context-window

LLM agent sessions fail due to context-window overrun when token budgets are unmanaged. This Skill prevents silent failures by enforcing budget zoning across System, Skill injection, and Working data, along with an 80% compaction rule.

Session recovery after compaction requires pre-compact checkpointing of decision trails and durable artifacts, followed by post-compact selective recovery using continuation signals to rebuild the working state without losing critical progress.

Token-cost reduction is achieved through deterministic CLI shaping, targeted file reads using offset and limit, search-before-read patterns, progressive skill disclosure, and count-mode exploration to minimize heavy tool-result injection.

Context window compaction preserves decision trails by checkpointing critical session artifacts before compaction and applying post-compact selective recovery to ensure multi-step audit continuity across the 80% threshold.

Context window budgeting applies to long-running agent sessions across various context sizes including 1M, 200K, and 128K tokens, dynamically allocating budgets across System, injected skill content, and working data zones.

Compaction should be triggered when the context health state reaches the 80% compaction rule threshold, indicated by tool results or file reads truncating mid-task, transitioning the session from the ok state to the compact state.