hud-python 0.3.4__py3-none-any.whl → 0.4.0__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.

Potentially problematic release.


This version of hud-python might be problematic. Click here for more details.

Files changed (192) hide show
  1. hud/__init__.py +22 -89
  2. hud/agents/__init__.py +17 -0
  3. hud/agents/art.py +101 -0
  4. hud/agents/base.py +599 -0
  5. hud/{mcp → agents}/claude.py +373 -321
  6. hud/{mcp → agents}/langchain.py +250 -250
  7. hud/agents/misc/__init__.py +7 -0
  8. hud/{agent → agents}/misc/response_agent.py +80 -80
  9. hud/{mcp → agents}/openai.py +352 -334
  10. hud/agents/openai_chat_generic.py +154 -0
  11. hud/{mcp → agents}/tests/__init__.py +1 -1
  12. hud/agents/tests/test_base.py +742 -0
  13. hud/agents/tests/test_claude.py +324 -0
  14. hud/{mcp → agents}/tests/test_client.py +363 -324
  15. hud/{mcp → agents}/tests/test_openai.py +237 -238
  16. hud/cli/__init__.py +617 -0
  17. hud/cli/__main__.py +8 -0
  18. hud/cli/analyze.py +371 -0
  19. hud/cli/analyze_metadata.py +230 -0
  20. hud/cli/build.py +427 -0
  21. hud/cli/clone.py +185 -0
  22. hud/cli/cursor.py +92 -0
  23. hud/cli/debug.py +392 -0
  24. hud/cli/docker_utils.py +83 -0
  25. hud/cli/init.py +281 -0
  26. hud/cli/interactive.py +353 -0
  27. hud/cli/mcp_server.py +756 -0
  28. hud/cli/pull.py +336 -0
  29. hud/cli/push.py +379 -0
  30. hud/cli/remote_runner.py +311 -0
  31. hud/cli/runner.py +160 -0
  32. hud/cli/tests/__init__.py +3 -0
  33. hud/cli/tests/test_analyze.py +284 -0
  34. hud/cli/tests/test_cli_init.py +265 -0
  35. hud/cli/tests/test_cli_main.py +27 -0
  36. hud/cli/tests/test_clone.py +142 -0
  37. hud/cli/tests/test_cursor.py +253 -0
  38. hud/cli/tests/test_debug.py +453 -0
  39. hud/cli/tests/test_mcp_server.py +139 -0
  40. hud/cli/tests/test_utils.py +388 -0
  41. hud/cli/utils.py +263 -0
  42. hud/clients/README.md +143 -0
  43. hud/clients/__init__.py +16 -0
  44. hud/clients/base.py +354 -0
  45. hud/clients/fastmcp.py +202 -0
  46. hud/clients/mcp_use.py +278 -0
  47. hud/clients/tests/__init__.py +1 -0
  48. hud/clients/tests/test_client_integration.py +111 -0
  49. hud/clients/tests/test_fastmcp.py +342 -0
  50. hud/clients/tests/test_protocol.py +188 -0
  51. hud/clients/utils/__init__.py +1 -0
  52. hud/clients/utils/retry_transport.py +160 -0
  53. hud/datasets.py +322 -192
  54. hud/misc/__init__.py +1 -0
  55. hud/{agent → misc}/claude_plays_pokemon.py +292 -283
  56. hud/otel/__init__.py +35 -0
  57. hud/otel/collector.py +142 -0
  58. hud/otel/config.py +164 -0
  59. hud/otel/context.py +536 -0
  60. hud/otel/exporters.py +366 -0
  61. hud/otel/instrumentation.py +97 -0
  62. hud/otel/processors.py +118 -0
  63. hud/otel/tests/__init__.py +1 -0
  64. hud/otel/tests/test_processors.py +197 -0
  65. hud/server/__init__.py +5 -5
  66. hud/server/context.py +114 -0
  67. hud/server/helper/__init__.py +5 -0
