rune-context-engine

Track tool call counts as a proxy for context usage and classify health into GREEN, YELLOW, ORANGE, or RED levels. At RED, save decisions and progress to project files before triggering compaction so work resumes exactly where it stopped.

Multiply total character count by 0.25 to estimate tokens without a tokenizer dependency. Compare the estimate against per-caller warn and block thresholds to decide whether to proceed, warn, or abort the dispatch.

Claude Code does not expose actual token usage to skills, so tool call count serves as a directional proxy. Sessions with many large file reads consume context faster, so thresholds adjust downward by 0.8x when files average over 500 lines.

Compaction during implementation is safe only at clean task boundaries where a file is fully written and tested. State must be flushed to session storage first, and tool_use/tool_result pairs must never be split across the compaction boundary.

The 0.25 chars-to-tokens ratio is calibrated for English code and markdown, underestimating Japanese or Chinese content by 30-50% and overestimating repetitive code by 20-30%. Both error directions remain safe because dispatch-time hard caps catch underestimates.