context-compression

Use anchored iterative summarization with explicit sections for files modified, files read, and decisions made. Summarize only newly truncated spans and merge them into the existing summary rather than regenerating, which prevents cumulative detail loss across compression cycles.

Use probe-based evaluation: after compression, ask questions testing recall, artifact tracking, continuation, and decision retention. Score responses across six dimensions including accuracy and artifact trail, since metrics like ROUGE miss functional information loss.

Anchored iterative summarization maintains interpretable structured summaries scoring 3.70 quality at 98.6% compression, while opaque compression achieves 99.3% ratios but sacrifices interpretability and scores 3.35. Use opaque only for short sessions with low re-fetching costs.

Trigger at 70-80% context utilization for fixed thresholds, or use a sliding window keeping the last N turns plus a summary for predictable context size. Task-boundary triggers work best when sessions have clear phase transitions like research then implementation.

Artifact trail is the weakest dimension across all compression methods, scoring only 2.2-2.5 out of 5. General summarization cannot reliably track files, so implement a separate artifact index or explicit file-state tracking outside the summarizer.

Never compress tool definitions, function call schemas, or API specs, since summarizing parameter names destroys the agent's ability to invoke tools. Also protect early session turns containing task constraints and preserve code blocks and identifiers verbatim.