hud-python 0.4.45__py3-none-any.whl → 0.5.13__py3-none-any.whl
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- hud/__init__.py +27 -7
- hud/agents/__init__.py +70 -5
- hud/agents/base.py +238 -500
- hud/agents/claude.py +236 -247
- hud/agents/gateway.py +42 -0
- hud/agents/gemini.py +264 -0
- hud/agents/gemini_cua.py +324 -0
- hud/agents/grounded_openai.py +98 -100
- hud/agents/misc/integration_test_agent.py +51 -20
- hud/agents/misc/response_agent.py +48 -36
- hud/agents/openai.py +282 -296
- hud/agents/{openai_chat_generic.py → openai_chat.py} +63 -33
- hud/agents/operator.py +199 -0
- hud/agents/resolver.py +70 -0
- hud/agents/tests/conftest.py +133 -0
- hud/agents/tests/test_base.py +300 -622
- hud/agents/tests/test_base_runtime.py +233 -0
- hud/agents/tests/test_claude.py +381 -214
- hud/agents/tests/test_client.py +9 -10
- hud/agents/tests/test_gemini.py +369 -0
- hud/agents/tests/test_grounded_openai_agent.py +65 -50
- hud/agents/tests/test_openai.py +377 -140
- hud/agents/tests/test_operator.py +362 -0
- hud/agents/tests/test_resolver.py +192 -0
- hud/agents/tests/test_run_eval.py +179 -0
- hud/agents/types.py +148 -0
- hud/cli/__init__.py +493 -546
- hud/cli/analyze.py +43 -5
- hud/cli/build.py +699 -113
- hud/cli/debug.py +8 -5
- hud/cli/dev.py +889 -732
- hud/cli/eval.py +793 -667
- hud/cli/flows/dev.py +167 -0
- hud/cli/flows/init.py +191 -0
- hud/cli/flows/tasks.py +153 -56
- hud/cli/flows/templates.py +151 -0
- hud/cli/flows/tests/__init__.py +1 -0
- hud/cli/flows/tests/test_dev.py +126 -0
- hud/cli/init.py +60 -58
- hud/cli/pull.py +1 -1
- hud/cli/push.py +38 -13
- hud/cli/rft.py +311 -0
- hud/cli/rft_status.py +145 -0
- hud/cli/tests/test_analyze.py +5 -5
- hud/cli/tests/test_analyze_metadata.py +3 -2
- hud/cli/tests/test_analyze_module.py +120 -0
- hud/cli/tests/test_build.py +110 -8
- hud/cli/tests/test_build_failure.py +41 -0
- hud/cli/tests/test_build_module.py +50 -0
- hud/cli/tests/test_cli_init.py +6 -1
- hud/cli/tests/test_cli_more_wrappers.py +30 -0
- hud/cli/tests/test_cli_root.py +140 -0
- hud/cli/tests/test_convert.py +361 -0
- hud/cli/tests/test_debug.py +12 -10
- hud/cli/tests/test_dev.py +197 -0
- hud/cli/tests/test_eval.py +251 -0
- hud/cli/tests/test_eval_bedrock.py +51 -0
- hud/cli/tests/test_init.py +124 -0
- hud/cli/tests/test_main_module.py +11 -5
- hud/cli/tests/test_mcp_server.py +12 -100
- hud/cli/tests/test_push.py +1 -1
- hud/cli/tests/test_push_happy.py +74 -0
- hud/cli/tests/test_push_wrapper.py +23 -0
- hud/cli/tests/test_registry.py +1 -1
- hud/cli/tests/test_utils.py +1 -1
- hud/cli/{rl → utils}/celebrate.py +14 -12
- hud/cli/utils/config.py +18 -1
- hud/cli/utils/docker.py +130 -4
- hud/cli/utils/env_check.py +9 -9
- hud/cli/utils/git.py +136 -0
- hud/cli/utils/interactive.py +39 -5
- hud/cli/utils/metadata.py +70 -1
- hud/cli/utils/runner.py +1 -1
- hud/cli/utils/server.py +2 -2
- hud/cli/utils/source_hash.py +3 -3
- hud/cli/utils/tasks.py +4 -1
- hud/cli/utils/tests/__init__.py +0 -0
- hud/cli/utils/tests/test_config.py +58 -0
- hud/cli/utils/tests/test_docker.py +93 -0
- hud/cli/utils/tests/test_docker_hints.py +71 -0
