@lmmslab/harness 0.0.0-stage → 0.81.1
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.
- package/LICENSE +21 -0
- package/README.md +124 -2
- package/bin/harness.js +71 -0
- package/package.json +34 -4
- package/python/harness/README.md +72 -0
- package/python/harness/__init__.py +8 -0
- package/python/harness/__main__.py +3 -0
- package/python/harness/bench.py +124 -0
- package/python/harness/builtin/benches/README.md +5 -0
- package/python/harness/builtin/benches/none/bench.toml +6 -0
- package/python/harness/builtin/configs/README.md +26 -0
- package/python/harness/builtin/configs/assistant.toml +11 -0
- package/python/harness/builtin/configs/default.toml +33 -0
- package/python/harness/builtin/configs/deploy/example-api.toml +10 -0
- package/python/harness/builtin/configs/models/deepseek-api.toml +21 -0
- package/python/harness/builtin/configs/models/deepseek-flash.toml +12 -0
- package/python/harness/builtin/configs/models/grok-low.toml +8 -0
- package/python/harness/builtin/configs/models/grok-medium.toml +43 -0
- package/python/harness/builtin/configs/models/opus-5.5.toml +11 -0
- package/python/harness/builtin/environment/README.md +53 -0
- package/python/harness/builtin/environment/constraints.txt +31 -0
- package/python/harness/builtin/model-settings.js +143 -0
- package/python/harness/builtin/prompts/conversation/core.md +5 -0
- package/python/harness/builtin/prompts/conversation/pack.toml +12 -0
- package/python/harness/builtin/prompts/core/core.md +88 -0
- package/python/harness/builtin/prompts/core/pack.toml +9 -0
- package/python/harness/builtin/prompts/reflect/core.md +45 -0
- package/python/harness/builtin/prompts/reflect/pack.toml +7 -0
- package/python/harness/builtin/prompts/shared/agent-skills.md +5 -0
- package/python/harness/builtin/prompts/shared/agents.md +14 -0
- package/python/harness/builtin/prompts/shared/checklist.md +21 -0
- package/python/harness/builtin/prompts/shared/entries.md +86 -0
- package/python/harness/builtin/prompts/shared/environments.md +21 -0
- package/python/harness/builtin/prompts/shared/learning-goal.md +31 -0
- package/python/harness/builtin/prompts/shared/learning.md +27 -0
- package/python/harness/builtin/prompts/shared/mcp.md +9 -0
- package/python/harness/builtin/prompts/shared/world.md +1 -0
- package/python/harness/builtin/skills/README.md +10 -0
- package/python/harness/builtin/skills/agents/__init__.py +88 -0
- package/python/harness/builtin/skills/agents/skill.toml +11 -0
- package/python/harness/builtin/skills/checklist/__init__.py +32 -0
- package/python/harness/builtin/skills/checklist/skill.toml +9 -0
- package/python/harness/builtin/skills/conversation/__init__.py +23 -0
- package/python/harness/builtin/skills/conversation/skill.toml +8 -0
- package/python/harness/builtin/skills/entries/__init__.py +62 -0
- package/python/harness/builtin/skills/entries/skill.toml +8 -0
- package/python/harness/builtin/skills/environments/__init__.py +69 -0
- package/python/harness/builtin/skills/environments/skill.toml +9 -0
- package/python/harness/builtin/skills/experience/__init__.py +27 -0
- package/python/harness/builtin/skills/experience/skill.toml +9 -0
- package/python/harness/builtin/skills/instructions/__init__.py +29 -0
- package/python/harness/builtin/skills/instructions/skill.toml +9 -0
- package/python/harness/builtin/skills/learning/__init__.py +189 -0
- package/python/harness/builtin/skills/learning/skill.toml +7 -0
- package/python/harness/builtin/skills/mcp/__init__.py +54 -0
- package/python/harness/builtin/skills/mcp/skill.toml +10 -0
- package/python/harness/builtin/skills/memory/__init__.py +130 -0
- package/python/harness/builtin/skills/memory/skill.toml +13 -0
- package/python/harness/builtin/skills/verify/__init__.py +31 -0
- package/python/harness/builtin/skills/verify/skill.toml +7 -0
- package/python/harness/builtin/third_party/sources.toml +40 -0
- package/python/harness/cli.py +1182 -0
- package/python/harness/config/__init__.py +224 -0
- package/python/harness/config/agent.py +777 -0
- package/python/harness/core/__init__.py +52 -0
- package/python/harness/core/artifacts.py +46 -0
- package/python/harness/core/codec.py +160 -0
- package/python/harness/core/contracts.py +75 -0
- package/python/harness/core/episodes.py +150 -0
- package/python/harness/core/format.py +51 -0
- package/python/harness/core/http.py +124 -0
- package/python/harness/core/lanes.py +324 -0
- package/python/harness/core/linux_sandbox.py +55 -0
- package/python/harness/core/media.py +96 -0
- package/python/harness/core/memory.py +147 -0
- package/python/harness/core/ports.py +63 -0
- package/python/harness/core/procs.py +227 -0
- package/python/harness/core/remote.py +102 -0
- package/python/harness/core/spans.py +166 -0
- package/python/harness/core/trace.py +139 -0
- package/python/harness/core/wait.py +12 -0
- package/python/harness/core/worker.py +388 -0
- package/python/harness/core/worker_pipe.py +119 -0
- package/python/harness/embodiments.py +217 -0
- package/python/harness/extensions.py +112 -0
- package/python/harness/host/__init__.py +397 -0