  68. hud/server/low_level.py +132 -0
  69. hud/server/server.py +166 -0
  70. hud/server/tests/__init__.py +3 -0
  71. hud/settings.py +73 -79
  72. hud/shared/__init__.py +5 -0
  73. hud/{exceptions.py → shared/exceptions.py} +180 -180
  74. hud/{server → shared}/requests.py +264 -264
  75. hud/shared/tests/test_exceptions.py +157 -0
  76. hud/{server → shared}/tests/test_requests.py +275 -275
  77. hud/telemetry/__init__.py +25 -30
  78. hud/telemetry/instrument.py +379 -0
  79. hud/telemetry/job.py +309 -141
  80. hud/telemetry/replay.py +74 -0
  81. hud/telemetry/trace.py +83 -0
  82. hud/tools/__init__.py +33 -34
  83. hud/tools/base.py +365 -65
  84. hud/tools/bash.py +161 -137
  85. hud/tools/computer/__init__.py +15 -13
  86. hud/tools/computer/anthropic.py +437 -414
  87. hud/tools/computer/hud.py +376 -328
  88. hud/tools/computer/openai.py +295 -286
  89. hud/tools/computer/settings.py +82 -0
  90. hud/tools/edit.py +314 -290
  91. hud/tools/executors/__init__.py +30 -30
  92. hud/tools/executors/base.py +539 -532
  93. hud/tools/executors/pyautogui.py +621 -619
  94. hud/tools/executors/tests/__init__.py +1 -1
  95. hud/tools/executors/tests/test_base_executor.py +338 -338
  96. hud/tools/executors/tests/test_pyautogui_executor.py +165 -165
  97. hud/tools/executors/xdo.py +511 -503
  98. hud/tools/{playwright_tool.py → playwright.py} +412 -379
  99. hud/tools/tests/__init__.py +3 -3
  100. hud/tools/tests/test_base.py +282 -0
  101. hud/tools/tests/test_bash.py +158 -152
  102. hud/tools/tests/test_bash_extended.py +197 -0
  103. hud/tools/tests/test_computer.py +425 -52
  104. hud/tools/tests/test_computer_actions.py +34 -34
  105. hud/tools/tests/test_edit.py +259 -240
  106. hud/tools/tests/test_init.py +27 -27
  107. hud/tools/tests/test_playwright_tool.py +183 -183
  108. hud/tools/tests/test_tools.py +145 -157
  109. hud/tools/tests/test_utils.py +156 -156
  110. hud/tools/types.py +72 -0
  111. hud/tools/utils.py +50 -50
  112. hud/types.py +136 -89
  113. hud/utils/__init__.py +10 -16
  114. hud/utils/async_utils.py +65 -0
  115. hud/utils/design.py +168 -0
  116. hud/utils/mcp.py +55 -0
  117. hud/utils/progress.py +149 -149
  118. hud/utils/telemetry.py +66 -66
  119. hud/utils/tests/test_async_utils.py +173 -0
  120. hud/utils/tests/test_init.py +17 -21
  121. hud/utils/tests/test_progress.py +261 -225
  122. hud/utils/tests/test_telemetry.py +82 -37
  123. hud/utils/tests/test_version.py +8 -8
  124. hud/version.py +7 -7
  125. hud_python-0.4.0.dist-info/METADATA +474 -0
  126. hud_python-0.4.0.dist-info/RECORD +132 -0
  127. hud_python-0.4.0.dist-info/entry_points.txt +3 -0
  128. {hud_python-0.3.4.dist-info → hud_python-0.4.0.dist-info}/licenses/LICENSE +21 -21
  129. hud/adapters/__init__.py +0 -8
  130. hud/adapters/claude/__init__.py +0 -5
  131. hud/adapters/claude/adapter.py +0 -180