- hud/cli/utils/tests/test_env_check.py +74 -0
- hud/cli/utils/tests/test_environment.py +42 -0
- hud/cli/utils/tests/test_git.py +142 -0
- hud/cli/utils/tests/test_interactive_module.py +60 -0
- hud/cli/utils/tests/test_local_runner.py +50 -0
- hud/cli/utils/tests/test_logging_utils.py +23 -0
- hud/cli/utils/tests/test_metadata.py +49 -0
- hud/cli/utils/tests/test_package_runner.py +35 -0
- hud/cli/utils/tests/test_registry_utils.py +49 -0
- hud/cli/utils/tests/test_remote_runner.py +25 -0
- hud/cli/utils/tests/test_runner_modules.py +52 -0
- hud/cli/utils/tests/test_source_hash.py +36 -0
- hud/cli/utils/tests/test_tasks.py +80 -0
- hud/cli/utils/version_check.py +258 -0
- hud/cli/{rl → utils}/viewer.py +2 -2
- hud/clients/README.md +12 -11
- hud/clients/__init__.py +4 -3
- hud/clients/base.py +166 -26
- hud/clients/environment.py +51 -0
- hud/clients/fastmcp.py +13 -6
- hud/clients/mcp_use.py +45 -15
- hud/clients/tests/test_analyze_scenarios.py +206 -0
- hud/clients/tests/test_protocol.py +9 -3
- hud/datasets/__init__.py +23 -20
- hud/datasets/loader.py +326 -0
- hud/datasets/runner.py +198 -105
- hud/datasets/tests/__init__.py +0 -0
- hud/datasets/tests/test_loader.py +221 -0
- hud/datasets/tests/test_utils.py +315 -0
- hud/datasets/utils.py +270 -90
- hud/environment/__init__.py +52 -0
- hud/environment/connection.py +258 -0
- hud/environment/connectors/__init__.py +33 -0
- hud/environment/connectors/base.py +68 -0
- hud/environment/connectors/local.py +177 -0
- hud/environment/connectors/mcp_config.py +137 -0
- hud/environment/connectors/openai.py +101 -0
- hud/environment/connectors/remote.py +172 -0
- hud/environment/environment.py +835 -0
- hud/environment/integrations/__init__.py +45 -0
- hud/environment/integrations/adk.py +67 -0
- hud/environment/integrations/anthropic.py +196 -0
- hud/environment/integrations/gemini.py +92 -0
- hud/environment/integrations/langchain.py +82 -0
- hud/environment/integrations/llamaindex.py +68 -0
- hud/environment/integrations/openai.py +238 -0
- hud/environment/mock.py +306 -0
- hud/environment/router.py +263 -0
- hud/environment/scenarios.py +620 -0
- hud/environment/tests/__init__.py +1 -0
- hud/environment/tests/test_connection.py +317 -0
- hud/environment/tests/test_connectors.py +205 -0
- hud/environment/tests/test_environment.py +593 -0
- hud/environment/tests/test_integrations.py +257 -0
- hud/environment/tests/test_local_connectors.py +242 -0
- hud/environment/tests/test_scenarios.py +1086 -0
- hud/environment/tests/test_tools.py +208 -0
- hud/environment/types.py +23 -0
- hud/environment/utils/__init__.py +35 -0
- hud/environment/utils/formats.py +215 -0
- hud/environment/utils/schema.py +171 -0
- hud/environment/utils/tool_wrappers.py +113 -0
- hud/eval/__init__.py +67 -0
- hud/eval/context.py +727 -0
- hud/eval/display.py +299 -0
- hud/eval/instrument.py +187 -0
- hud/eval/manager.py +533 -0
- hud/eval/parallel.py +268 -0
- hud/eval/task.py +372 -0
- hud/eval/tests/__init__.py +1 -0
- hud/eval/tests/test_context.py +178 -0
- hud/eval/tests/test_eval.py +210 -0
- hud/eval/tests/test_manager.py +152 -0
- hud/eval/tests/test_parallel.py +168 -0
