nemo_gym.agents.agent_outcome

Agent-reported causes, independent of environment grading and retry policy.

Module Contents

Classes

Name	Description
`AgentOutcome`	Additional evidence distinguishing causes of a failed agent response.
`AgentOutcomePolicyConfig`	Existing Environment Server defaults for classified failed agent responses.

Functions

Name	Description
`agent_outcome_policy`	Apply configured grading/retry rules to a classified failed response.
`confirmed_policy_response_timeout`	Whether the selected model operation timed out after upload on every observed attempt.
`policy_timeout_failure`	The agent’s policy-timeout failure for a failed model reply, or `None`.

Data

AGENT_OUTCOME_METADATA_KEY

POLICY_TIMEOUT_OPERATION

API

class nemo_gym.agents.agent_outcome.AgentOutcome()

Bases: BaseModel

Additional evidence distinguishing causes of a failed agent response.

Ordinary completion and limits already use response.status. HTTP status here belongs to the upstream model request, not the agent endpoint’s HTTP 200.

failure_kind

str

model_config

= ConfigDict(extra='forbid', frozen=True)

status_code

int | None = Field(default=None, ge=400, le=599, strict=True)

nemo_gym.agents.agent_outcome.AgentOutcome.require_registered_kind(
    value: str
) -> str

classmethod

Accept the shared vocabulary plus an explicitly undetermined cause.

nemo_gym.agents.agent_outcome.AgentOutcome.require_request_status() -> nemo_gym.agents.agent_outcome.AgentOutcome

class nemo_gym.agents.agent_outcome.AgentOutcomePolicyConfig()

Bases: OutcomePolicyConfig

Existing Environment Server defaults for classified failed agent responses.

by_failure_kind

dict[str, OutcomeAction]

by_status_code

dict[Annotated[int, Field(ge=400, le=599)], OutcomeAction]

classification

FailureClassificationConfig = Field(default_factory=FailureClassificationConfig)

default_action

OutcomeAction = OutcomeAction.FAIL

supported_actions

frozenset[OutcomeAction]

nemo_gym.agents.agent_outcome.agent_outcome_policy(
    response: nemo_gym.core.responses_api_types.NeMoGymResponse,
    config: nemo_gym.agents.agent_outcome.AgentOutcomePolicyConfig | None = None,
    classification: nemo_gym.core.failure_kinds.FailureClassificationConfig | None = None
) -> nemo_gym.core.outcome_policy.OutcomeAction

Apply configured grading/retry rules to a classified failed response.

Unclassified legacy responses retain their existing grading behavior; malformed metadata is an ungraded, nonretryable failure regardless of configuration.

nemo_gym.agents.agent_outcome.confirmed_policy_response_timeout(
    report: nemo_gym.core.failure_records.FailureReport | None
) -> bool

Whether the selected model operation timed out after upload on every observed attempt.

Unidentified, mixed, omitted or connecting attempts do not establish a final response timeout.

nemo_gym.agents.agent_outcome.policy_timeout_failure(
    error: aiohttp.ClientResponseError
) -> nemo_gym.core.typed_exceptions.NgFailureEvent | None

The agent’s policy-timeout failure for a failed model reply, or None.

The reply’s report must show that the final model call timed out after its upload on every attempt (confirmed_policy_response_timeout). The failure is caused by error, so its reply keeps the model’s attempts; raise it from error.

nemo_gym.agents.agent_outcome.AGENT_OUTCOME_METADATA_KEY = 'ng_agent_outcome'

nemo_gym.agents.agent_outcome.POLICY_TIMEOUT_OPERATION = 'agent.policy_request'