context-compression

Use anchored iterative summarization: maintain a persistent structured summary with sections for intent, files modified, decisions, and next steps, then merge only newly truncated content on each compression trigger. This prevents the cumulative detail loss caused by regenerating summaries from scratch.

Anchored iterative summarization scores highest for coding agents because it preserves file paths, function names, and decision rationale. Sliding window triggers at 70-80% context utilization provide the best balance of predictability and quality.

Use probe-based evaluation instead of metrics like ROUGE. After compression, ask recall, artifact, continuation, and decision questions, then score responses across six dimensions including accuracy, artifact trail, and continuity.

General summarization cannot reliably maintain artifact trails, which benchmark as the weakest dimension across all methods. Implement a separate artifact index or explicit file-state tracking outside the summarizer to preserve file paths and changes.

Avoid aggressive or opaque compression when debugging, file tracking, or verification matters, since it sacrifices interpretability and forces costly re-fetching. Also never compress tool definitions, code blocks, or early turns containing task constraints.