Context Compression

Context Compression is a critical skill for managing the massive conversation histories that accumulate in long-running agent sessions. When sessions generate millions of tokens, compression becomes mandatory to maintain performance and control costs. This skill teaches production-ready approaches to context management without losing essential information.

The skill reframes the optimization goal from minimizing tokens per request to minimizing tokens per task—the total tokens consumed to complete a task including re-fetching costs when compression loses critical information. This distinction is crucial for making the right compression tradeoffs.

Three production-ready compression approaches are covered in depth. Anchored Iterative Summarization maintains structured summaries with dedicated sections for session intent, file modifications, decisions, and next steps, with new content merged incrementally rather than regenerated. This structured approach forces preservation through explicit checklists. Opaque Compression achieves maximum compression ratios of 99%+ but sacrifices interpretability and verification capabilities. Regenerative Full Summary produces readable, detailed outputs but risks information drift across repeated compression cycles.

A critical finding highlighted by this skill: file tracking remains the weakest dimension across all compression methods, scoring only 2.2-2.5 out of 5.0. Agents struggle to maintain complete records of which files were created, modified, or merely read—information essential for coding tasks.

The skill recommends triggering compression at 70-80% context utilization and provides a three-phase workflow: Research (compress exploration into structured analysis), Planning (convert findings into specifications), and Implementation (execute against specifications). Essential for anyone building long-running agent systems or working with large codebases exceeding 5M tokens.