  132. hud/adapters/claude/tests/__init__.py +0 -1
  133. hud/adapters/claude/tests/test_adapter.py +0 -519
  134. hud/adapters/common/__init__.py +0 -6
  135. hud/adapters/common/adapter.py +0 -178
  136. hud/adapters/common/tests/test_adapter.py +0 -289
  137. hud/adapters/common/types.py +0 -446
  138. hud/adapters/operator/__init__.py +0 -5
  139. hud/adapters/operator/adapter.py +0 -108
  140. hud/adapters/operator/tests/__init__.py +0 -1
  141. hud/adapters/operator/tests/test_adapter.py +0 -370
  142. hud/agent/__init__.py +0 -19
  143. hud/agent/base.py +0 -126
  144. hud/agent/claude.py +0 -271
  145. hud/agent/langchain.py +0 -215
  146. hud/agent/misc/__init__.py +0 -3
  147. hud/agent/operator.py +0 -268
  148. hud/agent/tests/__init__.py +0 -1
  149. hud/agent/tests/test_base.py +0 -202
  150. hud/env/__init__.py +0 -11
  151. hud/env/client.py +0 -35
  152. hud/env/docker_client.py +0 -349
  153. hud/env/environment.py +0 -446
  154. hud/env/local_docker_client.py +0 -358
  155. hud/env/remote_client.py +0 -212
  156. hud/env/remote_docker_client.py +0 -292
  157. hud/gym.py +0 -130
  158. hud/job.py +0 -773
  159. hud/mcp/__init__.py +0 -17
  160. hud/mcp/base.py +0 -631
  161. hud/mcp/client.py +0 -312
  162. hud/mcp/tests/test_base.py +0 -512
  163. hud/mcp/tests/test_claude.py +0 -294
  164. hud/task.py +0 -149
  165. hud/taskset.py +0 -237
  166. hud/telemetry/_trace.py +0 -347
  167. hud/telemetry/context.py +0 -230
  168. hud/telemetry/exporter.py +0 -575
  169. hud/telemetry/instrumentation/__init__.py +0 -3
  170. hud/telemetry/instrumentation/mcp.py +0 -259
  171. hud/telemetry/instrumentation/registry.py +0 -59
  172. hud/telemetry/mcp_models.py +0 -270
  173. hud/telemetry/tests/__init__.py +0 -1
  174. hud/telemetry/tests/test_context.py +0 -210
  175. hud/telemetry/tests/test_trace.py +0 -312
  176. hud/tools/helper/README.md +0 -56
  177. hud/tools/helper/__init__.py +0 -9
  178. hud/tools/helper/mcp_server.py +0 -78
  179. hud/tools/helper/server_initialization.py +0 -115
  180. hud/tools/helper/utils.py +0 -58
  181. hud/trajectory.py +0 -94
  182. hud/utils/agent.py +0 -37
  183. hud/utils/common.py +0 -256
  184. hud/utils/config.py +0 -120
  185. hud/utils/deprecation.py +0 -115
  186. hud/utils/misc.py +0 -53
  187. hud/utils/tests/test_common.py +0 -277
  188. hud/utils/tests/test_config.py +0 -129
  189. hud_python-0.3.4.dist-info/METADATA +0 -284
  190. hud_python-0.3.4.dist-info/RECORD +0 -120
  191. /hud/{adapters/common → shared}/tests/__init__.py +0 -0
  192. {hud_python-0.3.4.dist-info → hud_python-0.4.0.dist-info}/WHEEL +0 -0
hud/otel/collector.py ADDED
@@ -0,0 +1,142 @@
1
+ """Global span collector for building in-memory traces.
2
+
3
+ This module provides a way to collect spans during execution
4
+ and retrieve them as a Trace object, enabling replay functionality
5
+ without modifying agent code.
