cc-dumb-zone-statusline

Claude Code dumb zone statusline.sh

Know when your Claude Code session is about to get dumb — before you waste an hour wondering why it keeps making the same mistake.

A three-line statusline for Claude Code that color-codes your context usage against model-aware "dumb zone" thresholds — so you can /clear *before* quality degrades, not after.

One bash script. No Node, no npm, no config. Cross-platform (macOS, Linux, Windows), one-liner install, automatic settings.json patch with backup, validated end-to-end before it tells you "Done".

Why this exists

Every frontier LLM has a context window it *advertises* (200k, 1M, …) and a context window it *actually performs well in*. They are not the same number.

As your conversation fills up, attention dilutes, retrieval over the prompt degrades, and the model starts:

- forgetting instructions you gave 30 turns ago,
- contradicting code it just wrote,
- "fixing" things that weren't broken,
- inventing APIs that don't exist.

The community calls this the dumb zone. The Dictionary of AI Coding anchors the threshold around ~100k tokens for frontier models — but the exact line moves with the model:

- Sonnet drifts earlier.
- Opus holds longer.
- Opus with the 1M-context beta holds longer still, but eventually degrades too.
- Haiku is the most fragile.

By the time Claude Code's built-in % indicator shows 50%, a Sonnet session may already be drifting. You want to /clear (or compact, or hand off) before you hit the dumb zone, not after. That's the entire job of this statusline.

What you see

Line 1 — the important one
- Zone label (🧐 / 🥱 / 🤪) with color
- Tokens used / window size, plus % of window
- A 20-segment progress bar:
  - █ solid = tokens used (colored by zone)
  - ▓ dim = safe headroom (same color as current zone)
 …