- package/python/harness/host/client.py +201 -0
- package/python/harness/host/http.py +720 -0
- package/python/harness/host/listing.py +68 -0
- package/python/harness/host/server.py +189 -0
- package/python/harness/host/sessions.py +406 -0
- package/python/harness/host/tasks.py +1356 -0
- package/python/harness/host/workers.py +257 -0
- package/python/harness/host/workspaces.py +256 -0
- package/python/harness/model_list.py +117 -0
- package/python/harness/packages/__init__.py +73 -0
- package/python/harness/packages/addons.py +121 -0
- package/python/harness/packages/archive.py +122 -0
- package/python/harness/packages/describe.py +46 -0
- package/python/harness/packages/environments.py +448 -0
- package/python/harness/packages/fetch.py +57 -0
- package/python/harness/packages/manager.py +587 -0
- package/python/harness/packages/manifest.py +232 -0
- package/python/harness/packages/migrate.py +68 -0
- package/python/harness/packages/provision.py +38 -0
- package/python/harness/packages/registry.py +275 -0
- package/python/harness/packages/resolve.py +89 -0
- package/python/harness/packages/sources.py +141 -0
- package/python/harness/packages/state.py +213 -0
- package/python/harness/packages/versions.py +112 -0
- package/python/harness/paths.py +218 -0
- package/python/harness/providers.py +303 -0
- package/python/harness/runtime/__init__.py +8 -0
- package/python/harness/runtime/decision/__init__.py +1 -0
- package/python/harness/runtime/decision/loop.py +198 -0
- package/python/harness/runtime/decision/messages.py +271 -0
- package/python/harness/runtime/decision/model.py +424 -0
- package/python/harness/runtime/decision/prompts.py +248 -0
- package/python/harness/runtime/decision/protocol.py +169 -0
- package/python/harness/runtime/decision/thoughts.py +199 -0
- package/python/harness/runtime/events.py +213 -0
- package/python/harness/runtime/execution/__init__.py +1 -0
- package/python/harness/runtime/execution/attached.py +96 -0
- package/python/harness/runtime/execution/control.py +211 -0
- package/python/harness/runtime/execution/entries.py +665 -0
- package/python/harness/runtime/execution/kernel.py +811 -0
- package/python/harness/runtime/execution/noenv.py +84 -0
- package/python/harness/runtime/execution/results.py +40 -0
- package/python/harness/runtime/execution/session.py +871 -0
- package/python/harness/runtime/execution/skills.py +436 -0
- package/python/harness/runtime/learning/__init__.py +1 -0
- package/python/harness/runtime/learning/active.py +314 -0
- package/python/harness/runtime/learning/experience.py +48 -0
- package/python/harness/runtime/learning/fixtures.py +167 -0
- package/python/harness/runtime/learning/learner.py +651 -0
- package/python/harness/runtime/learning/library.py +504 -0
- package/python/harness/runtime/learning/reflect.py +151 -0
- package/python/harness/runtime/learning/usage.py +30 -0
- package/python/harness/runtime/learning/views.py +89 -0
- package/python/harness/runtime/reactive/__init__.py +1 -0
- package/python/harness/runtime/reactive/rules.py +549 -0
- package/python/harness/runtime/safety/__init__.py +1 -0
- package/python/harness/runtime/safety/approval.py +131 -0
- package/python/harness/runtime/safety/input.py +62 -0
- package/python/harness/runtime/safety/isolation.py +419 -0
- package/python/harness/runtime/safety/shield.py +129 -0
- package/python/harness/runtime/status.py +23 -0
- package/python/harness/runtime/task/__init__.py +7 -0
- package/python/harness/runtime/task/context.py +431 -0
- package/python/harness/runtime/task/delegation.py +468 -0
- package/python/harness/runtime/task/environments.py +268 -0
- package/python/harness/runtime/task/run.py +2878 -0
- package/python/harness/runtime/tick.py +909 -0
- package/python/harness/sdk/__init__.py +281 -0
- package/python/harness/services/__init__.py +6 -0
- package/python/harness/services/broker.py +35 -0
- package/python/harness/services/connectors.py +88 -0
- package/python/harness/services/mcp.py +298 -0
- package/python/harness/services/mcp_stdio.py +131 -0
- package/python/harness/services/placement.py +82 -0
- package/python/harness/services/registry.py +156 -0
- package/python/harness/settings.py +536 -0
- package/python/harness/terminal.py +79 -0
- package/python/harness/transcript.py +297 -0
- package/python/harness/tui/__init__.py +68 -0
- package/python/harness/tui/chat.py +1181 -0
- package/python/harness/tui/pty.py +81 -0
- package/python/harness/tui/renderer.py +45 -0
- package/python/harness/tui/timeline_rows.py +58 -0
- package/python/harness/tui/views.py +434 -0
|
@@ -0,0 +1,213 @@
|
|
|
1
|
+
"""Events: the signals of a run, in one catalog.
|
|
2
|
+
|
|
3
|
+
An execution's event (Event) ends every bounded execution and goes up the tree of entries to
|
|
4
|
+
the rules that answer it (on()); one nobody answers escalates to the model as an Escalation.
|
|
5
|
+
A trace record (EVENTS) is what the run writes to its trace (core/trace.py), which the
|
|
6
|
+
timeline, the overlay, replay and the exported video all draw. A family of records is one
|
|
7
|
+
kind with an `action` (library), which record_kind() reads.
