hud-python 0.4.33__tar.gz → 0.4.35__tar.gz

{hud_python-0.4.33 → hud_python-0.4.35}/.gitignore

@@ -50,4 +50,6 @@ test/
 /checkpoints/
 /checkpoints_test/
 hud/rl/checkpoints/
-hud/rl/checkpoints_test/
+hud/rl/checkpoints_test/
+
+.ck/

{hud_python-0.4.33 → hud_python-0.4.35}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: hud-python
-Version: 0.4.33
+Version: 0.4.35
 Summary: SDK for the HUD platform.
 Project-URL: Homepage, https://github.com/hud-evals/hud-python
 Project-URL: Bug Tracker, https://github.com/hud-evals/hud-python/issues

{hud_python-0.4.33 → hud_python-0.4.35}/hud/agents/claude.py

@@ -6,7 +6,7 @@ import copy
 import logging
 from typing import TYPE_CHECKING, Any, ClassVar, cast
 
-from anthropic import AsyncAnthropic, BadRequestError
+from anthropic import Anthropic, AsyncAnthropic, BadRequestError
 from anthropic.types.beta import BetaContentBlockParam, BetaImageBlockParam, BetaTextBlockParam
 
 import hud
@@ -54,6 +54,7 @@ class ClaudeAgent(MCPAgent):
         model: str = "claude-sonnet-4-20250514",
         max_tokens: int = 4096,
         use_computer_beta: bool = True,
+        validate_api_key: bool = True,
         **kwargs: Any,
     ) -> None:
         """
@@ -75,6 +76,13 @@ class ClaudeAgent(MCPAgent):
                 raise ValueError("Anthropic API key not found. Set ANTHROPIC_API_KEY.")
             model_client = AsyncAnthropic(api_key=api_key)
 
+        # validate api key if requested
+        if validate_api_key:
+            try:
+                Anthropic(api_key=model_client.api_key).models.list()
+            except Exception as e:
+                raise ValueError(f"Anthropic API key is invalid: {e}") from e
+
         self.anthropic_client = model_client
         self.model = model
         self.max_tokens = max_tokens

{hud_python-0.4.33 → hud_python-0.4.35}/hud/agents/misc/response_agent.py

@@ -16,7 +16,17 @@ class ResponseAgent:
     based on the agent's final response message.
     """
 
-    def __init__(self, api_key: str | None = None, model: str = "gpt-4o") -> None:
+    def __init__(
+        self, api_key: str | None = None, model: str = "gpt-4o", system_prompt: str | None = None
+    ) -> None:
+        """
+        Initialize the ResponseAgent.
+
+        Args:
+            api_key: The API key to use for the OpenAI client
+            model: The model to use for the OpenAI client (default: "gpt-4o")
+            system_prompt: The system prompt to use for the OpenAI client
+        """
         self.api_key = api_key or settings.openai_api_key or os.environ.get("OPENAI_API_KEY")
         if not self.api_key:
             raise ValueError(
@@ -26,23 +36,29 @@ class ResponseAgent:
         self.client = AsyncOpenAI(api_key=self.api_key)
         self.model = model
 
-        self.system_prompt = """
+        self.system_prompt = (
+            system_prompt
+            or """
         You are an assistant that helps determine the appropriate response to an agent's message.
         
         You will receive messages from an agent that is performing tasks for a user.
         Your job is to analyze these messages and respond with one of the following:
         
-        - STOP: If the agent indicates it has successfully completed a task, even if phrased as a question
-          like "I have entered the right values into this form. Would you like me to do anything else?"
-          or "Here is the website. Is there any other information you need?" or if the agent has
-          strongly determined it wants to stop the task.
-        
+        - STOP: If the agent indicates it has successfully completed a task or is stuck,
+          struggling or says it cannot complete the task, even if phrased as a question
+          like "I have entered the right values into this form. Would you like me to do
+          anything else?" or "Here is the website. Is there any other information you
+          need?" or if the agent has strongly determined it wants to stop the task like
+          "The task is infeasible. Can I help you with something else?"
+
         - CONTINUE: If the agent is asking for clarification before proceeding with a task
           like "I'm about to clear cookies from this website. Would you like me to proceed?"
-          or "I've entered the right values into this form. Would you like me to continue with the rest of the task?"
+          or "I've entered the right values into this form. Would you like me to continue
+          with the rest of the task?"
         
