nemo_rl.models.generation.capture_context

Framework-owned serving-prefix decisions, independent of capture storage.

Module Contents

Classes

CaptureInputDecision

Commit a continuation link or start a new context segment.

Functions

decide_capture_input

Verify the converted boundary and renderer for a Gym-selected candidate.

Data

LOGGER

API

nemo_rl.models.generation.capture_context.LOGGER

‘getLogger(…)’

class nemo_rl.models.generation.capture_context.CaptureInputDecision

Commit a continuation link or start a new context segment.

storage: nemo_gym.token_id_capture.staging.records.CaptureAdmission

None

serving: CaptureAdmission | None

None

verify_retained_media: bool

None

nemo_rl.models.generation.capture_context.decide_capture_input(

admission: nemo_gym.token_id_capture.staging.records.CaptureAdmission,

*,

messages: list[dict[str, Any]],

allow_empty_tool_content: bool = False,

) → nemo_rl.models.generation.capture_context.CaptureInputDecision

Verify the converted boundary and renderer for a Gym-selected candidate.

Gym proves source continuity and effective render-option equality. RL installs exact captured tokens only when conversion preserves the cut. Managed roots retain media validation; legacy admissions are unchanged. allow_empty_tool_content is a worker-proven rendering fact, not a harness option. It permits only null/empty Chat assistant tool content.