- hud/eval/tests/test_task.py +291 -0
- hud/eval/types.py +65 -0
- hud/eval/utils.py +194 -0
- hud/patches/__init__.py +19 -0
- hud/patches/mcp_patches.py +308 -0
- hud/patches/warnings.py +54 -0
- hud/samples/browser.py +4 -4
- hud/server/__init__.py +2 -1
- hud/server/low_level.py +2 -1
- hud/server/router.py +164 -0
- hud/server/server.py +567 -80
- hud/server/tests/test_mcp_server_integration.py +11 -11
- hud/server/tests/test_mcp_server_more.py +1 -1
- hud/server/tests/test_server_extra.py +2 -0
- hud/settings.py +45 -3
- hud/shared/exceptions.py +36 -10
- hud/shared/hints.py +26 -1
- hud/shared/requests.py +15 -3
- hud/shared/tests/test_exceptions.py +40 -31
- hud/shared/tests/test_hints.py +167 -0
- hud/telemetry/__init__.py +20 -19
- hud/telemetry/exporter.py +201 -0
- hud/telemetry/instrument.py +165 -253
- hud/telemetry/tests/test_eval_telemetry.py +356 -0
- hud/telemetry/tests/test_exporter.py +258 -0
- hud/telemetry/tests/test_instrument.py +401 -0
- hud/tools/__init__.py +18 -2
- hud/tools/agent.py +223 -0
- hud/tools/apply_patch.py +639 -0
- hud/tools/base.py +54 -4
- hud/tools/bash.py +2 -2
- hud/tools/computer/__init__.py +36 -3
- hud/tools/computer/anthropic.py +2 -2
- hud/tools/computer/gemini.py +385 -0
- hud/tools/computer/hud.py +23 -6
- hud/tools/computer/openai.py +20 -21
- hud/tools/computer/qwen.py +434 -0
- hud/tools/computer/settings.py +37 -0
- hud/tools/edit.py +3 -7
- hud/tools/executors/base.py +4 -2
- hud/tools/executors/pyautogui.py +1 -1
- hud/tools/grounding/grounded_tool.py +13 -18
- hud/tools/grounding/grounder.py +10 -31
- hud/tools/grounding/tests/test_grounded_tool.py +26 -44
- hud/tools/jupyter.py +330 -0
- hud/tools/playwright.py +18 -3
- hud/tools/shell.py +308 -0
- hud/tools/tests/test_agent_tool.py +355 -0
- hud/tools/tests/test_apply_patch.py +718 -0
- hud/tools/tests/test_computer.py +4 -9
- hud/tools/tests/test_computer_actions.py +24 -2
- hud/tools/tests/test_jupyter_tool.py +181 -0
- hud/tools/tests/test_shell.py +596 -0
- hud/tools/tests/test_submit.py +85 -0
- hud/tools/tests/test_types.py +193 -0
- hud/tools/types.py +21 -1
- hud/types.py +194 -56
- hud/utils/__init__.py +2 -0
- hud/utils/env.py +67 -0
- hud/utils/hud_console.py +89 -18
- hud/utils/mcp.py +15 -58
- hud/utils/strict_schema.py +162 -0
- hud/utils/tests/test_init.py +1 -2
- hud/utils/tests/test_mcp.py +1 -28
- hud/utils/tests/test_pretty_errors.py +186 -0
- hud/utils/tests/test_tool_shorthand.py +154 -0
- hud/utils/tests/test_version.py +1 -1
- hud/utils/types.py +20 -0
- hud/version.py +1 -1
- hud_python-0.5.13.dist-info/METADATA +264 -0
- hud_python-0.5.13.dist-info/RECORD +305 -0
- {hud_python-0.4.45.dist-info → hud_python-0.5.13.dist-info}/WHEEL +1 -1
- hud/agents/langchain.py +0 -261
- hud/agents/lite_llm.py +0 -72
- hud/cli/rl/__init__.py +0 -180
- hud/cli/rl/config.py +0 -101
- hud/cli/rl/display.py +0 -133
- hud/cli/rl/gpu.py +0 -63
- hud/cli/rl/gpu_utils.py +0 -321
- hud/cli/rl/local_runner.py +0 -595
- hud/cli/rl/presets.py +0 -96
- hud/cli/rl/remote_runner.py +0 -463
- hud/cli/rl/rl_api.py +0 -150
- hud/cli/rl/vllm.py +0 -177
- hud/cli/rl/wait_utils.py +0 -89
