Skip to main content

context.window_config

Context window management configuration.

Determines how the context window is managed for long-running durable agents — auto-summarization, eviction, and history archive.

ContextWindowConfig Objects​

@dataclass
class ContextWindowConfig()

Configuration for context window management.

Spec example::

context_management:
  enabled: true
  max_tokens: 200000
  compress_threshold: 0.9
  summarization_model: "openai:gpt-4.1-mini"
  eviction_token_limit: 20000

Attributes​

enabled : bool Whether automatic context management is active. max_tokens : int Maximum context window size in tokens (auto-detected from model if 0). compress_threshold : float Fraction of max_tokens at which summarization triggers (default 0.9). summarization_model : str | None Model to use for summarization (cheaper than the agent model). eviction_token_limit : int | None Tool outputs larger than this are evicted to files. history_archive : bool Whether to persist the full conversation history for search.

max_tokens​

0 = auto-detect from model

from_spec​

@classmethod
def from_spec(cls,
              spec_context: dict[str, Any] | None) -> "ContextWindowConfig"

Build from AgentSpec context_management dict.

from_env​

@classmethod
def from_env(cls) -> "ContextWindowConfig"

Build from environment variables.