|
|
8
|
+
"""
|
|
9
|
+
|
|
10
|
+
from dataclasses import asdict, dataclass, field
|
|
11
|
+
from typing import Any
|
|
12
|
+
|
|
13
|
+
|
|
14
|
+
# The kinds of an execution's event (Event.kind; a guard may raise one of its own name):
|
|
15
|
+
# nominal ends (NOMINAL), a time bound, the execution's scope, a guard, the shield, reactive
|
|
16
|
+
# code that broke its rules, a skill's error, the arm held back short of its target (contact
|
|
17
|
+
# or a load), the episode's end, the user's pause or message (both always reach the model; the
|
|
18
|
+
# message stops the program's own code, not the entries it started), an entry that failed,
|
|
19
|
+
# and a channel another entry has.
|
|
20
|
+
BUILTIN = ("COMPLETED", "PREDICATE_SATISFIED", "TIMEOUT", "REGION_EXIT", "GUARD_TRIGGERED", "SHIELD_REJECTED",
|
|
21
|
+
"REACTIVE_FAULT", "SKILL_ERROR", "BLOCKED", "EPISODE_ENDED", "USER_PAUSED", "USER_MESSAGE",
|
|
22
|
+
"ENTRY_FAILED", "CHANNEL_CONFLICT")
|
|
23
|
+
(COMPLETED, PREDICATE_SATISFIED, TIMEOUT, REGION_EXIT, GUARD_TRIGGERED, SHIELD_REJECTED, REACTIVE_FAULT, SKILL_ERROR,
|
|
24
|
+
BLOCKED, EPISODE_ENDED, USER_PAUSED, USER_MESSAGE, ENTRY_FAILED, CHANNEL_CONFLICT) = BUILTIN
|
|
25
|
+
# A program may continue after these without consulting a handler.
|
|
26
|
+
NOMINAL = frozenset({COMPLETED, PREDICATE_SATISFIED})
|
|
27
|
+
|
|
28
|
+
|
|
29
|
+
@dataclass
|
|
30
|
+
class Event:
|
|
31
|
+
"""Why an execution ended: its kind (a built-in kind above or a guard's own action),
|
|
32
|
+
the source that raised it (a skill, a guard, "shield", "environment"), details, and
|
|
33
|
+
the robot step it happened at."""
|
|
34
|
+
|
|
35
|
+
kind: str
|
|
36
|
+
source: str
|
|
37
|
+
detail: dict[str, Any] = field(default_factory=dict)
|
|
38
|
+
tick: int | None = None
|
|
39
|
+
|
|
40
|
+
def to_dict(self) -> dict:
|
|
41
|
+
"""The event as a plain dict."""
|
|
42
|
+
return asdict(self)
|
|
43
|
+
|
|
44
|
+
|
|
45
|
+
def input_refusal(receipt, step):
|
|
46
|
+
if conflict := receipt.get("conflict"):
|
|
47
|
+
return Event(CHANNEL_CONFLICT, "channels", conflict, step)
|
|
48
|
+
if not receipt["executed"]:
|
|
49
|
+
return Event(SHIELD_REJECTED, "shield", {"reasons": receipt["shield"]["reasons"]}, step)
|
|
50
|
+
|
|
51
|
+
|
|
52
|
+
@dataclass
|
|
53
|
+
class ExecutionTrace:
|
|
54
|
+
"""What one bounded execution did, and why it ended."""
|
|
55
|
+
|
|
56
|
+
skill: str
|
|
57
|
+
parameters: dict[str, Any]
|
|
58
|
+
start_tick: int
|
|
59
|
+
end_tick: int
|
|
60
|
+
event: Event
|
|
61
|
+
samples: list[dict] = field(default_factory=list)
|
|
62
|
+
# A value for the calling program, never part of an event or serialized trace.
|
|
63
|
+
result: Any = field(default=None, repr=False)
|
|
64
|
+
|
|
65
|
+
@property
|
|
66
|
+
def ok(self) -> bool:
|
|
67
|
+
"""Whether the execution ended nominally (COMPLETED or PREDICATE_SATISFIED)."""
|
|
68
|
+
return self.event.kind in NOMINAL
|
|
69
|
+
|
|
70
|
+
@property
|
|
71
|
+
def ticks(self) -> int:
|
|
72
|
+
"""Robot steps the execution took."""
|
|
73
|
+
return self.end_tick - self.start_tick
|
|
74
|
+
|
|
75
|
+
def to_dict(self) -> dict:
|
|
76
|
+
"""The trace as a plain dict, with `ok`."""
|
|
77
|
+
data = {"skill": self.skill, "parameters": self.parameters, "start_tick": self.start_tick,
|
|
78
|
+
"end_tick": self.end_tick, "event": self.event.to_dict(), "samples": self.samples}
|
|
79
|
+
data["ok"] = self.ok
|
|
80
|
+
return data
|
|
81
|
+
|
|
82
|
+
def __repr__(self) -> str:
|
|
83
|
+
return (f"Trace({self.skill}, {self.event.kind} from {self.event.source}, "
|
|
84
|
+
f"ticks={self.ticks})")
|
|
85
|
+
|
|
86
|
+
|
|
87
|
+
class Escalation(Exception):
|
|
88
|
+
"""An event no handler resolved. It ends the current program."""
