File size: 3,688 Bytes
6aa6c12 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 | """Runtime protocols implemented by benchmark and policy adapters."""
from __future__ import annotations
from pathlib import Path
from typing import Mapping, Protocol, Sequence
from benchkit.contracts import (
ActionRejectionFeedback,
ActionSpec,
BenchmarkDescriptor,
EpisodeContext,
EpisodeSpec,
ExecutionFeedback,
JsonValue,
Observation,
ObservationSpec,
PolicyDecision,
PolicyDescriptor,
Transition,
)
class EpisodeSession(Protocol):
"""One native benchmark episode opened by an adapter."""
def reset(self) -> Observation: ...
def step(self, action: tuple[float, ...]) -> Transition: ...
def metrics(self) -> Mapping[str, JsonValue]: ...
def artifacts(self) -> Mapping[str, str]: ...
def close(self) -> None: ...
class ArtifactBoundEpisodeSession(EpisodeSession, Protocol):
"""Optional session capability for runner-owned video/observation artifacts.
The runner calls this once, before reset, with a newly created directory
inside the episode's run directory. Sessions without this capability keep
the original EpisodeSession interface.
"""
def bind_artifact_dir(self, path: Path) -> None: ...
class BenchmarkAdapter(Protocol):
"""Factory and protocol metadata for one pinned benchmark runtime."""
@property
def descriptor(self) -> BenchmarkDescriptor: ...
@property
def observation_spec(self) -> ObservationSpec: ...
@property
def action_spec(self) -> ActionSpec: ...
def episodes(self, seed: int, limit: int | None = None) -> Sequence[EpisodeSpec]: ...
def open_episode(self, episode: EpisodeSpec) -> EpisodeSession: ...
class Policy(Protocol):
"""A model constrained to one observation-to-decision call at a time."""
@property
def descriptor(self) -> PolicyDescriptor: ...
def compatibility_errors(
self, observation_spec: ObservationSpec, action_spec: ActionSpec
) -> Sequence[str]: ...
def start_episode(self, context: EpisodeContext) -> None: ...
def act(self, observation: Observation) -> PolicyDecision: ...
def end_episode(self) -> None: ...
class RuntimeCheckablePolicy(Policy, Protocol):
"""Optional model-owned local readiness, before any native attempt.
No model request, task context, simulator, action generation or usage event
is allowed. Check local dependencies/credential presence only; release any
temporary clients. Success returns None; failures use PolicyPreflightError.
A profile opts in through its captured runtime_preflight declaration.
"""
def preflight_runtime(self) -> None: ...
class ExecutionAwarePolicy(Policy, Protocol):
"""Optional public receipt, sent after a control chunk is executed.
The runner neither knows nor enforces a policy's persistence/stop rule.
Policies can use actual execution receipts for their own state and limits.
A separately declared robot-control sensor channel may supply ordered
control acknowledgements. No private simulator metadata or scoring crosses
this capability. The hook records feedback without calling a model or SIM.
"""
def observe_execution(self, feedback: ExecutionFeedback) -> None: ...
class ActionRejectionAwarePolicy(Policy, Protocol):
"""Required opt-in for feedback-mode invalid actions, without another call.
For a rejected call this replaces observe_execution, including when a
prefix already executed. The next Policy.act receives the unchanged last
acknowledged observation. The hook must not request a model or simulator.
"""
def observe_rejection(self, feedback: ActionRejectionFeedback) -> None: ...
|