What problem does it solve? Auto-compaction in long AI coding sessions triggers at arbitrary points, often mid-task, losing important context like file paths, variable names, and partial reasoning. This Skill helps you compact at logical phase boundaries instead, preserving the context that matters. ## Core Features & Use Cases - Context-size monitoring: Reads token usage from the session transcript and suggests /compact at window-scaled thresholds (160k on a 200k window, 250k on a 1M window), re-reminding every 60k tokens of growth. - Tool-call counting: Tracks tool invocations per session and suggests compaction at a configurable threshold (default 50 calls, then every 25). - Compaction decision guidance: Provides a phase-transition table (research → planning → implementation → testing) and a list of what survives compaction (CLAUDE.md, TodoWrite, memory files, git state) versus what is lost. - Use Case: During a long multi-phase feature build, the PreToolUse hook fires before an Edit and reminds you to run /compact after finalizing your plan, so implementation starts with a fresh context window. ## Quick Start Add the suggest-compact.js hook to your Claude settings for Edit and Write tool calls, then follow its suggestions to run /compact at logical phase boundaries.