|
|
89
|
+
|
|
90
|
+
def __init__(self, trace: ExecutionTrace):
|
|
91
|
+
super().__init__(f"{trace.event.kind} in {trace.skill}")
|
|
92
|
+
self.trace = trace
|
|
93
|
+
|
|
94
|
+
|
|
95
|
+
class ProgramFinished(Exception):
|
|
96
|
+
"""A program called finish(): its own checks say the task is done. The episode
|
|
97
|
+
ends without another model call."""
|
|
98
|
+
|
|
99
|
+
def __init__(self, summary: str = ""):
|
|
100
|
+
super().__init__(summary)
|
|
101
|
+
self.summary = str(summary)
|
|
102
|
+
|
|
103
|
+
|
|
104
|
+
class EpisodeEnded(Exception):
|
|
105
|
+
"""The environment terminated or truncated the episode during a program."""
|
|
106
|
+
|
|
107
|
+
def __init__(self, trace: ExecutionTrace):
|
|
108
|
+
super().__init__("episode ended")
|
|
109
|
+
self.trace = trace
|
|
110
|
+
|
|
111
|
+
|
|
112
|
+
# Every kind of trace record: its source (the user; the agent, i.e. the main model; the program
|
|
113
|
+
# running in the kernel; the environment; the harness itself) and whether what it holds is part
|
|
114
|
+
# of what the model is shown. What each record holds is in the traces reference.
|
|
115
|
+
EVENTS = {
|
|
116
|
+
"execution_event": {"source": "program", "shown": False},
|
|
117
|
+
"program_output": {"source": "program", "shown": False},
|
|
118
|
+
"state": {"source": "harness", "shown": False},
|
|
119
|
+
"reset": {"source": "environment", "shown": True},
|
|
120
|
+
"screen_recording": {"source": "harness", "shown": False},
|
|
121
|
+
"library": {"source": "harness", "shown": True},
|
|
122
|
+
"user_message": {"source": "user", "shown": True},
|
|
123
|
+
"model_request": {"source": "harness", "shown": True},
|
|
124
|
+
"model_changed": {"source": "user", "shown": False},
|
|
125
|
+
"ahead": {"source": "harness", "shown": False},
|
|
126
|
+
"context_shortened": {"source": "harness", "shown": False},
|
|
127
|
+
"side_program": {"source": "harness", "shown": False},
|
|
128
|
+
"side_program_end": {"source": "harness", "shown": False},
|
|
129
|
+
"program_stopped": {"source": "harness", "shown": False},
|
|
130
|
+
"model_preempted": {"source": "harness", "shown": False},
|
|
131
|
+
"model_reply": {"source": "agent", "shown": True},
|
|
132
|
+
"agent_message": {"source": "agent", "shown": True},
|
|
133
|
+
"empty_reply": {"source": "harness", "shown": False},
|
|
134
|
+
"protocol_error": {"source": "harness", "shown": True},
|
|
135
|
+
"malformed_dropped": {"source": "harness", "shown": False},
|
|
136
|
+
"verify": {"source": "environment", "shown": True},
|
|
137
|
+
"confirmation_request": {"source": "harness", "shown": False},
|
|
138
|
+
"confirmation": {"source": "user", "shown": True},
|
|
139
|
+
"physical_request": {"source": "harness", "shown": False},
|
|
140
|
+
"physical_answer": {"source": "user", "shown": False},
|
|
141
|
+
"worker_approval_result": {"source": "harness", "shown": False},
|
|
142
|
+
"input": {"source": "environment", "shown": False},
|
|
143
|
+
"cell": {"source": "program", "shown": True},
|
|
144
|
+
"private_paths": {"source": "harness", "shown": False},
|
|
145
|
+
"file_access": {"source": "program", "shown": False},
|
|
146
|
+
"modules": {"source": "harness", "shown": False},
|
|
147
|
+
"experience": {"source": "program", "shown": False},
|
|
148
|
+
"experience_folded": {"source": "harness", "shown": False},
|
|
149
|
+
"tick_periods": {"source": "program", "shown": False},
|
|
150
|
+
"guard": {"source": "program", "shown": False},
|
|
151
|
+
"program_held": {"source": "harness", "shown": False},
|
|
152
|
+
"program_dropped": {"source": "harness", "shown": False},
|
|
153
|
+
"user_took_over": {"source": "harness", "shown": False},
|
|
154
|
+
"stuck": {"source": "harness", "shown": False},
|
|
155
|
+
"policy_to_end": {"source": "program", "shown": False},
|
|
156
|
+
"finish": {"source": "agent", "shown": False},
|
|
157
|
+
"finish_refused": {"source": "harness", "shown": True},
|
|
158
|
+
"wrap_up": {"source": "program", "shown": False},
|
|
159
|
+
"kernel_died": {"source": "harness", "shown": False},
|
|
160
|
+
"interrupted": {"source": "harness", "shown": False},
|
|
161
|
+
"error": {"source": "harness", "shown": False},
|
|
162
|
+
"cleanup_error": {"source": "harness", "shown": False},
|
|
163
|
+
"outcome": {"source": "environment", "shown": False},
|
|
164
|
+
"episode_video": {"source": "environment", "shown": False},
|
|
165
|
+
"late_receipt": {"source": "environment", "shown": False},
|
|
166
|
+
"plan": {"source": "program", "shown": False},
|
|
167
|
+
"stay_on_screen": {"source": "program", "shown": False},
|
|
168
|
+
"decision": {"source": "program", "shown": False},
|
|
169
|
+
"page_read": {"source": "environment", "shown": False},
