cd67d8b3e0e56222a2e7b7f7da9031e30ecd1fe251c023340b9f12caca85ab60"},
+    {file = "orjson-3.10.16-cp312-cp312-macosx_10_15_x86_64.macosx_11_0_arm64.macosx_10_15_universal2.whl", hash = "sha256:6d3444abbfa71ba21bb042caa4b062535b122248259fdb9deea567969140abca"},
+    {file = "orjson-3.10.16-cp312-cp312-macosx_15_0_arm64.whl", hash = "sha256:30245c08d818fdcaa48b7d5b81499b8cae09acabb216fe61ca619876b128e184"},
+    {file = "orjson-3.10.16-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:a0ba1d0baa71bf7579a4ccdcf503e6f3098ef9542106a0eca82395898c8a500a"},
+    {file = "orjson-3.10.16-cp312-cp312-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:eb0beefa5ef3af8845f3a69ff2a4aa62529b5acec1cfe5f8a6b4141033fd46ef"},
+    {file = "orjson-3.10.16-cp312-cp312-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:6daa0e1c9bf2e030e93c98394de94506f2a4d12e1e9dadd7c53d5e44d0f9628e"},
+    {file = "orjson-3.10.16-cp312-cp312-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:9da9019afb21e02410ef600e56666652b73eb3e4d213a0ec919ff391a7dd52aa"},
+    {file = "orjson-3.10.16-cp312-cp312-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:daeb3a1ee17b69981d3aae30c3b4e786b0f8c9e6c71f2b48f1aef934f63f38f4"},
+    {file = "orjson-3.10.16-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:80fed80eaf0e20a31942ae5d0728849862446512769692474be5e6b73123a23b"},
+    {file = "orjson-3.10.16-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:73390ed838f03764540a7bdc4071fe0123914c2cc02fb6abf35182d5fd1b7a42"},
+    {file = "orjson-3.10.16-cp312-cp312-musllinux_1_2_armv7l.whl", hash = "sha256:a22bba012a0c94ec02a7768953020ab0d3e2b884760f859176343a36c01adf87"},
+    {file = "orjson-3.10.16-cp312-cp312-musllinux_1_2_i686.whl", hash = "sha256:5385bbfdbc90ff5b2635b7e6bebf259652db00a92b5e3c45b616df75b9058e88"},
+    {file = "orjson-3.10.16-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:02c6279016346e774dd92625d46c6c40db687b8a0d685aadb91e26e46cc33e1e"},
+    {file = "orjson-3.10.16-cp312-cp312-win32.whl", hash = "sha256:7ca55097a11426db80f79378e873a8c51f4dde9ffc22de44850f9696b7eb0e8c"},
+    {file = "orjson-3.10.16-cp312-cp312-win_amd64.whl", hash = "sha256:86d127efdd3f9bf5f04809b70faca1e6836556ea3cc46e662b44dab3fe71f3d6"},
+    {file = "orjson-3.10.16-cp313-cp313-macosx_10_15_x86_64.macosx_11_0_arm64.macosx_10_15_universal2.whl", hash = "sha256:148a97f7de811ba14bc6dbc4a433e0341ffd2cc285065199fb5f6a98013744bd"},
+    {file = "orjson-3.10.16-cp313-cp313-macosx_15_0_arm64.whl", hash = "sha256:1d960c1bf0e734ea36d0adc880076de3846aaec45ffad29b78c7f1b7962516b8"},
+    {file = "orjson-3.10.16-cp313-cp313-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:a318cd184d1269f68634464b12871386808dc8b7c27de8565234d25975a7a137"},
+    {file = "orjson-3.10.16-cp313-cp313-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:df23f8df3ef9223d1d6748bea63fca55aae7da30a875700809c500a05975522b"},
+    {file = "orjson-3.10.16-cp313-cp313-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:b94dda8dd6d1378f1037d7f3f6b21db769ef911c4567cbaa962bb6dc5021cf90"},
+    {file = "orjson-3.10.16-cp313-cp313-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:f12970a26666a8775346003fd94347d03ccb98ab8aa063036818381acf5f523e"},
