dreadnode 2.0.31__tar.gz → 2.0.32__tar.gz

{dreadnode-2.0.31 → dreadnode-2.0.32}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: dreadnode
-Version: 2.0.31
+Version: 2.0.32
 Summary: Dreadnode SDK
 Project-URL: Homepage, https://dreadnode.io
 Project-URL: Documentation, https://docs.dreadnode.io
@@ -27,14 +27,13 @@ Requires-Dist: loguru>=0.7.3
 Requires-Dist: markdown-it-py[linkify]>=2.1.0
 Requires-Dist: markdownify<2.0.0,>=1.1.0
 Requires-Dist: mcp<2.0.0,>=1.25.0
-Requires-Dist: moviepy<3.0.0,>=2.1.2
 Requires-Dist: networkx>=3.6.1
 Requires-Dist: numpy<=2.3.5,>=2.3.0
 Requires-Dist: optuna<5.0.0,>=4.5.0
 Requires-Dist: orjson>=3.11.5
 Requires-Dist: packaging>=24.0
 Requires-Dist: pandas<3.0.0,>=2.2.3
-Requires-Dist: pillow<12.3.0,>=12.2.0
+Requires-Dist: pillow<13.0.0,>=11.0.0
 Requires-Dist: prompt-toolkit<4.0.0,>=3.0.52
 Requires-Dist: pyarrow>=4.0.0
 Requires-Dist: pydantic-xml>=2.14.0
@@ -50,6 +49,7 @@ Requires-Dist: websockets>=14.0
 Requires-Dist: xmltodict>=1.0.2
 Provides-Extra: all
 Requires-Dist: confusables<2.0.0,>=1.2.0; extra == 'all'
+Requires-Dist: moviepy<3.0.0,>=2.1.2; extra == 'all'
 Requires-Dist: nltk<4.0.0,>=3.9.1; extra == 'all'
 Requires-Dist: presidio-analyzer<3.0.0,>=2.2.359; extra == 'all'
 Requires-Dist: python-jsonpath>=2.0.1; extra == 'all'
@@ -75,6 +75,8 @@ Requires-Dist: textstat<1.0.0,>=0.7.10; extra == 'scoring'
 Provides-Extra: training
 Requires-Dist: ray[default,serve]<3.0.0,>=2.10.0; extra == 'training'
 Requires-Dist: torch>=2.9.0; extra == 'training'
+Provides-Extra: video
+Requires-Dist: moviepy<3.0.0,>=2.1.2; extra == 'video'
 Description-Content-Type: text/markdown
 
 <p align="center">

{dreadnode-2.0.31 → dreadnode-2.0.32}/dreadnode/agents/agent.py

@@ -194,6 +194,11 @@ class Agent(Executor[AgentEvent, Trajectory]):
     # Private state
     _generator: Generator | None = PrivateAttr(None, init=False)
     _current_input: str = PrivateAttr("", init=False)
+    _capability: tuple[str, str] | None = PrivateAttr(None, init=False)
+    """The (name, version) capability this agent runs under, if any. Bound to
+    ``current_capability`` for the duration of a run (and reset afterward) so
+    runtime tools attribute emitted items to the right producer without leaking
+    attribution into later agents that share the async context."""
     _permission_bridge: PermissionBridge | None = PrivateAttr(None, init=False)
     """Tool-approval bridge injected by the runtime so foreign engines wire their
     permission callback into the native HITL path. ``None`` for bare-SDK use."""
@@ -1204,6 +1209,13 @@ class Agent(Executor[AgentEvent, Trajectory]):
             active_trajectory = self.trajectory
 
         async with AsyncExitStack() as stack:
+            # Scope capability attribution to THIS run (reset on exit) so it never
+            # leaks into a later agent sharing the async context.
+            if self._capability is not None:
+                from dreadnode.tracing.span import bind_capability
+
+                stack.enter_context(bind_capability(self._capability))
+
             # Enter tool contexts
             for tool_container in self.tools:
                 if hasattr(tool_container, "__aenter__") and hasattr(tool_container, "__aexit__"):

{dreadnode-2.0.31 → dreadnode-2.0.32}/dreadnode/agents/subagent.py

@@ -5,6 +5,7 @@ Similar to Claude Code's Task tool, this allows spawning specialized agents
 to handle specific subtasks autonomously.
 """
 
+import asyncio
 import typing as t
 from textwrap import dedent
 
@@ -93,8 +94,11 @@ class SubAgentToolset(Toolset):
     parent_agent: t.Any = Config(default=None)
     """The parent agent to clone sub-agents from."""
 