- hud/datasets/parallel.py +0 -687
- hud/misc/__init__.py +0 -1
- hud/misc/claude_plays_pokemon.py +0 -292
- hud/otel/__init__.py +0 -35
- hud/otel/collector.py +0 -142
- hud/otel/config.py +0 -181
- hud/otel/context.py +0 -570
- hud/otel/exporters.py +0 -369
- hud/otel/instrumentation.py +0 -135
- hud/otel/processors.py +0 -121
- hud/otel/tests/__init__.py +0 -1
- hud/otel/tests/test_processors.py +0 -197
- hud/rl/README.md +0 -30
- hud/rl/__init__.py +0 -1
- hud/rl/actor.py +0 -176
- hud/rl/buffer.py +0 -405
- hud/rl/chat_template.jinja +0 -101
- hud/rl/config.py +0 -192
- hud/rl/distributed.py +0 -132
- hud/rl/learner.py +0 -637
- hud/rl/tests/__init__.py +0 -1
- hud/rl/tests/test_learner.py +0 -186
- hud/rl/train.py +0 -382
- hud/rl/types.py +0 -101
- hud/rl/utils/start_vllm_server.sh +0 -30
- hud/rl/utils.py +0 -524
- hud/rl/vllm_adapter.py +0 -143
- hud/telemetry/job.py +0 -352
- hud/telemetry/replay.py +0 -74
- hud/telemetry/tests/test_replay.py +0 -40
- hud/telemetry/tests/test_trace.py +0 -63
- hud/telemetry/trace.py +0 -158
- hud/utils/agent_factories.py +0 -86
- hud/utils/async_utils.py +0 -65
- hud/utils/group_eval.py +0 -223
- hud/utils/progress.py +0 -149
- hud/utils/tasks.py +0 -127
- hud/utils/tests/test_async_utils.py +0 -173
- hud/utils/tests/test_progress.py +0 -261
- hud_python-0.4.45.dist-info/METADATA +0 -552
- hud_python-0.4.45.dist-info/RECORD +0 -228
- {hud_python-0.4.45.dist-info → hud_python-0.5.13.dist-info}/entry_points.txt +0 -0
- {hud_python-0.4.45.dist-info → hud_python-0.5.13.dist-info}/licenses/LICENSE +0 -0
hud/__init__.py
CHANGED
|
@@ -5,16 +5,36 @@ tools for building, evaluating, and training AI agents.
|
|
|
5
5
|
|
|
6
6
|
from __future__ import annotations
|
|
7
7
|
|
|
8
|
-
|
|
8
|
+
import warnings
|
|
9
|
+
|
|
10
|
+
# Apply patches to third-party libraries early, before other imports
|
|
11
|
+
from . import patches as _patches # noqa: F401
|
|
12
|
+
from .environment import Environment
|
|
13
|
+
from .eval import EvalContext
|
|
14
|
+
from .eval import run_eval as eval
|
|
15
|
+
from .telemetry.instrument import instrument
|
|
16
|
+
|
|
17
|
+
|
|
18
|
+
def trace(*args: object, **kwargs: object) -> EvalContext:
|
|
19
|
+
"""Deprecated: Use hud.eval() instead.
|
|
20
|
+
|
|
21
|
+
.. deprecated:: 0.5.2
|
|
22
|
+
hud.trace() is deprecated. Use hud.eval() or env.eval() instead.
|
|
23
|
+
"""
|
|
24
|
+
warnings.warn(
|
|
25
|
+
"hud.trace() is deprecated. Use hud.eval() or env.eval() instead.",
|
|
26
|
+
DeprecationWarning,
|
|
27
|
+
stacklevel=2,
|
|
28
|
+
)
|
|
29
|
+
return eval(*args, **kwargs) # type: ignore[arg-type]
|
|
30
|
+
|
|
9
31
|
|
|
10
32
|
__all__ = [
|
|
11
|
-
"
|
|
12
|
-
"
|
|
13
|
-
"
|
|
14
|
-
"get_trace",
|
|
33
|
+
"Environment",
|
|
34
|
+
"EvalContext",
|
|
35
|
+
"eval",
|
|
15
36
|
"instrument",
|
|
16
|
-
"
|
|
17
|
-
"trace",
|
|
37
|
+
"trace", # Deprecated alias for eval
|
|
18
38
|
]
|
|
19
39
|
|
|
20
40
|
try:
|
hud/agents/__init__.py
CHANGED
|
@@ -1,13 +1,78 @@
|
|
|
1
1
|
from __future__ import annotations
|
|
2
2
|
|
|
3
|
+
from typing import Any
|
|
4
|
+
|
|
3
5
|
from .base import MCPAgent
|
|
4
|
-
from .