+    {file = "orjson-3.10.16-cp313-cp313-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:15a1431a245d856bd56e4d29ea0023eb4d2c8f71efe914beb3dee8ab3f0cd7fb"},
+    {file = "orjson-3.10.16-cp313-cp313-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:c83655cfc247f399a222567d146524674a7b217af7ef8289c0ff53cfe8db09f0"},
+    {file = "orjson-3.10.16-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:fa59ae64cb6ddde8f09bdbf7baf933c4cd05734ad84dcf4e43b887eb24e37652"},
+    {file = "orjson-3.10.16-cp313-cp313-musllinux_1_2_armv7l.whl", hash = "sha256:ca5426e5aacc2e9507d341bc169d8af9c3cbe88f4cd4c1cf2f87e8564730eb56"},
+    {file = "orjson-3.10.16-cp313-cp313-musllinux_1_2_i686.whl", hash = "sha256:6fd5da4edf98a400946cd3a195680de56f1e7575109b9acb9493331047157430"},
+    {file = "orjson-3.10.16-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:980ecc7a53e567169282a5e0ff078393bac78320d44238da4e246d71a4e0e8f5"},
+    {file = "orjson-3.10.16-cp313-cp313-win32.whl", hash = "sha256:28f79944dd006ac540a6465ebd5f8f45dfdf0948ff998eac7a908275b4c1add6"},
+    {file = "orjson-3.10.16-cp313-cp313-win_amd64.whl", hash = "sha256:fe0a145e96d51971407cb8ba947e63ead2aa915db59d6631a355f5f2150b56b7"},
+    {file = "orjson-3.10.16-cp39-cp39-macosx_10_15_x86_64.macosx_11_0_arm64.macosx_10_15_universal2.whl", hash = "sha256:c35b5c1fb5a5d6d2fea825dec5d3d16bea3c06ac744708a8e1ff41d4ba10cdf1"},
+    {file = "orjson-3.10.16-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:c9aac7ecc86218b4b3048c768f227a9452287001d7548500150bb75ee21bf55d"},
+    {file = "orjson-3.10.16-cp39-cp39-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:6e19f5102fff36f923b6dfdb3236ec710b649da975ed57c29833cb910c5a73ab"},
+    {file = "orjson-3.10.16-cp39-cp39-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:17210490408eb62755a334a6f20ed17c39f27b4f45d89a38cd144cd458eba80b"},
+    {file = "orjson-3.10.16-cp39-cp39-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:fbbe04451db85916e52a9f720bd89bf41f803cf63b038595674691680cbebd1b"},
+    {file = "orjson-3.10.16-cp39-cp39-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:6a966eba501a3a1f309f5a6af32ed9eb8f316fa19d9947bac3e6350dc63a6f0a"},
+    {file = "orjson-3.10.16-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:01e0d22f06c81e6c435723343e1eefc710e0510a35d897856766d475f2a15687"},
+    {file = "orjson-3.10.16-cp39-cp39-musllinux_1_2_aarch64.whl", hash = "sha256:7c1e602d028ee285dbd300fb9820b342b937df64d5a3336e1618b354e95a2569"},
+    {file = "orjson-3.10.16-cp39-cp39-musllinux_1_2_armv7l.whl", hash = "sha256:d230e5020666a6725629df81e210dc11c3eae7d52fe909a7157b3875238484f3"},
+    {file = "orjson-3.10.16-cp39-cp39-musllinux_1_2_i686.whl", hash = "sha256:0f8baac07d4555f57d44746a7d80fbe6b2c4fe2ed68136b4abb51cfec512a5e9"},
+    {file = "orjson-3.10.16-cp39-cp39-musllinux_1_2_x86_64.whl", hash = "sha256:524e48420b90fc66953e91b660b3d05faaf921277d6707e328fde1c218b31250"},
+    {file = "orjson-3.10.16-cp39-cp39-win32.whl", hash = "sha256:a9f614e31423d7292dbca966a53b2d775c64528c7d91424ab2747d8ab8ce5c72"},
+    {file = "orjson-3.10.16-cp39-cp39-win_amd64.whl", hash = "sha256:c338dc2296d1ed0d5c5c27dfb22d00b330555cb706c2e0be1e1c3940a0895905"},
+    {file = "orjson-3.10.16.tar.gz", hash = "sha256:d2aaa5c495e11d17b9b93205f5fa196737ee3202f000aaebf028dc9a73750f10"},
 ]
 