|
|
170
|
+
"workspace_reclaimed": {"source": "harness", "shown": False},
|
|
171
|
+
"environment_attached": {"source": "harness", "shown": True},
|
|
172
|
+
"environment_refused": {"source": "harness", "shown": True},
|
|
173
|
+
"environment_released": {"source": "harness", "shown": True},
|
|
174
|
+
"agents_enabled": {"source": "harness", "shown": True},
|
|
175
|
+
"views": {"source": "program", "shown": False},
|
|
176
|
+
"agent_delegated": {"source": "program", "shown": False},
|
|
177
|
+
"agent_waited": {"source": "program", "shown": False},
|
|
178
|
+
"agent_sent": {"source": "program", "shown": False},
|
|
179
|
+
"agent_cancelled": {"source": "program", "shown": False},
|
|
180
|
+
"agent_report": {"source": "harness", "shown": False},
|
|
181
|
+
"span_start": {"source": "harness", "shown": False},
|
|
182
|
+
"span_end": {"source": "harness", "shown": False},
|
|
183
|
+
"model_text": {"source": "agent", "shown": False},
|
|
184
|
+
"mcp_submitted": {"source": "agent", "shown": False},
|
|
185
|
+
"mcp_dispatched": {"source": "agent", "shown": False},
|
|
186
|
+
"mcp_receipt": {"source": "agent", "shown": False},
|
|
187
|
+
"thought_queued": {"source": "agent", "shown": False},
|
|
188
|
+
"thought_started": {"source": "agent", "shown": False},
|
|
189
|
+
"thought_failed": {"source": "agent", "shown": False},
|
|
190
|
+
"thought_exited": {"source": "agent", "shown": False},
|
|
191
|
+
"thought_cancelled": {"source": "agent", "shown": False},
|
|
192
|
+
"escalation": {"source": "program", "shown": False},
|
|
193
|
+
"entry_start": {"source": "program", "shown": False},
|
|
194
|
+
"entry_claims": {"source": "program", "shown": False},
|
|
195
|
+
"entry_end": {"source": "program", "shown": False},
|
|
196
|
+
"skill_step": {"source": "program", "shown": False},
|
|
197
|
+
"world": {"source": "program", "shown": False},
|
|
198
|
+
}
|
|
199
|
+
|
|
200
|
+
|
|
201
|
+
def record_kind(record: dict) -> tuple[str, str | None]:
|
|
202
|
+
"""A trace record's kind and, for a family (library), its action."""
|
|
203
|
+
kind = record.get("event") or ""
|
|
204
|
+
return kind, record.get("action") if kind == "library" else None
|
|
205
|
+
|
|
206
|
+
|
|
207
|
+
def step_of(start: dict, end: dict, call: str | None = None) -> dict:
|
|
208
|
+
"""An execution's span_start and span_end (lane "skill") as a step of its skill, in the
|
|
209
|
+
shape of a skill_step record (core/spans.py step), so a reader takes both alike."""
|
|
210
|
+
event = end.get("event_kind") or ("error" if end.get("error") else COMPLETED)
|
|
211
|
+
return {"skill": start.get("name"), "call": call, "ok": event in NOMINAL, "event": event,
|
|
212
|
+
"seconds": float(end.get("time") or 0.0) - float(start.get("time") or 0.0),
|
|
213
|
+
"reason": end.get("reason") or end.get("error")}
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
"""Bounded programs, skills and policy execution."""
|
|
@@ -0,0 +1,96 @@
|
|
|
1
|
+
"""Another environment taken into a running conversation (use_environment).
|
|
2
|
+
|
|
3
|
+
A conversation starts in one environment, whose session is the kernel's and whose calls
|
|
4
|
+
are the program's plain names. When a step cannot be done there, the model asks for
|
|
5
|
+
another kind (skills/environments): the session host asks the user as their settings say,
|
|
6
|
+
takes a workspace of that kind and answers with what its session needs. The kernel then
|
|
7
|
+
builds that kind's session beside its own, each with its tick (runtime/tick.py) and its
|
|
8
|
+
entries, and gives programs its calls under the kind's name:
|
|
9
|
+
|
|
10
|
+
desktop.click(x, y) project.run_command("npm test") browser.read_page()
|
|
11
|
+
|
|
12
|
+
The attached session never connects to its worker itself. Every call goes over the
|
|
13
|
+
kernel's connection to the host (RelayedEnvironment), which holds the worker's key, checks
|
|
14
|
+
the call as it checks the conversation's own and keeps the user's approvals: a step with
|
|
15
|
+
the real mouse and keyboard stays under the physical-input policy after the desktop is
|
|
16
|
+
attached. So the kernel's network guard is as it was when the kernel started.
|
|
17
|
+
"""
|
|
18
|
+
|
|
19
|
+
from harness.runtime.events import EpisodeEnded
|
|
20
|
+
from harness.runtime.execution import control
|
|
21
|
+
|
|
22
|
+
# Names of an attached session that programs do not get: ending the task is the
|
|
23
|
+
# conversation's own finish(), and library entries are the conversation's.
|
|
24
|
+
WITHHELD = ("finish", "library", "ClosedLoop")
|
|
25
|
+
|
|
26
|
+
|
|
27
|
+
class RelayedEnvironment:
|
|
28
|
+
"""An attached environment's worker, reached through the session host."""