|
|
5
|
-
from .
|
|
6
|
-
from .
|
|
6
|
+
from .openai import OpenAIAgent
|
|
7
|
+
from .openai_chat import OpenAIChatAgent
|
|
8
|
+
from .operator import OperatorAgent
|
|
7
9
|
|
|
8
10
|
__all__ = [
|
|
9
|
-
"ClaudeAgent",
|
|
10
|
-
"GenericOpenAIChatAgent",
|
|
11
11
|
"MCPAgent",
|
|
12
|
+
"OpenAIAgent",
|
|
13
|
+
"OpenAIChatAgent",
|
|
12
14
|
"OperatorAgent",
|
|
15
|
+
"create_agent",
|
|
13
16
|
]
|
|
17
|
+
|
|
18
|
+
|
|
19
|
+
def create_agent(model: str, **kwargs: Any) -> MCPAgent:
|
|
20
|
+
"""Create an agent for a gateway model.
|
|
21
|
+
|
|
22
|
+
This routes ALL requests through the HUD gateway. For direct API access
|
|
23
|
+
(using your own API keys), use the agent classes directly.
|
|
24
|
+
|
|
25
|
+
Args:
|
|
26
|
+
model: Model name (e.g., "gpt-4o", "claude-sonnet-4-5").
|
|
27
|
+
**kwargs: Additional params passed to agent.create().
|
|
28
|
+
|
|
29
|
+
Returns:
|
|
30
|
+
Configured MCPAgent instance with gateway routing.
|
|
31
|
+
|
|
32
|
+
Example:
|
|
33
|
+
```python
|
|
34
|
+
# Gateway routing (recommended)
|
|
35
|
+
agent = create_agent("gpt-4o")
|
|
36
|
+
agent = create_agent("claude-sonnet-4-5", temperature=0.7)
|
|
37
|
+
|
|
38
|
+
# Direct API access (use agent classes)
|
|
39
|
+
from hud.agents.claude import ClaudeAgent
|
|
40
|
+
|
|
41
|
+
agent = ClaudeAgent.create(model="claude-sonnet-4-5")
|
|
42
|
+
```
|
|
43
|
+
"""
|
|
44
|
+
from hud.agents.gateway import build_gateway_client
|
|
45
|
+
from hud.agents.resolver import resolve_cls
|
|
46
|
+
|
|
47
|
+
# Resolve class and gateway info
|
|
48
|
+
agent_cls, gateway_info = resolve_cls(model)
|
|
49
|
+
|
|
50
|
+
# Get model ID from gateway info or use input
|
|
51
|
+
model_id = model
|
|
52
|
+
if gateway_info:
|
|
53
|
+
model_id = gateway_info.get("model") or gateway_info.get("id") or model
|
|
54
|
+
|
|
55
|
+
# Determine provider: from gateway info, or infer from agent class
|
|
56
|
+
if gateway_info:
|
|
57
|
+
provider = gateway_info.get("provider") or "openai"
|
|
58
|
+
else:
|
|
59
|
+
provider = "openai"
|
|
60
|
+
if agent_cls.__name__ == "ClaudeAgent":
|
|
61
|
+
provider = "anthropic"
|
|
62
|
+
elif agent_cls.__name__ in ("GeminiAgent", "GeminiCUAAgent"):
|
|
63
|
+
provider = "gemini"
|
|
64
|
+
|
|
65
|
+
client = build_gateway_client(provider)
|
|
66
|
+
|
|
67
|
+
# Set up kwargs
|
|
68
|
+
kwargs.setdefault("model", model_id)
|
|
69
|
+
|
|
70
|
+
# Use correct client key based on agent type
|
|
71
|
+
if agent_cls == OpenAIChatAgent:
|
|
72
|
+
kwargs.setdefault("openai_client", client)
|
|
73
|
+
else:
|
|
74
|
+
# Claude and other agents use model_client and validate_api_key
|
|
75
|
+
kwargs.setdefault("model_client", client)
|
|
76
|
+
kwargs.setdefault("validate_api_key", False)
|
|
77
|
+
|
|
78
|
+
return agent_cls.create(**kwargs)
|