context-management

Context management optimizes session history by trimming non-essential information while preserving critical context. Use the /compact command before sessions end to prune history, apply must_keep and can_remove rules to prioritize what stays, and monitor context usage at 70% (warning), 80% (critical), and 90% (danger) thresholds to stay within limits.

Structured session summaries store essential information in .claude/session-history with pre- and post-compact checks. The Skill tracks prior and current phases, externalizes context via context-log records, and maintains a Single Source of Truth through current-implementation.md updates to preserve reasoning while reducing token overhead.

Yes. The Skill integrates on_session_start and on_session_end hooks into your workflow, centralizes logging via context-log records, and stores standardized session summaries for traceability. This enables monitoring context drift across interactive agents and long-running sessions with full auditability.

Define must_keep rules for critical context—reasoning, decisions, system state—and can_remove rules for noise or redundant exchanges. The Skill applies these prioritized retention lists during /compact to balance information density with token efficiency, guided by threshold-based warnings.

Context drift occurs when history grows unchecked and important prior reasoning gets deprioritized or dropped. Proactive context trimming, phase tracking, and structured summaries prevent drift by maintaining essential information while compacting redundant content, ensuring continuity across session boundaries.