|
|
29
|
+
|
|
30
|
+
approval_key = None
|
|
31
|
+
program_controls = frozenset()
|
|
32
|
+
|
|
33
|
+
def __init__(self, name: str, timeout: float = 300.0, host_controls=()):
|
|
34
|
+
self.name = name
|
|
35
|
+
self.timeout = float(timeout)
|
|
36
|
+
self.host_controls = frozenset(host_controls)
|
|
37
|
+
|
|
38
|
+
def call(self, name: str, **arguments):
|
|
39
|
+
return control.host_call({"environment": self.name, "name": name, "arguments": arguments}, self.timeout)
|
|
40
|
+
|
|
41
|
+
def close(self) -> None:
|
|
42
|
+
pass
|
|
43
|
+
|
|
44
|
+
|
|
45
|
+
class Attached:
|
|
46
|
+
"""The calls of an attached environment as attributes: desktop.click, project.read_file.
|
|
47
|
+
A call made after it was released, or after its episode ended, says so."""
|
|
48
|
+
|
|
49
|
+
def __init__(self, name: str, session):
|
|
50
|
+
calls = {key: value for key, value in session.namespace.items()
|
|
51
|
+
if not key.startswith("_") and key not in WITHHELD}
|
|
52
|
+
object.__setattr__(self, "_name", name)
|
|
53
|
+
object.__setattr__(self, "_calls", calls)
|
|
54
|
+
object.__setattr__(self, "_live", True)
|
|
55
|
+
|
|
56
|
+
def __getattr__(self, key):
|
|
57
|
+
calls = object.__getattribute__(self, "_calls")
|
|
58
|
+
name = object.__getattribute__(self, "_name")
|
|
59
|
+
if key not in calls:
|
|
60
|
+
raise AttributeError(f"{name} has no call {key}; it has {', '.join(sorted(calls))}")
|
|
61
|
+
value = calls[key]
|
|
62
|
+
if not callable(value):
|
|
63
|
+
return value
|
|
64
|
+
|
|
65
|
+
def call(*args, **kwargs):
|
|
66
|
+
if not object.__getattribute__(self, "_live"):
|
|
67
|
+
raise RuntimeError(f"{name} was released; use_environment({name!r}, why) takes it again")
|
|
68
|
+
try:
|
|
69
|
+
return value(*args, **kwargs)
|
|
70
|
+
except EpisodeEnded:
|
|
71
|
+
# Its worker ended its episode (a step limit): this conversation goes on.
|
|
72
|
+
raise RuntimeError(f"{name} has ended; release_environment({name!r}) gives it back") from None
|
|
73
|
+
|
|
74
|
+
call.__name__ = key
|
|
75
|
+
call.__doc__ = getattr(value, "__doc__", None)
|
|
76
|
+
return call
|
|
77
|
+
|
|
78
|
+
def __setattr__(self, key, value):
|
|
79
|
+
raise AttributeError("an attached environment's calls cannot be rebound")
|
|
80
|
+
|
|
81
|
+
def __dir__(self):
|
|
82
|
+
return sorted(object.__getattribute__(self, "_calls"))
|
|
83
|
+
|
|
84
|
+
def __repr__(self):
|
|
85
|
+
name = object.__getattribute__(self, "_name")
|
|
86
|
+
return f"<environment {name}: {', '.join(dir(self))}>"
|
|
87
|
+
|
|
88
|
+
|
|
89
|
+
def build(name: str, given: dict):
|
|
90
|
+
"""The session of an attached environment from what the host answered: {"embodiment",
|
|
91
|
+
"knowledge", "episode_id", "skills", "timeout_seconds", "host_controls", and the keys a
|
|
92
|
+
session of that kind reads}. Its worker is reached through the host."""
|
|
93
|
+
from harness.runtime.execution.session import session_class
|
|
94
|
+
|
|
95
|
+
environment = RelayedEnvironment(name, given.get("timeout_seconds", 300.0), given.get("host_controls", ()))
|
|
96
|
+
return session_class(given["embodiment"], given.get("knowledge"))(dict(given), environment=environment)
|
|
@@ -0,0 +1,211 @@
|
|
|
1
|
+
"""What the host asks of a running program besides its cells.
|
|
2
|
+
|
|
3
|
+
The kernel process reads the host's requests on a thread of its own, so a request
|
|
4
|
+
reaches a program while it runs: a pause (the user paused the conversation) makes the
|
|
5
|
+
next step of a bounded execution end it with USER_PAUSED (runtime/execution/session.py). A
|
|
6
|
+
message for the model (the user's, or the parent agent's) interrupts the program's own
|
|
7
|
+
code, not the entries it started: its execution ends at the next step with USER_MESSAGE,
|
|
8
|
+
and a wait of it (an entry's wait(), wait_agents()) raises that event at once
|
|
9
|
+
(runtime/execution/entries.py). The flags are cleared when the next cell starts. An answer is the host's reply to a request a
|
|
10
|
+
program sent it as a note (the user's answer to a physical-input request,
|
|
11
|
+
runtime/safety/input.py): the program waits for it, or for a pause. When a sub-agent the
|
|
12
|
+
program started ends, the host says so, and its entry ends (runtime/execution/entries.py).