 [[package]]
@@ -2825,14 +2854,14 @@ testing = ["pytest", "pytest-benchmark"]
 
 [[package]]
 name = "pre-commit"
-version = "4.1.0"
+version = "4.2.0"
 description = "A framework for managing and maintaining multi-language pre-commit hooks."
 optional = false
 python-versions = ">=3.9"
 groups = ["dev"]
 files = [
-    {file = "pre_commit-4.1.0-py2.py3-none-any.whl", hash = "sha256:d29e7cb346295bcc1cc75fc3e92e343495e3ea0196c9ec6ba53f49f10ab6ae7b"},
-    {file = "pre_commit-4.1.0.tar.gz", hash = "sha256:ae3f018575a588e30dfddfab9a05448bfbd6b73d78709617b5a2b853549716d4"},
+    {file = "pre_commit-4.2.0-py2.py3-none-any.whl", hash = "sha256:a009ca7205f1eb497d10b845e52c838a98b6cdd2102a6c8e4540e94ee75c58bd"},
+    {file = "pre_commit-4.2.0.tar.gz", hash = "sha256:601283b9757afd87d40c4c4a9b2b5de9637a8ea02eaff7adc2d0fb4e04841146"},
 ]
 
 [package.dependencies]
@@ -3241,6 +3270,27 @@ files = [
 [package.dependencies]
 typing-extensions = ">=4.6.0,<4.7.0 || >4.7.0"
 
+[[package]]
+name = "pydantic-settings"
+version = "2.8.1"
+description = "Settings management using Pydantic"
+optional = false
+python-versions = ">=3.8"
+groups = ["main"]
+files = [
+    {file = "pydantic_settings-2.8.1-py3-none-any.whl", hash = "sha256:81942d5ac3d905f7f3ee1a70df5dfb62d5569c12f51a5a647defc1c3d9ee2e9c"},
+    {file = "pydantic_settings-2.8.1.tar.gz", hash = "sha256:d5c663dfbe9db9d5e1c646b2e161da12f0d734d422ee56f567d0ea2cee4e8585"},
+]
+
+[package.dependencies]
+pydantic = ">=2.7.0"
+python-dotenv = ">=0.21.0"
+
+[package.extras]
+azure-key-vault = ["azure-identity (>=1.16.0)", "azure-keyvault-secrets (>=4.8.0)"]
+toml = ["tomli (>=2.0.1)"]
+yaml = ["pyyaml (>=6.0.1)"]
+
 [[package]]
 name = "pyfiglet"
 version = "1.0.2"
@@ -3415,6 +3465,21 @@ files = [
 [package.dependencies]
 six = ">=1.5"
 
+[[package]]
+name = "python-dotenv"
+version = "1.1.0"
+description = "Read key-value pairs from a .env file and set them as environment variables"
+optional = false
+python-versions = ">=3.9"
+groups = ["main"]
+files = [
+    {file = "python_dotenv-1.1.0-py3-none-any.whl", hash = "sha256:d7c01d9e2293916c18baf562d95698754b0dbbb5e74d457c45d4f6561fb9d55d"},
+    {file = "python_dotenv-1.1.0.tar.gz", hash = "sha256:41f90bc6f5f177fb41f53e87666db362025010eb28f60a01c9143bfa33a2b2d5"},
+]
+
+[package.extras]
+cli = ["click (>=5.0)"]
+
 [[package]]
 name = "python-multipart"
 version = "0.0.20"
@@ -3686,19 +3751,19 @@ rpds-py = ">=0.7.0"
 