6
+ """
7
+
8
+ from __future__ import annotations
9
+
10
+ import logging
11
+ import threading
12
+ from contextvars import ContextVar
13
+ from typing import TYPE_CHECKING
14
+
15
+ from opentelemetry import trace
16
+ from opentelemetry.sdk.trace.export import SpanExporter, SpanExportResult
17
+
18
+ from hud.types import Trace
19
+
20
+ if TYPE_CHECKING:
21
+ from opentelemetry.sdk.trace import ReadableSpan
22
+
23
+ logger = logging.getLogger(__name__)
24
+
25
+ # Global storage for collected spans by task_run_id
26
+ _TRACE_STORAGE: dict[str, TraceCollector] = {}
27
+ _LOCK = threading.Lock()
28
+
29
+ # Context variable to track if collection is enabled
30
+ _collecting_enabled: ContextVar[bool] = ContextVar("collecting_enabled", default=False)
31
+
32
+
33
+ class TraceCollector:
34
+ """Collects spans for a single task run."""
35
+
36
+ def __init__(self, task_run_id: str) -> None:
37
+ self.task_run_id = task_run_id
38
+ self.spans: list[ReadableSpan] = []
39
+ self._lock = threading.Lock()
40
+
41
+ def add_span(self, span: ReadableSpan) -> None:
42
+ """Thread-safe span addition."""
43
+ with self._lock:
44
+ self.spans.append(span)
45
+
46
+ def to_trace(self) -> Trace:
47
+ """Convert collected spans to a Trace object."""
48
+ from .exporters import HudSpan, _span_to_dict
49
+
50
+ trace = Trace()
51
+
52
+ # Convert spans to TraceSteps
53
+ for span in self.spans:
54
+ try:
55
+ # Use the same conversion logic as the exporter
56
+ span_dict = _span_to_dict(span)
57
+ hud_span = HudSpan.model_validate(span_dict)
58
+
59
+ # The attributes field is already a TraceStep
60
+ step = hud_span.attributes
61
+ # Add timing from the span itself
62
+ step.start_timestamp = hud_span.start_time
63
+ step.end_timestamp = hud_span.end_time
64
+ trace.append(step)
65
+
66
+ except Exception as e:
67
+ # Log but don't fail the whole trace
68
+ logger.debug("Failed to convert span: %s", e)
69
+
70
+ return trace
71
+
72
+
73
+ class CollectingSpanExporter(SpanExporter):
74
+ """A span exporter that collects spans in memory for replay."""
75
+
76
+ def export(self, spans: list[ReadableSpan]) -> SpanExportResult:
77
+ """Collect spans if collection is enabled."""
78
+ if not _collecting_enabled.get():
79
+ return SpanExportResult.SUCCESS
80
+
81
+ for span in spans:
82
+ # Extract task_run_id from span
83
+ task_run_id = span.attributes.get("hud.task_run_id") if span.attributes else None
84
+ if not task_run_id or not isinstance(task_run_id, str):
85
+ continue
86
+
87
+ # Get or create collector
88
+ with _LOCK:
89
+ if task_run_id not in _TRACE_STORAGE:
90
+ _TRACE_STORAGE[task_run_id] = TraceCollector(task_run_id)
91
+ collector = _TRACE_STORAGE[task_run_id]
92
+
93
+ # Add span
94
+ collector.add_span(span)
95
+
96
+ return SpanExportResult.SUCCESS
97
+
98
+ def shutdown(self) -> None:
99
+ """Clean up resources."""
100
+ with _LOCK:
101
+ _TRACE_STORAGE.clear()
102
+
103
+
104
+ def enable_trace_collection(enabled: bool = True) -> None:
105
+ """Enable or disable in-memory trace collection."""
106
+ _collecting_enabled.set(enabled)
107
+
108
+
109
+ def get_trace(task_run_id: str) -> Trace | None:
110
+ """Retrieve collected trace for a task run ID.
111
+
112
+ Returns None if no trace was collected or collection was disabled.
113
+ """
114
+ with _LOCK:
115
+ collector = _TRACE_STORAGE.get(task_run_id)
116
+ if collector:
117
+ return collector.to_trace()
118
+ return None
119
+
120
+
121
+ def clear_trace(task_run_id: str) -> None:
122
+ """Clear collected trace for a task run ID."""