|
|
13
|
+
"""
|
|
14
|
+
|
|
15
|
+
import sys
|
|
16
|
+
import threading
|
|
17
|
+
import time
|
|
18
|
+
import uuid
|
|
19
|
+
|
|
20
|
+
|
|
21
|
+
_pause = threading.Event()
|
|
22
|
+
_message = threading.Event()
|
|
23
|
+
_answers = {}
|
|
24
|
+
_answered = threading.Condition()
|
|
25
|
+
_agent_listeners = []
|
|
26
|
+
_thoughts = set()
|
|
27
|
+
_physical = set()
|
|
28
|
+
_worker_calls = set()
|
|
29
|
+
_mcp_calls = set()
|
|
30
|
+
host_thoughts = False
|
|
31
|
+
|
|
32
|
+
|
|
33
|
+
def request_pause() -> None:
|
|
34
|
+
"""Ask the running program to stop at its next execution step."""
|
|
35
|
+
with _answered:
|
|
36
|
+
_pause.set()
|
|
37
|
+
_answered.notify_all()
|
|
38
|
+
|
|
39
|
+
|
|
40
|
+
def request_message() -> None:
|
|
41
|
+
"""Tell the running program a message for the model came."""
|
|
42
|
+
with _answered:
|
|
43
|
+
_message.set()
|
|
44
|
+
_answered.notify_all()
|
|
45
|
+
|
|
46
|
+
|
|
47
|
+
def clear() -> None:
|
|
48
|
+
"""Forget earlier requests (a new cell starts)."""
|
|
49
|
+
_pause.clear()
|
|
50
|
+
_message.clear()
|
|
51
|
+
|
|
52
|
+
|
|
53
|
+
def message_arrived() -> bool:
|
|
54
|
+
"""Whether a message for the model came while the program runs."""
|
|
55
|
+
return _message.is_set()
|
|
56
|
+
|
|
57
|
+
|
|
58
|
+
def pause_requested() -> bool:
|
|
59
|
+
"""Whether the host asked the running program to pause."""
|
|
60
|
+
return _pause.is_set()
|
|
61
|
+
|
|
62
|
+
|
|
63
|
+
def wake_waiters():
|
|
64
|
+
with _answered:
|
|
65
|
+
_answered.notify_all()
|
|
66
|
+
|
|
67
|
+
|
|
68
|
+
def answer(identifier, value) -> None:
|
|
69
|
+
"""The host's answer to the request `identifier`."""
|
|
70
|
+
with _answered:
|
|
71
|
+
identifier = str(identifier)
|
|
72
|
+
if isinstance(value, dict) and value.get("thought") and identifier not in _thoughts:
|
|
73
|
+
return
|
|
74
|
+
if isinstance(value, dict) and value.get("physical") and identifier not in _physical:
|
|
75
|
+
return
|
|
76
|
+
if isinstance(value, dict) and value.get("worker") and identifier not in _worker_calls:
|
|
77
|
+
return
|
|
78
|
+
if isinstance(value, dict) and value.get("mcp") and identifier not in _mcp_calls:
|
|
79
|
+
return
|
|
80
|
+
_answers[identifier] = value
|
|
81
|
+
_answered.notify_all()
|
|
82
|
+
|
|
83
|
+
|
|
84
|
+
def wait_answer(identifier, poll: float | None = None, stop=None, *, deadline: float | None = None,
|
|
85
|
+
interruptible=True):
|
|
86
|
+
"""Wait for the host's answer, an interruption or an absolute monotonic deadline."""
|
|
87
|
+
identifier = str(identifier)
|
|
88
|
+
with _answered:
|
|
89
|
+
while identifier not in _answers:
|
|
90
|
+
if poll == 0 or (interruptible and _pause.is_set()) or (stop is not None and stop()):
|
|
91
|
+
return None
|
|
92
|
+
remaining = None if deadline is None else deadline - time.monotonic()
|
|
93
|
+
if remaining is not None and remaining <= 0:
|
|
94
|
+
return None
|
|
95
|
+
_answered.wait(poll if remaining is None else remaining if poll is None else min(poll, remaining))
|
|
96
|
+
return _answers.pop(identifier)
|
|
97
|
+
|
|
98
|
+
|
|
99
|
+
def worker_call(send, arguments, timeout):
|
|
100
|
+
"""Relay a worker RPC over the kernel connection, with a finite receipt deadline."""
|
|
101
|
+
identifier = uuid.uuid4().hex
|
|
102
|
+
with _answered:
|
|
103
|
+
_worker_calls.add(identifier)
|
|
104
|
+
try:
|
|
105
|
+
deadline = time.monotonic() + timeout + 5
|
|
106
|
+
send({"op": "worker_request", "id": identifier, **arguments})
|
|
107
|
+
answer = wait_answer(identifier, deadline=deadline, interruptible=False)
|
|
108
|
+
finally:
|
|
109
|
+
with _answered:
|
|
110
|
+
_worker_calls.discard(identifier)
|
|
111
|
+
_answers.pop(identifier, None)
|
|
112
|
+
if answer is None:
|
|
113
|
+
raise RuntimeError("Worker receipt timed out; delivery may be unknown")
|
|
114
|
+
if "error" in answer:
|
|
115
|
+
raise RuntimeError(answer["error"])
|
|
116
|
+
return answer["value"]
|
|
117
|
+
|
|
118
|
+
|
|
119
|
+
def host_call(arguments, timeout):
|
|
120
|
+
"""A worker call of an attached environment (runtime/execution/attached.py), made by the
|
|
121
|
+
session host: the kernel's own connection carries it, as it carries the host's controls."""