 [[package]]
 name = "requests"
-version = "2.32.3"
+version = "2.32.4"
 description = "Python HTTP for Humans."
 optional = false
 python-versions = ">=3.8"
 groups = ["main", "dev"]
 files = [
-    {file = "requests-2.32.3-py3-none-any.whl", hash = "sha256:70761cfe03c773ceb22aa2f671b4757976145175cdfca038c02654d061d6dcc6"},
-    {file = "requests-2.32.3.tar.gz", hash = "sha256:55365417734eb18255590a9ff9eb97e9e1da868d4ccd6402399eaf68af20a760"},
+    {file = "requests-2.32.4-py3-none-any.whl", hash = "sha256:27babd3cda2a6d50b30443204ee89830707d396671944c998b5975b031ac2b2c"},
+    {file = "requests-2.32.4.tar.gz", hash = "sha256:27d0316682c8a29834d3264820024b62a36942083d52caf2f14c0591336d3422"},
 ]
 
 [package.dependencies]
 certifi = ">=2017.4.17"
-charset-normalizer = ">=2,<4"
+charset_normalizer = ">=2,<4"
 idna = ">=2.5,<4"
 urllib3 = ">=1.21.1,<3"
 
@@ -4062,6 +4127,26 @@ files = [
     {file = "soupsieve-2.6.tar.gz", hash = "sha256:e2e68417777af359ec65daac1057404a3c8a5455bb8abc36f1a9866ab1a51abb"},
 ]
 
+[[package]]
+name = "sse-starlette"
+version = "2.1.3"
+description = "SSE plugin for Starlette"
+optional = false
+python-versions = ">=3.8"
+groups = ["main"]
+files = [
+    {file = "sse_starlette-2.1.3-py3-none-any.whl", hash = "sha256:8ec846438b4665b9e8c560fcdea6bc8081a3abf7942faa95e5a744999d219772"},
+    {file = "sse_starlette-2.1.3.tar.gz", hash = "sha256:9cd27eb35319e1414e3d2558ee7414487f9529ce3b3cf9b21434fd110e017169"},
+]
+
+[package.dependencies]
+anyio = "*"
+starlette = "*"
+uvicorn = "*"
+
+[package.extras]
+examples = ["fastapi"]
+
 [[package]]
 name = "stack-data"
 version = "0.6.3"
@@ -4644,4 +4729,4 @@ propcache = ">=0.2.0"
 [metadata]
 lock-version = "2.1"
 python-versions = "^3.11"
-content-hash = "35e03dba41d30cf6129a4a4f3107eca560f779205b21d3ffb2871eeffc5d5a64"
+content-hash = "ff925fc5fa9bb5fdf0f36dc9399f0117e2ee824742373b3f2c4cf45192955853"

pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "agentic_security"
-version = "0.7.0"
+version = "0.7.4"
 description = "Agentic LLM vulnerability scanner"
 authors = ["Alexander Miasoiedov <msoedov@gmail.com>"]
 maintainers = ["Alexander Miasoiedov <msoedov@gmail.com>"]
@@ -52,6 +52,7 @@ sentry_sdk = "^2.22.0"
 orjson = "^3.10"
 pyfiglet = "^1.0.2"
 termcolor = "^2.4.0"
+mcp = "^1.4.1"
 
 # garak = { version = "*", optional = true }
 pytest-xdist = "3.6.1"
@@ -69,7 +70,7 @@ pytest-mock = "^3.14.0"
 black = ">=24.10,<26.0"
 mypy = "^1.12.0"
 pre-commit = "^4.0.1"
-huggingface-hub = ">=0.25.1,<0.29.0"
+huggingface-hub = ">=0.25.1,<0.30.0"
 
 # Docs
 mkdocs = ">=1.4.2"
@@ -91,3 +92,9 @@ addopts = "--durations=5 -m 'not slow' -n 3"
 asyncio_mode = "auto"
 asyncio_default_fixture_loop_scope = "function"
 markers = "slow: marks tests as slow"
+
+[project]
+# MCP requires the following fields to be present in the pyproject.toml file
+name = "agentic_security"
+version = "1.0.0"
+requires-python = ">=3.11"

tests/conftest.py

@@ -1,8 +1,17 @@
 import os
 
 import pytest
+from cache_to_disk import delete_old_disk_caches
+
+from agentic_security.logutils import logger
 
 
 def pytest_runtest_setup(item):
     if "slow" in item.keywords and not os.getenv("RUN_SLOW_TESTS"):
         pytest.skip("Skipping slow test")
+
+
+@pytest.fixture(autouse=True, scope="session")
+def setup_delete_old_disk_caches():
+    logger.info("delete_old_disk_caches")
+    delete_old_disk_caches()