-    run_in_background: bool = Config(default=False)
-    """Whether to run sub-agents in background (not yet implemented)."""
+    timeout: float | None = Config(default=3600.0)
+    """Maximum seconds a spawned sub-agent may run before it is cancelled and
+    reported as a timeout. ``None`` disables the ceiling (unbounded). A wedged
+    tool inside a sub-agent (hung MCP transport, stuck LLM stream) would
+    otherwise hang the parent session indefinitely."""
 
     @tool_method
     async def spawn_agent(
@@ -126,7 +130,6 @@ class SubAgentToolset(Toolset):
         - Complex tasks requiring focused work
         - Exploration that might take many steps
         - Tasks where you want isolated context
-        - Parallel work (with run_in_background)
 
         ## Examples
 
@@ -173,7 +176,37 @@ class SubAgentToolset(Toolset):
         sub_agent.reset()
 
         try:
-            trajectory = await sub_agent.run(task)
+            # Bound the sub-agent run so a wedged tool inside it can't hang the
+            # parent session forever. ``asyncio.timeout(None)`` is a no-op, so a
+            # single path covers both the bounded and unbounded cases. On expiry
+            # it cancels the inner run and raises ``TimeoutError`` here; the
+            # ``CancelledError`` it uses internally never escapes. We check
+            # ``cm.expired()`` to tell our ceiling breach apart from a builtin
+            # ``TimeoutError`` raised *inside* the sub-agent, and deliberately do
+            # NOT catch ``CancelledError`` (a genuine parent-step cancellation
+            # must propagate, not be swallowed into a tool result).
+            async with asyncio.timeout(self.timeout) as cm:
+                trajectory = await sub_agent.run(task)
+        except TimeoutError as e:
+            if not cm.expired():
+                # A builtin ``TimeoutError`` from within the sub-agent (socket,
+                # inner ``async_timeout``, …) — not our ceiling. Surface it as an
+                # ordinary failure rather than a false "cancelled at ceiling".
+                logger.error(f"Sub-agent failed: {e}")
+                return f"Sub-agent failed: {e}"
+            logger.warning("Sub-agent '{}' timed out after {}s", config["name"], self.timeout)
+            # The cancelled run still burned tokens; carry its partial cost so
+            # the parent session's accounting isn't understated.
+            return self._tool_result(
+                f"Sub-agent '{config['name']}' timed out after {self.timeout} seconds "
+                "and was cancelled. Consider narrowing the task or raising the "
+                "SubAgentToolset timeout.",
+                sub_agent.trajectory,
+            )
+        except Exception as e:
+            logger.error(f"Sub-agent failed: {e}")
+            return f"Sub-agent failed: {e}"
+        else:
             # Get the last assistant message content as the response
             last_message = trajectory.messages[-1] if trajectory.messages else None
             response = str(last_message.content) if last_message else ""
@@ -185,17 +218,17 @@ class SubAgentToolset(Toolset):
             result += f"**Result:**\n{response}"
 
             logger.info(f"Sub-agent completed in {len(trajectory.steps)} steps")
-        except Exception as e:
-            logger.error(f"Sub-agent failed: {e}")
-            return f"Sub-agent failed: {e}"
-        else:
-            # Stash the sub-agent's LLM cost on the tool result metadata so
-            # the agent framework can lift it onto the ``ToolEnd`` event and
-            # the TUI can display it in the parent session's footer.
-            msg = Message(role="tool", content=result)
-            if trajectory.usage.cost_usd is not None:
-                msg.metadata["subagent_cost_usd"] = trajectory.usage.cost_usd
-            return msg
+            return self._tool_result(result, trajectory)
+
+    @staticmethod
+    def _tool_result(content: str, trajectory: t.Any) -> Message:
+        """Build the tool-result message, stashing the sub-agent's LLM cost on
+        its metadata so the agent framework can lift it onto the ``ToolEnd``
+        event and the TUI can display it in the parent session's footer."""
+        msg = Message(role="tool", content=content)
+        if trajectory.usage.cost_usd is not None:
+            msg.metadata["subagent_cost_usd"] = trajectory.usage.cost_usd
+        return msg
 
 
 def create_subagent_tool(parent_agent: "Agent") -> SubAgentToolset:

dreadnode-2.0.32/dreadnode/airt/message_parts.py

@@ -0,0 +1,263 @@
+"""
+Canonical multimodal message parts for AI Red Teaming.
+
+A trial's input (candidate) and output (response) are represented as an ordered list of
+``MessagePart`` dicts. Text parts carry their text inline; media parts (image/audio/video)
+carry a *content-addressed reference* to bytes stored once on the trial span
+(``span.log_artifact`` → ``sha256:`` oid in ``dreadnode.artifacts``), never inline base64.
+
+The shape is a lossless superset of the OpenAI / Anthropic / Gemini content-part models
+(discriminated ``source`` union: artifact | url | file_id) so it round-trips faithfully and
+the platform can resolve refs to presigned inline URLs for display.
+"""
+
+import io
+import tempfile
+import typing as t
+from pathlib import Path
+
+from loguru import logger
+
+if t.TYPE_CHECKING:
+    from dreadnode.core.types import Audio, Image, Video
+    from dreadnode.tracing.span import TaskSpan
+
+_MIME_BY_EXT = {
+    "png": "image/png",
+    "jpg": "image/jpeg",
+    "jpeg": "image/jpeg",
+    "gif": "image/gif",
+    "webp": "image/webp",
+    "bmp": "image/bmp",
+    "mp3": "audio/mpeg",
+    "wav": "audio/wav",
+    "ogg": "audio/ogg",
+    "flac": "audio/flac",
+    "m4a": "audio/mp4",
+    "mp4": "video/mp4",
+    "webm": "video/webm",
+    "mov": "video/quicktime",
+}
+
+
+def _mime_for_extension(ext: str | None, default: str) -> str:
+    if not ext:
+        return default
+    return _MIME_BY_EXT.get(ext.lower().lstrip("."), default)
+
+
+def text_part(
+    text: str,
+    *,
+    transform: str | None = None,
+    variant: str | None = None,
+) -> dict[str, t.Any]:
+    """Build a canonical text message part.
+
+    ``transform``/``variant`` record provenance when a *text* transform was
+    applied, so the UI can attribute the transform to the text part (not only
+    to image/audio/video parts).
+    """
+    part: dict[str, t.Any] = {"kind": "text", "text": text}
+    if transform:
+        part["transform"] = transform
+    if variant:
+        part["variant"] = variant
+    return part
+
+
+def _store_bytes_as_artifact(
+    span: "TaskSpan",
+    data: bytes,
+    *,
+    filename: str,
+) -> dict[str, t.Any] | None:
+    """Persist bytes as a content-addressed span artifact, return the artifact metadata.
+
+    The artifact's MIME type is inferred from ``filename`` by ``log_artifact``; the
+    caller supplies the canonical ``media_type`` on the resulting part separately.
+    """
+    suffix = Path(filename).suffix or ""
+    try:
+        with tempfile.NamedTemporaryFile(suffix=suffix, delete=False) as tmp:
+            tmp.write(data)
+            tmp_path = Path(tmp.name)
+        try:
+            return span.log_artifact(tmp_path, name=filename)
+        finally:
+            tmp_path.unlink(missing_ok=True)
+    except Exception:
+        # Storage may be unconfigured (e.g. local unit tests) — degrade to no media ref
+        # rather than failing the trial. The text parts still carry the attack content.
+        logger.debug("Failed to store media artifact for message part", exc_info=True)
+        return None
+
+
+def _media_part(
+    span: "TaskSpan",
+    data: bytes,
+    *,
+    kind: str,
+    media_type: str,
+    filename: str,
+    transform: str | None,
+    variant: str | None,
+    caption: str | None,
+) -> dict[str, t.Any]:
+    """Build a media message part, storing bytes as a content-addressed artifact."""
+    meta = _store_bytes_as_artifact(span, data, filename=filename)
+    source: dict[str, t.Any] = {
+        "ref": "artifact",
+        "media_type": (meta.get("mime_type") if meta else None) or media_type,
+        "byte_size": (meta.get("size") if meta else None) or len(data),
+        "filename": (meta.get("name") if meta else None) or filename,
+    }
+    if meta and meta.get("oid"):
+        source["artifact_oid"] = meta["oid"]
+    part: dict[str, t.Any] = {"kind": kind, "source": source}
+    if transform:
+        part["transform"] = transform
+    if variant:
+        part["variant"] = variant
+    if caption:
+        part["caption"] = caption
+    return part
+
+
+def image_part(
+    span: "TaskSpan",
+    image: "Image",
+    *,
+    transform: str | None = None,
+    variant: str | None = None,
+    filename: str = "image.png",
+) -> dict[str, t.Any]:
+    """Build an image message part from an ``Image``, storing its bytes as an artifact."""