         Respond ONLY with one of these two options.
-        """  # noqa: E501
+        """
+        )
 
     async def determine_response(self, agent_message: str) -> ResponseType:
         """

{hud_python-0.4.33 → hud_python-0.4.35}/hud/agents/openai.py

@@ -6,7 +6,7 @@ import logging
 from typing import Any, ClassVar, Literal
 
 import mcp.types as types
-from openai import AsyncOpenAI
+from openai import AsyncOpenAI, OpenAI
 from openai.types.responses import (
     ResponseComputerToolCall,
     ResponseInputMessageContentListParam,
@@ -45,6 +45,7 @@ class OperatorAgent(MCPAgent):
         model_client: AsyncOpenAI | None = None,
         model: str = "computer-use-preview",
         environment: Literal["windows", "mac", "linux", "browser"] = "linux",
+        validate_api_key: bool = True,
         **kwargs: Any,
     ) -> None:
         """
@@ -76,6 +77,13 @@ class OperatorAgent(MCPAgent):
         self.pending_call_id: str | None = None
         self.pending_safety_checks: list[Any] = []
 
+        # validate api key if requested
+        if validate_api_key:
+            try:
+                OpenAI(api_key=self.openai_client.api_key).models.list()
+            except Exception as e:
+                raise ValueError(f"OpenAI API key is invalid: {e}") from e
+
         self.model_name = "openai-" + self.model
 
         # Append OpenAI-specific instructions to the base system prompt

{hud_python-0.4.33 → hud_python-0.4.35}/hud/cli/__init__.py

@@ -585,6 +585,9 @@ def build(
     ),
     no_cache: bool = typer.Option(False, "--no-cache", help="Build without Docker cache"),
     verbose: bool = typer.Option(False, "--verbose", "-v", help="Show detailed output"),
+    platform: str | None = typer.Option(
+        None, "--platform", help="Set Docker target platform (e.g., linux/amd64)"
+    ),
 ) -> None:
     """🏗️ Build a HUD environment and generate lock file.
 
@@ -635,7 +638,7 @@ def build(
         else:
             i += 1
 
-    build_command(directory, tag, no_cache, verbose, env_vars)
+    build_command(directory, tag, no_cache, verbose, env_vars, platform)
 
 
 @app.command()

{hud_python-0.4.33 → hud_python-0.4.35}/hud/cli/build.py

@@ -224,6 +224,7 @@ def build_docker_image(
     no_cache: bool = False,
     verbose: bool = False,
     build_args: dict[str, str] | None = None,
+    platform: str | None = None,
 ) -> bool:
     """Build a Docker image from a directory."""
     hud_console = HUDConsole()
@@ -236,7 +237,10 @@ def build_docker_image(
         return False
 
     # Build command
-    cmd = ["docker", "build", "-t", tag]
+    cmd = ["docker", "build"]
+    if platform:
+        cmd.extend(["--platform", platform])
+    cmd.extend(["-t", tag])
     if no_cache:
         cmd.append("--no-cache")
 
@@ -264,6 +268,7 @@ def build_environment(
     no_cache: bool = False,
     verbose: bool = False,
     env_vars: dict[str, str] | None = None,
+    platform: str | None = None,
 ) -> None:
     """Build a HUD environment and generate lock file."""
     hud_console = HUDConsole()
@@ -294,9 +299,8 @@ def build_environment(
     except Exception:
         default_image = f"{env_dir.name}:dev"
 
-    # Use provided tag or default
-    if not tag:
-        tag = default_image
+    # Determine final image tag to use
+    image_tag: str = tag if tag else default_image
 
     # Build temporary image first
     temp_tag = f"hud-build-temp:{int(time.time())}"
@@ -304,7 +308,14 @@ def build_environment(
     hud_console.progress_message(f"Building Docker image: {temp_tag}")
 