|
|
122
|
+
from harness.core import spans
|
|
123
|
+
|
|
124
|
+
if not spans.connected():
|
|
125
|
+
raise RuntimeError("An attached environment needs the conversation that runs this program")
|
|
126
|
+
return worker_call(spans._emit, arguments, timeout)
|
|
127
|
+
|
|
128
|
+
|
|
129
|
+
def ask_host(note, *, stop=None, **detail):
|
|
130
|
+
"""Send the host a request as a note `note` and wait for its answer: the value, None
|
|
131
|
+
when the wait was interrupted (a pause, or `stop()`), ValueError with the host's reason
|
|
132
|
+
when it refused."""
|
|
133
|
+
from harness.core import spans
|
|
134
|
+
|
|
135
|
+
if not spans.connected():
|
|
136
|
+
raise RuntimeError(f"{note} needs the conversation that runs this program; none is connected")
|
|
137
|
+
identifier = uuid.uuid4().hex[:12]
|
|
138
|
+
spans.note(note, id=identifier, **detail)
|
|
139
|
+
answer = wait_answer(identifier, stop=stop)
|
|
140
|
+
if answer is None:
|
|
141
|
+
return None
|
|
142
|
+
if answer.get("error"):
|
|
143
|
+
raise ValueError(answer["error"])
|
|
144
|
+
return answer.get("value")
|
|
145
|
+
|
|
146
|
+
|
|
147
|
+
def think(role, text="", *, images=(), system=None, entry=None, frame=None, stop=None, **detail):
|
|
148
|
+
"""Ask the session's host on behalf of a kernel entry."""
|
|
149
|
+
from harness.core import spans
|
|
150
|
+
from harness.runtime.decision.model import ChatReply
|
|
151
|
+
from harness.runtime.execution.entries import Cancelled, current_entry
|
|
152
|
+
|
|
153
|
+
sys.audit("harness.model_request", role)
|
|
154
|
+
if not host_thoughts or not spans.connected():
|
|
155
|
+
return None
|
|
156
|
+
if stop is None and threading.current_thread() is threading.main_thread():
|
|
157
|
+
stop = message_arrived
|
|
158
|
+
owner = current_entry()
|
|
159
|
+
|
|
160
|
+
def owner_cancelled():
|
|
161
|
+
return owner is not None and any(entry._cancel.is_set() for entry in owner.lineage())
|
|
162
|
+
|
|
163
|
+
identifier = uuid.uuid4().hex
|
|
164
|
+
with _answered:
|
|
165
|
+
_thoughts.add(identifier)
|
|
166
|
+
answer = None
|
|
167
|
+
try:
|
|
168
|
+
spans.note("thought", id=identifier, role=role, text=str(text), images=list(images),
|
|
169
|
+
system=system, entry=entry or spans.current_entry(), frame=frame, **detail)
|
|
170
|
+
answer = wait_answer(identifier, stop=lambda: owner_cancelled() or (stop is not None and stop()))
|
|
171
|
+
finally:
|
|
172
|
+
with _answered:
|
|
173
|
+
_thoughts.discard(identifier)
|
|
174
|
+
_answers.pop(identifier, None)
|
|
175
|
+
if answer is None:
|
|
176
|
+
spans.note("thought_cancel", id=identifier)
|
|
177
|
+
if owner_cancelled():
|
|
178
|
+
raise Cancelled(f"entry {owner.id} was cancelled")
|
|
179
|
+
if answer is None:
|
|
180
|
+
raise stopped(role)
|
|
181
|
+
if answer.get("error"):
|
|
182
|
+
raise RuntimeError(answer["error"])
|
|
183
|
+
value = answer.get("value")
|
|
184
|
+
return None if value is None else ChatReply(text=value["text"], usage=value.get("usage") or {},
|
|
185
|
+
seconds=float(value.get("seconds") or 0))
|
|
186
|
+
|
|
187
|
+
|
|
188
|
+
def stopped(skill: str = "program"):
|
|
189
|
+
"""The typed escalation for an interrupted program wait."""
|
|
190
|
+
from harness.runtime.events import USER_MESSAGE, USER_PAUSED, Escalation, Event, ExecutionTrace
|
|
191
|
+
|
|
192
|
+
event = Event(USER_PAUSED if pause_requested() else USER_MESSAGE, "user")
|
|
193
|
+
return Escalation(ExecutionTrace(skill, {}, 0, 0, event))
|
|
194
|
+
|
|
195
|
+
|
|
196
|
+
def on_agent_ended(listener, remove: bool = False) -> None:
|
|
197
|
+
"""Call listener(agent, report) when the host says a sub-agent ended (`remove`: stop)."""
|
|
198
|
+
if remove:
|
|
199
|
+
if listener in _agent_listeners:
|
|
200
|
+
_agent_listeners.remove(listener)
|
|
201
|
+
else:
|
|
202
|
+
_agent_listeners.append(listener)
|
|
203
|
+
|
|
204
|
+
|
|
205
|
+
def agent_ended(agent, report, *, cancelled=False) -> None:
|
|
206
|
+
"""The host says the sub-agent `agent` ended, with its report."""
|
|
207
|
+
for listener in list(_agent_listeners):
|
|
208
|
+
try:
|
|
209
|
+
listener(agent, report, cancelled=cancelled)
|
|
210
|
+
except Exception:
|
|
211
|
+
pass
|