tests/probe_actor/test_fuzzer.py

@@ -7,6 +7,7 @@ import pytest
 
 from agentic_security.primitives import Scan
 from agentic_security.probe_actor.fuzzer import (
+    FuzzerState,
     generate_prompts,
     perform_many_shot_scan,
     perform_single_shot_scan,
@@ -207,9 +208,7 @@ class TestProcessPrompt(unittest.IsolatedAsyncioTestCase):
             prompt="test prompt",
             tokens=0,
             module_name="module_a",
-            refusals=[],
-            errors=[],
-            outputs=[],
+            fuzzer_state=FuzzerState(),
         )
 
         self.assertEqual(tokens, 3)  # Tokens from "Valid response text"
@@ -226,20 +225,17 @@ class TestProcessPrompt(unittest.IsolatedAsyncioTestCase):
             )
         )
 
-        refusals = []
-        outputs = []
+        fuzzer_state = FuzzerState()
         tokens, refusal = await process_prompt(
             request_factory=mock_request_factory,
             prompt="test prompt",
             tokens=0,
             module_name="module_a",
-            refusals=refusals,
-            errors=[],
-            outputs=outputs,
+            fuzzer_state=fuzzer_state,
         )
 
         self.assertEqual(tokens, 3)  # Tokens from "Response indicating refusal"
-        self.assertFalse(refusal)
+        # self.assertFalse(fuzzer_state.refusals)
 
     async def test_http_error_response(self):
         mock_request_factory = Mock()
@@ -252,15 +248,13 @@ class TestProcessPrompt(unittest.IsolatedAsyncioTestCase):
             )
         )
 
-        refusals = []
+        fuzzer_state = FuzzerState()
         await process_prompt(
             request_factory=mock_request_factory,
             prompt="test prompt",
             tokens=0,
             module_name="module_a",
-            refusals=refusals,
-            errors=[],
-            outputs=[],
+            fuzzer_state=fuzzer_state,
         )
 
     async def test_request_error(self):
@@ -269,18 +263,14 @@ class TestProcessPrompt(unittest.IsolatedAsyncioTestCase):
             side_effect=httpx.RequestError("Connection error")
         )
 
-        errors = []
+        fuzzer_state = FuzzerState()
         tokens, refusal = await process_prompt(
             request_factory=mock_request_factory,
             prompt="test prompt",
             tokens=0,
             module_name="module_a",
-            refusals=[],
-            errors=errors,
-            outputs=[],
+            fuzzer_state=fuzzer_state,
         )
 
         self.assertEqual(tokens, 0)
         self.assertTrue(refusal)
-        self.assertEqual(len(errors), 1)
-        self.assertIn("Connection error", errors[0][3])

tests/test_mcp.py

@@ -0,0 +1,12 @@
+import pytest
+
+from agentic_security.mcp.client import run
+
+
+@pytest.mark.asyncio
+async def test_mcp_echo_tool():
+    """Test the echo tool functionality"""
+    prompts, resources, tools = await run()
+    assert prompts
+    assert resources
+    assert tools

ui/package-lock.json

@@ -6891,9 +6891,9 @@
       }
     },
     "node_modules/http-proxy-middleware": {
-      "version": "2.0.7",
-      "resolved": "https://registry.npmjs.org/http-proxy-middleware/-/http-proxy-middleware-2.0.7.tgz",
-      "integrity": "sha512-fgVY8AV7qU7z/MmXJ/rxwbrtQH4jBQ9m7kp3llF0liB7glmFeVZFBepQb32T3y8n8k2+AEYuMPCpinYW+/CuRA==",
+      "version": "2.0.9",
+      "resolved": "https://registry.npmjs.org/http-proxy-middleware/-/http-proxy-middleware-2.0.9.tgz",
+      "integrity": "sha512-c1IyJYLYppU574+YI7R4QyX2ystMtVXZwIdzazUIPIJsHuWNd+mho2j+bKoHftndicGj9yh+xjd+l0yj7VeT1Q==",
       "dev": true,
       "license": "MIT",
       "dependencies": {