context-compression_skill

Overview

This skill helps design and implement context compression for long-running agent sessions and large codebase interactions. It focuses on minimizing tokens-per-task by preserving critical information (file artifacts, decisions, next steps) using structured summarization and incremental merging. Use it to avoid agent forgetfulness, reduce re-fetching costs, and maintain actionable summaries across many compressions.

How this skill works

The skill implements multiple compression strategies—anchored iterative summarization, opaque compression, and regenerative full summaries—and recommends triggers and evaluation methods. It produces structured summaries with explicit sections (session intent, files modified, decisions, current state, next steps) and supports incremental merging so only newly-truncated content is summarized and appended. Probe-based evaluation assesses functional retention across accuracy, artifact trail, continuity, and instruction-following dimensions.

When to use it

• Agent sessions exceed model context windows or approach limits
• Managing multi-million-token codebase exploration or long debugging runs
• You observe agents losing track of files, decisions, or recent changes
• Designing or validating conversation summarization strategies
• Optimizing tokens-per-task to reduce re-fetching costs

Best practices

• Optimize for tokens-per-task, not tokens-per-request
• Define explicit summary sections for intent, files, decisions, state, and next steps
• Trigger compression at 70–80% context utilization or use a sliding-window plus summary approach
• Use anchored iterative summarization for file-tracking and verification needs
• Evaluate compression quality with probe-based tests (recall, artifact, continuation, decision)
• Track artifact provenance separately when file integrity is critical

Example use cases

• Long debugging session where tracking modified files and test status prevents rework
• Refactoring a 5M+ token codebase into a 2,000-word implementation specification
• Production agents that must keep stable next-step plans across many interactions
• Evaluation harness that measures tokens-per-task impact and re-fetch frequency
• Migration task using a reference PR as a seed to preserve invariants and constraints