context-compression

Use anchored iterative summarization: maintain a structured summary with explicit sections for session intent, files modified, decisions, and next steps. On each compression trigger, summarize only the newly truncated span and merge it into existing sections rather than regenerating the whole summary.

Use probe-based evaluation instead of lexical metrics like ROUGE. Generate recall, artifact, continuation, and decision probes from the original history, collect model responses using the compressed context, and score them across six dimensions including accuracy and artifact trail.

Anchored iterative summarization merges new spans into persistent structured sections and scores highest on quality (3.70). Regenerative summaries regenerate fully each cycle and suit sessions with clear phase boundaries. Opaque compression achieves the highest ratio (99.3%) but sacrifices interpretability and verifiability.

Default to a sliding window trigger at 70-80% context utilization for coding agents, keeping recent turns plus a structured summary. Use task-boundary triggers when sessions have clear phase transitions such as research, implementation, and testing.

Artifact trail is the weakest dimension across all compression methods, scoring 2.19-2.45 in benchmarks. General summarization cannot reliably track file operations, 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 or types breaks tool invocation entirely. Also protect early turns containing task constraints and preserve code blocks, file paths, and error codes verbatim.