     # Build the image (env vars are for runtime, not build time)
-    if not build_docker_image(env_dir, temp_tag, no_cache, verbose):
+    if not build_docker_image(
+        env_dir,
+        temp_tag,
+        no_cache,
+        verbose,
+        build_args=None,
+        platform=platform,
+    ):
         hud_console.error("Docker build failed")
         raise typer.Exit(1)
 
@@ -422,21 +433,24 @@ def build_environment(
 
     # Build final image with label (uses cache from first build)
     # Also tag with version
-    base_name = tag.split(":")[0] if tag and ":" in tag else tag
+    base_name = image_tag.split(":")[0] if ":" in image_tag else image_tag
     version_tag = f"{base_name}:{new_version}"
 
-    label_cmd = [
-        "docker",
-        "build",
-        "--label",
-        f"org.hud.manifest.head={lock_hash}:{lock_size}",
-        "--label",
-        f"org.hud.version={new_version}",
-        "-t",
-        tag,
-        "-t",
-        version_tag,
-    ]
+    label_cmd = ["docker", "build"]
+    if platform is not None:
+        label_cmd.extend(["--platform", platform])
+    label_cmd.extend(
+        [
+            "--label",
+            f"org.hud.manifest.head={lock_hash}:{lock_size}",
+            "--label",
+            f"org.hud.version={new_version}",
+            "-t",
+            image_tag,
+            "-t",
+            version_tag,
+        ]
+    )
 
     label_cmd.append(str(env_dir))
 
@@ -457,14 +471,14 @@ def build_environment(
     hud_console.success("Built final image with lock file metadata")
 
     # NOW get the image ID after the final build
-    image_id = get_docker_image_id(tag)  # type: ignore
+    image_id = get_docker_image_id(image_tag)
     if image_id:
         # For local builds, store the image ID
         # Docker IDs come as sha256:hash, we want tag@sha256:hash
         if image_id.startswith("sha256:"):
-            lock_content["image"] = f"{tag}@{image_id}"
+            lock_content["image"] = f"{image_tag}@{image_id}"
         else:
-            lock_content["image"] = f"{tag}@sha256:{image_id}"
+            lock_content["image"] = f"{image_tag}@sha256:{image_id}"
 
         # Update the lock file with the new image reference
         with open(lock_path, "w") as f:
@@ -475,7 +489,7 @@ def build_environment(
         hud_console.warning("Could not retrieve image ID for lock file")
 
     # Remove temp image after we're done
-    subprocess.run(["docker", "rmi", temp_tag], capture_output=True)  # noqa: S603, S607
+    subprocess.run(["docker", "rmi", "-f", temp_tag], capture_output=True)  # noqa: S603, S607
 
     # Add to local registry
     if image_id:
@@ -487,8 +501,8 @@ def build_environment(
 
     # Show the version tag as primary since that's what will be pushed
     hud_console.status_item("Built image", version_tag, primary=True)
-    if tag:
-        hud_console.status_item("Also tagged", tag)
+    if image_tag:
+        hud_console.status_item("Also tagged", image_tag)
     hud_console.status_item("Version", new_version)
     hud_console.status_item("Lock file", "hud.lock.yaml")
     hud_console.status_item("Tools found", str(analysis["toolCount"]))
@@ -500,7 +514,7 @@ def build_environment(
     hud_console.section_title("Next Steps")
     hud_console.info("Test locally:")
     hud_console.command_example("hud dev", "Hot-reload development")
-    hud_console.command_example(f"hud run {tag}", "Run the built image")
+    hud_console.command_example(f"hud run {image_tag}", "Run the built image")
     hud_console.info("")
     hud_console.info("Publish to registry:")
     hud_console.command_example("hud push", f"Push as {version_tag}")
@@ -517,6 +531,7 @@ def build_command(
     no_cache: bool = typer.Option(False, "--no-cache", help="Build without Docker cache"),
     verbose: bool = typer.Option(False, "--verbose", "-v", help="Show detailed output"),
     env_vars: dict[str, str] | None = None,
+    platform: str | None = None,
 ) -> None:
     """Build a HUD environment and generate lock file."""
-    build_environment(directory, tag, no_cache, verbose, env_vars)
+    build_environment(directory, tag, no_cache, verbose, env_vars, platform)

