Rollout

Chain Generation

Base class for chain-based generation:

ChainRollout()

Basic class for chain-based rollout. It starts multiple chains and runs them asynchronously.

Functions

get_trajectories() -> List[Trajectory]

Build :class:Trajectory objects for every active chain.

Maps chain.info keys into the typed Trajectory fields:

• reward (a RewardResult produced by calculate_reward) → Trajectory.reward (float) + Trajectory.metrics (extras)
• finish_reason, rollout_time_sec, chain_id, group_id, chain_idx, group_idx → first-class fields
• everything else (dataset-row passthrough fields like answer, task_id, etc.) → Trajectory.metadata

run_async(messages: List[Dict], max_turns: int, num_chains: int, generation_config: Optional[Dict[str, Any]] = None, enable_streaming: bool = False, max_concurrent_chains: Optional[int] = None, context_config: Optional[ContextConfig] = None) async

Run the chain-based rollout with optional streaming support.

Parameters:

• max_steps –

  Maximum number of steps for each chain.

• start_messages –

  List of messages to start the chains.

• num_chains (int) –

  Number of chains to run for each message.

• max_concurrent_chains (Optional[int], default: None ) –

  Maximum number of chains to execute concurrently across the whole rollout. If None, all chains are scheduled at once.

• generation_config (Optional[Dict[str, Any]], default: None ) –

  Generation configuration dictionary.

• enable_streaming (bool, default: False ) –

  Whether to enable streaming mode.

• streaming_callback –

  Optional callback for streaming events.

• context_config (Optional[ContextConfig], default: None ) –

  Optional :class:~agentfly.core.context_config.ContextConfig for tool resources (backend).

apply_context_folding(chain: Chain, messages: Messages, tool_name: str, observation: Any, tool_call_id: str, tool_result_name: str, new_content: List[Dict[str, Any]], context: Context) -> Tuple[Messages, bool]

Apply context folding based on the tool that was just executed.

Parameters:

• chain (Chain) –

  The current Chain object (used to record full histories).

• messages (Messages) –

  The current full Messages object before applying the tool.

• tool_name (str) –

  Name of the tool that was executed.

• observation (Any) –

  Tool observation returned by the tool.

• tool_call_id (str) –

  ID of the tool call.

• tool_result_name (str) –

  Name of the tool result (tool implementation name).

• new_content (List[Dict[str, Any]]) –

  Content payload for the tool message.

Returns:

• Messages –

  (updated Messages, did_fold) where did_fold is True when history was

• bool –

  replaced by a summarize fold (caller should refresh cumulative token counts).