When you estimate the context window is approximately 75% full, proactively say:
"Context is around 75% full. Run
/compactto compress before continuing."
Do not wait until the context is nearly exhausted — compact early to preserve quality.
When compacting (via /compact or auto-compaction), follow these rules:
Preserve the most recent 3 conversation turns exactly as-is. Do not summarize or paraphrase them.
Regardless of what else is summarized, the compacted context MUST include:
-
Active rubrics — the list of rubric tags (
#audience,#thesis, etc.) being evaluated in the current session, and which ones have already been evaluated vs. are still pending. -
Assignment summary — a 3–5 sentence summary of the current assignment's argument, structure, and discipline. Do not discard which assignment file or Google Doc is being reviewed.
-
Professor feedback highlights — for each rubric being evaluated, any specific issues the professor flagged in past assignments. Keep the exact wording of each flag, not a paraphrase.
-
Evaluation state — which dimensions (a/b/c) have been completed per rubric, and any partial findings already written.
- Tool call details (Glob results, file path lookups, scraper output logs)
- Repeated rubric definition text already loaded from reference files
- Early exploration steps (finding the PDF, confirming file names)
- Grammar/spelling subagent raw output (keep only the flagged items, not the full output)