123
+ with _LOCK:
124
+ _TRACE_STORAGE.pop(task_run_id, None)
125
+
126
+
127
+ def install_collector() -> None:
128
+ """Install the collecting span exporter.
129
+
130
+ This should be called after configure_telemetry().
131
+ """
132
+ provider = trace.get_tracer_provider()
133
+ # Guard for SDK tracer providers only
134
+ if hasattr(provider, "add_span_processor"):
135
+ from opentelemetry.sdk.trace.export import SimpleSpanProcessor
136
+
137
+ exporter = CollectingSpanExporter()
138
+ processor = SimpleSpanProcessor(exporter)
139
+ try:
140
+ provider.add_span_processor(processor) # type: ignore[attr-defined]
141
+ except Exception:
142
+ logger.warning("Failed to add span processor")
hud/otel/config.py ADDED
@@ -0,0 +1,164 @@
1
+ """Central configuration for OpenTelemetry inside HUD SDK.
2
+
3
+ This file is responsible for
4
+ 1. creating the global ``TracerProvider``
5
+ 2. attaching span processors (HUD enrichment, batch + exporter)
6
+ 3. activating the community MCP instrumentation so that *every* MCP
7
+ request/response/notification is traced automatically.
8
+
9
+ It is *idempotent*: calling :func:`configure_telemetry` more than once
10
+ returns the same provider and does nothing.
11
+ """
12
+
13
+ from __future__ import annotations
14
+
15
+ import logging
16
+ from typing import Any
17
+
18
+ from opentelemetry import trace
19
+ from opentelemetry.sdk.resources import Resource
20
+ from opentelemetry.sdk.trace import TracerProvider
21
+ from opentelemetry.sdk.trace.export import BatchSpanProcessor
22
+
23
+ from hud.settings import settings
24
+
25
+ from .collector import enable_trace_collection, install_collector
26
+ from .exporters import HudSpanExporter
27
+ from .instrumentation import install_mcp_instrumentation
28
+ from .processors import HudEnrichmentProcessor
29
+
30
+ logger = logging.getLogger(__name__)
31
+
32
+ # Global singleton provider so multiple calls do not create duplicates
33
+ _TRACER_PROVIDER: TracerProvider | None = None
34
+
35
+
36
+ def is_telemetry_configured() -> bool:
37
+ """Check if telemetry has been configured."""
38
+ return _TRACER_PROVIDER is not None
39
+
40
+
41
+ # ---------------------------------------------------------------------------
42
+ # Public API
43
+ # ---------------------------------------------------------------------------
44
+
45
+
46
+ def configure_telemetry(
47
+ *,
48
+ service_name: str = "hud-sdk",
49
+ service_version: str | None = None,
50
+ environment: str | None = None,
51
+ extra_resource_attributes: dict[str, Any] | None = None,
52
+ enable_otlp: bool = False,
53
+ otlp_endpoint: str | None = None,
54
+ otlp_headers: dict[str, str] | None = None,
55
+ enable_collection: bool = True,
56
+ ) -> TracerProvider:
57
+ """Initialise OpenTelemetry for the current Python process.
58
+
59
+ It is safe to call this in every entry-point; the provider will only
60
+ be created once.
