manage-contextlisted

# Manage Context Keep work stable when context grows long, the codebase is unfamiliar, attention drifts, or the model is stuck. Seven modes: context budget, session hygiene, zoom-out, deep triage, escalate, onboarding, post-compact re-anchor. ## Iron Rule <EXTREMELY-IMPORTANT> 1. NEVER continue editing when the same bug surfaces from 3+ angles — that is a context-loss signal. Zoom out first. 2. NEVER assume a fact from earlier in the session still holds for files edited since. Re-read before acting. 3. ESCALATE to a stronger model (or a fresh session) at the SECOND failed attempt at the same goal, not the tenth. In a debug flow this is debug-issue §9: one cross-model consult, one advisor-informed attempt, then the user. 4. ONBOARDING in an unfamiliar repo: detect stack and conventions from real files before writing a line. </EXTREMELY-IMPORTANT> ## When to use - An observable context signal: the context bar turns yellow/red, the harness warns about compaction, or recall degrades (re-reading files already read, forgetting stated constraints). ~70% of a meter is the line — act on the observable signal, not the estimate. - You just resumed from a compaction summary (auto or `/compact`). - A usage / quota limit is near — the session dies regardless of context; hand off, don't trim. - The same bug reappears at a different surface · you forgot a stated constraint · unfamiliar codebase · a multi-file refactor fans out beyond the plan · the model is stuck on architecture or root