As conversations lengthen, response quality degrades due to context issues, but users receive no signal. The user proposes adding a context health indicator that proactively suggests starting a new conversation when context degradation is detected.
## Problem As conversations grow longer, response quality degrades due to attention dilution, lost-in-the-middle effects, and context compression losses. Users have no signal that this is happening — they only notice when Claude starts giving worse answers. Currently Claude Code compresses prior messages when approaching context limits, but this is invisible to the user. There's no proactive nudge to start a fresh conversation. ## Proposal Add a context health indicator that suggests starting a new conversation when quality is likely degrading. Possible triggers: 1. **Token usage threshold** — e.g., "Context is 70%+ full, consider starting a new conversation to maintain quality" 2. **Compression event** — when prior messages get summarized, notify the user: "Earlier messages have been compressed — some details may be lost. Consider `/clear` or a new session for complex tasks." 3. **Turn count heuristic** — after N substantive back-and-forth turns, suggest a break ## Why this matt