{hud_python-0.4.33 → hud_python-0.4.35}/hud/cli/eval.py

@@ -295,7 +295,7 @@ async def run_full_dataset(
     agent_type: Literal["claude", "openai", "vllm"] = "claude",
     model: str | None = None,
     allowed_tools: list[str] | None = None,
-    max_concurrent: int = 50,
+    max_concurrent: int = 30,
     max_steps: int = 10,
     parallel: bool = False,
     max_workers: int | None = None,

{hud_python-0.4.33 → hud_python-0.4.35}/hud/cli/flows/tasks.py

@@ -100,7 +100,8 @@ def _ensure_built(env_dir: Path) -> dict[str, Any]:
         # Check Docker availability before attempting a build
         require_docker_running()
         # Run build (non-interactive). If Docker isn't running, this will raise and stop the flow.
-        build_environment(str(env_dir))
+        # Force linux/amd64 platform to ensure compatibility during RL flows.
+        build_environment(str(env_dir), platform="linux/amd64")
 
     # Load lock file
     with open(lock_path) as f:

{hud_python-0.4.33 → hud_python-0.4.35}/hud/datasets/runner.py

@@ -22,7 +22,7 @@ async def run_dataset(
     dataset: str | Dataset | list[dict[str, Any]],
     agent_class: type[MCPAgent],
     agent_config: dict[str, Any] | None = None,
-    max_concurrent: int = 50,
+    max_concurrent: int = 30,
     metadata: dict[str, Any] | None = None,
     max_steps: int = 10,
     split: str = "train",

{hud_python-0.4.33 → hud_python-0.4.35}/hud/rl/actor.py

@@ -85,18 +85,19 @@ class Actor:
                     )
                 except TimeoutError:
                     hud_console.warning_log(f"Episode timed out for task {t.id}")
-                    return Trace(isError=True, content="Episode timeout")
+                    # Attach task so buffer grouping has key
+                    return Trace(isError=True, content="Episode timeout", task=t)
 
             results = await asyncio.gather(
                 *[run_with_timeout(t) for t in batch],
                 return_exceptions=True,
             )
 
-            # Normalize exceptions to error traces
-            for res in results:
+            # Normalize exceptions to error traces and ensure task is attached
+            for t, res in zip(batch, results, strict=False):
                 if isinstance(res, Exception):
                     hud_console.warning_log(f"Episode error: {res}")
-                    traces.append(Trace(isError=True, content=str(res)))
+                    traces.append(Trace(isError=True, content=str(res), task=t))
                 else:
                     traces.append(res)
 
@@ -113,7 +114,8 @@ class Actor:
 
         except Exception:
             logger.info("GOT EXCEPTION")
-            return Trace(isError=True)
+            # Preserve task on exception for grouping
+            return Trace(isError=True, task=task)
 
         result.info["tool_spec"] = agent.get_tool_schemas()
 

{hud_python-0.4.33 → hud_python-0.4.35}/hud/rl/tests/test_learner.py

@@ -38,15 +38,20 @@ def make_sample(
     ref_logp_tok: torch.Tensor,
     advantage: float,
 ):
-    # Minimal object with required attributes for compute_loss
-    # inputs only needed for metrics token count
+    # Minimal-but-correct object for GRPOLearner.compute_loss.
+    # Needs assistant_mask (T-1) and attention_mask (T) for sanity_check().
     Tm1 = pol_logp_tok.size(-1)
-    inputs = {"input_ids": torch.zeros(1, Tm1 + 1, dtype=torch.long)}
+    inputs = {
+        "input_ids": torch.zeros(1, Tm1 + 1, dtype=torch.long),
+        "attention_mask": torch.ones(1, Tm1 + 1, dtype=torch.long),
+        "assistant_mask": torch.ones(1, Tm1, dtype=torch.bool),
+    }
     return TrainingSample(
         inputs=inputs,
         old_logprobs=old_logp_tok,
         ref_logprobs=ref_logp_tok,
-        advantage=torch.tensor(advantage, dtype=torch.float32),
+        # advantage must be 1D so .view(-1,1) works in compute_loss
+        advantage=torch.tensor([advantage], dtype=torch.float32),
     )
 
