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.

Anchored iterative summarization scores highest (3.70 overall) for long sessions where file tracking matters, achieving 98.6% compression. Opaque compression reaches 99.3% but sacrifices interpretability, while regenerative summaries suit sessions with clear phase boundaries.

Use probe-based evaluation instead of metrics like ROUGE. Generate recall, artifact, continuation, and decision probes from the original history, then score model responses across six dimensions including accuracy and artifact trail using an LLM judge.

Trigger compression at 70-80% context utilization as a default, or use a sliding window keeping the last N turns plus a summary. Task-boundary triggers work well when sessions have clear phase transitions like research, implementation, and testing.

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 in the agent scaffolding.

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