61
+ """
62
+ global _TRACER_PROVIDER
63
+
64
+ if _TRACER_PROVIDER is not None:
65
+ return _TRACER_PROVIDER
66
+
67
+ # ------------------------------------------------------------------
68
+ # 1. Resource (identity of this service)
69
+ # ------------------------------------------------------------------
70
+ res_attrs: dict[str, Any] = {
71
+ "service.name": service_name,
72
+ "telemetry.sdk.name": "hud-otel",
73
+ "telemetry.sdk.language": "python",
74
+ }
75
+ if service_version:
76
+ res_attrs["service.version"] = service_version
77
+ if environment:
78
+ res_attrs["deployment.environment"] = environment
79
+ if extra_resource_attributes:
80
+ res_attrs.update(extra_resource_attributes)
81
+
82
+ resource = Resource.create(res_attrs)
83
+
84
+ # ------------------------------------------------------------------
85
+ # 2. Provider
86
+ # ------------------------------------------------------------------
87
+ provider = TracerProvider(resource=resource)
88
+ _TRACER_PROVIDER = provider
89
+
90
+ # ------------------------------------------------------------------
91
+ # 3. Processors / exporters
92
+ # ------------------------------------------------------------------
93
+ provider.add_span_processor(HudEnrichmentProcessor())
94
+
95
+ # HUD exporter (only if enabled and API key is available)
96
+ if settings.telemetry_enabled and settings.api_key:
97
+ exporter = HudSpanExporter(
98
+ telemetry_url=settings.hud_telemetry_url, api_key=settings.api_key
99
+ )
100
+ provider.add_span_processor(BatchSpanProcessor(exporter))
101
+ elif settings.telemetry_enabled and not settings.api_key and not enable_otlp:
102
+ # Error if no exporters are configured
103
+ raise ValueError(
104
+ "No telemetry backend configured. Either:\n"
105
+ "1. Set HUD_API_KEY environment variable for HUD telemetry\n"
106
+ "2. Use enable_otlp=True with configure_telemetry() for alternative backends (e.g., Jaeger)\n" # noqa: E501
107
+ )
108
+ elif not settings.telemetry_enabled:
109
+ logger.info("HUD telemetry disabled via HUD_TELEMETRY_ENABLED=false")
110
+
111
+ # OTLP exporter (optional - for standard OTel viewers)
112
+ if enable_otlp:
113
+ try:
114
+ from opentelemetry.exporter.otlp.proto.http.trace_exporter import OTLPSpanExporter
115
+
116
+ otlp_config = {}
117
+ if otlp_endpoint:
118
+ otlp_config["endpoint"] = otlp_endpoint
119
+ # Default to HTTP endpoint if not specified
120
+ if not otlp_endpoint.startswith(("http://", "https://")):
121
+ otlp_config["endpoint"] = f"http://{otlp_endpoint}/v1/traces"
122
+ else:
123
+ # Default HTTP endpoint
124
+ otlp_config["endpoint"] = "http://localhost:4318/v1/traces"
125
+
126
+ if otlp_headers:
127
+ otlp_config["headers"] = otlp_headers
128
+
129
+ otlp_exporter = OTLPSpanExporter(**otlp_config)
130
+ provider.add_span_processor(BatchSpanProcessor(otlp_exporter))
131
+ logger.info("OTLP HTTP exporter enabled - endpoint: %s", otlp_config["endpoint"])
132
+ except ImportError:
133
+ logger.warning(
134
+ "OTLP export requested but opentelemetry-exporter-otlp-proto-http not installed. "
135
+ "Install with: pip install 'hud-python[agent]'"
136
+ )
137
+
138
+ # ------------------------------------------------------------------
139
+ # 4. Activate provider and instrumentation
140
+ # ------------------------------------------------------------------
141
+ trace.set_tracer_provider(provider)
142
+ install_mcp_instrumentation(provider)
143
+
144
+ # Install in-memory collector if requested
145
+ if enable_collection:
146
+ install_collector()
147
+ enable_trace_collection(True)
148
+ logger.debug("In-memory trace collection enabled")
149
+
150
+ # Agent instrumentation now handled by @hud.instrument decorators
151
+ logger.debug("OpenTelemetry configuration completed")
152
+
153
+ logger.debug("OpenTelemetry configured (provider id=%s)", id(provider))
154
+ return provider
155
+
156
+
157
+ def shutdown_telemetry() -> None:
158
+ """Flush and shutdown the global provider (if configured)."""
159
+ global _TRACER_PROVIDER
160
+ if _TRACER_PROVIDER is None:
161
+ return
162
+ _TRACER_PROVIDER.shutdown() # type: ignore[arg-type]
163
+ _TRACER_PROVIDER = None
164
+ logger.debug("OpenTelemetry shutdown complete")