context-compression

Compress context by producing token-efficient, structured summaries that retain critical files, decisions, and next steps. This ensures long-running agent sessions avoid token overflow and prevent forgetting critical details.

Anchored iterative summarization maintains persistent, sectioned summaries that merge new content without losing history. It enforces deterministic merges and explicit file-state tracking to reduce drift during code review tasks.

Evaluate compression trade-offs using probe-based evaluation and artifact tracking. This quantifies information fidelity against token savings to determine if opaque compression or regenerative full summaries suit your task.

Opaque compression produces highly compact representations for maximum token savings during multi-step debugging. It involves interpretability trade-offs, meaning the compressed output is less human-readable but preserves critical session data.

Use regenerative full summaries when you need to support review and auditing. This approach regenerates detailed summaries on each compression cycle, contrasting with anchored merging which maintains persistent sectioned summaries incrementally.