+    try:
+        data, metadata = image.to_serializable()
+        media_type = _mime_for_extension(metadata.get("extension"), "image/png")
+        caption = metadata.get("caption")
+    except Exception:
+        buffer = io.BytesIO()
+        image.to_pil().save(buffer, format="PNG")
+        data, media_type, caption = buffer.getvalue(), "image/png", None
+    return _media_part(
+        span,
+        data,
+        kind="image",
+        media_type=media_type,
+        filename=filename,
+        transform=transform,
+        variant=variant,
+        caption=caption,
+    )
+
+
+def audio_part(
+    span: "TaskSpan",
+    audio: "Audio",
+    *,
+    transform: str | None = None,
+    variant: str | None = None,
+    filename: str = "audio",
+) -> dict[str, t.Any]:
+    """Build an audio message part from an ``Audio``, storing its bytes as an artifact."""
+    data, metadata = audio.to_serializable()
+    ext = metadata.get("extension", "mp3")
+    media_type = _mime_for_extension(ext, "audio/mpeg")
+    name = filename if Path(filename).suffix else f"{filename}.{ext}"
+    return _media_part(
+        span,
+        data,
+        kind="audio",
+        media_type=media_type,
+        filename=name,
+        transform=transform,
+        variant=variant,
+        caption=metadata.get("transcript"),
+    )
+
+
+def video_part(
+    span: "TaskSpan",
+    video: "Video",
+    *,
+    transform: str | None = None,
+    variant: str | None = None,
+    filename: str = "video",
+) -> dict[str, t.Any]:
+    """Build a video message part from a ``Video``, storing its bytes as an artifact."""
+    data, metadata = video.to_serializable()
+    ext = metadata.get("extension", "mp4")
+    media_type = _mime_for_extension(ext, "video/mp4")
+    name = filename if Path(filename).suffix else f"{filename}.{ext}"
+    return _media_part(
+        span,
+        data,
+        kind="video",
+        media_type=media_type,
+        filename=name,
+        transform=transform,
+        variant=variant,
+        caption=None,
+    )
+
+
+def modality_from_parts(parts: list[dict[str, t.Any]]) -> str:
+    """Derive input_modality from the kinds present in a part list."""
+    kinds = {str(p.get("kind")) for p in parts if p.get("kind")}
+    media = kinds - {"text"}
+    if not media:
+        return "text"
+    if len(media) == 1:
+        return next(iter(media))  # "image" | "audio" | "video"
+    return "multimodal"
+
+
+def response_to_parts(
+    span: "TaskSpan",
+    response: t.Any,
+) -> list[dict[str, t.Any]]:
+    """
+    Build response message parts from a target's output.
+
+    Handles text-out targets (str), media-out targets (Image/Audio/Video), and mixed
+    outputs (a Message, or a list/dict containing media). Media bytes are stored as
+    content-addressed artifacts; text is inlined.
+    """
+    from dreadnode.core.types import Audio, Image, Video
+
+    parts: list[dict[str, t.Any]] = []
+
+    def _walk(value: t.Any) -> None:
+        if value is None:
+            return
+        if isinstance(value, str):
+            if value:
+                parts.append(text_part(value))
+        elif isinstance(value, Image):
+            parts.append(image_part(span, value, variant="output", filename="response.png"))
+        elif isinstance(value, Audio):
+            parts.append(audio_part(span, value, variant="output", filename="response"))
+        elif isinstance(value, Video):
+            parts.append(video_part(span, value, variant="output", filename="response"))
+        elif isinstance(value, dict):
+            # A Message-like dict or provider payload — walk content/parts if present.
+            for key in ("content", "content_parts", "parts", "text"):
+                if key in value:
+                    _walk(value[key])
+                    return
+            parts.append(text_part(str(value)))
+        elif isinstance(value, (list, tuple)):
+            for item in value:
+                _walk(item)
+        else:
+            # Message object or unknown — try content_parts, else stringify.
+            content = getattr(value, "content_parts", None) or getattr(value, "content", None)
+            if content is not None and not isinstance(content, str):
+                _walk(content)
+            else:
+                parts.append(text_part(str(value)))
+
+    _walk(response)
+    return parts or [text_part("")]