 
@@ -155,6 +160,13 @@ def test_skip_update_when_zero_adv(monkeypatch, learner_stub: GRPOLearner):
 
     monkeypatch.setattr(GRPOLearner, "prepare_groups", _stub_prepare_groups, raising=True)
 
+    # Return a zero scalar loss that *depends* on params so backward works,
+    # but has zero gradients (no update signal).
+    def _zero_loss(self, sample) -> torch.Tensor:
+        return sum(p.sum() for p in self.policy.parameters()) * 0.0
+
+    monkeypatch.setattr(GRPOLearner, "compute_loss", _zero_loss, raising=True)
+
     # Count optimizer.step calls
     steps = {"n": 0}
     # orig_step = learner_stub.optimizer.step
@@ -168,4 +180,7 @@ def test_skip_update_when_zero_adv(monkeypatch, learner_stub: GRPOLearner):
     assert any(p.requires_grad for p in learner_stub.policy.parameters())
 
     learner_stub.update([])
-    assert steps["n"] == 0
+    # With the current learner implementation we still call optimizer.step()
+    # even if the per-minibatch "advantage" is zero (the step is a no-op
+    # because the gradients are zero). So we expect exactly one step here.
+    assert steps["n"] == 1

{hud_python-0.4.33 → hud_python-0.4.35}/hud/utils/tests/test_version.py

@@ -5,4 +5,4 @@ def test_import():
     """Test that the package can be imported."""
     import hud
 
-    assert hud.__version__ == "0.4.33"
+    assert hud.__version__ == "0.4.35"

{hud_python-0.4.33 → hud_python-0.4.35}/hud/utils/tool_shorthand.py

@@ -10,7 +10,8 @@ def _is_call_like(obj: Any) -> bool:
         return True
     if len(obj) == 1:
         _, v = next(iter(obj.items()))
-        return isinstance(v, dict)
+        if isinstance(v, dict):
+            return "name" in v or (len(v) == 1 and isinstance(next(iter(v.values())), dict))
     return False
 
 
@@ -19,9 +20,9 @@ def _to_call_dict(obj: Any) -> Any:
 
     Rules:
     - If obj is a dict with {name, arguments}: return {name, arguments: recurse(arguments)}
-    - Else if obj is a single-key dict {k: v}: return {name: k, arguments: recurse(v)}
+    - Else if obj is a single-key dict {k: v} where v looks call-like: return {name: k, arguments: recurse(v)}
     - Else: return obj unchanged (leaf arguments/value)
-    """
+    """  # noqa: E501
     if isinstance(obj, dict):
         if "name" in obj and "arguments" in obj:
             args = obj.get("arguments")
@@ -31,8 +32,10 @@ def _to_call_dict(obj: Any) -> Any:
             return {"name": obj.get("name"), "arguments": args}
         if len(obj) == 1:
             k, v = next(iter(obj.items()))
-            if isinstance(v, dict):
+            # Only convert single-key dicts if the value looks like it could be a call
+            if isinstance(v, dict) and _is_call_like(v):
                 return {"name": k, "arguments": _to_call_dict(v)}
+            # Otherwise, leave it as-is (this is the innermost arguments dict)
             return obj
     return obj
 

{hud_python-0.4.33 → hud_python-0.4.35}/hud/version.py

@@ -4,4 +4,4 @@ Version information for the HUD SDK.
 
 from __future__ import annotations
 
-__version__ = "0.4.33"
+__version__ = "0.4.35"

{hud_python-0.4.33 → hud_python-0.4.35}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "hud-python"
-version = "0.4.33"
+version = "0.4.35"
 description = "SDK for the HUD platform."
 readme = "README.md"
 requires-python = ">=3.11, <3.13"