ContextCompressionGuard Reference¶

The ContextCompressionGuard is responsible for detecting and auditing silent context compression/summarization events performed by LLM providers.

Initialization¶

The guard is typically initialized automatically by TheseusWrapper based on the WrapperConfig.

from ct_toolkit.core.compression_guard import ContextCompressionGuard

# Manual initialization
guard = ContextCompressionGuard(
    threshold=0.85,
    embedding_client=my_client
)

Methods¶

`analyze_summary_drift(original_messages, new_summary) → dict`¶

Compares the original message history with a new summary/compacted version to detect semantic drift.

Parameters: - original_messages: List of message dictionaries representing the uncompressed history. - new_summary: The new summary string or list of compressed messages.

Returns: - similarity: Cosine similarity score (0.0–1.0). - drift_detected: True if similarity < threshold. - critical_drift: True if similarity is significantly below threshold (risk of identity loss).

`record_audit(audit_data) → None`¶

Records the results of a compression audit in the provenance log.

Integration with Middlewares¶

LangChain¶

Exposed via the compression_guard property on TheseusChatModel.

CrewAI¶

Settings are propagated to all agents; the guard handles internal audits for each agent's LLM calls.