Claude Code’s latest iteration still struggles with sluggish input, despite a patch that swapped non‑cached file reads for cached ones. The author, after profiling the tool, discovered that a single get function rebuilds the entire conversation buffer every frame, consuming roughly a quarter of CPU time.

Profiling revealed that Qt1 and get together consume 45% of CPU, each performing O(n) operations over the full scrollback. The renderer allocates a fresh 2‑D array per frame and replays all operations, while the LLM attends to the entire context window for every token, amplifying latency.

Attempts to patch the renderer with simple I/O fixes—caching reads, memoizing writes—only addressed surface symptoms. The root issue lies in the architectural choice of rebuilding the entire buffer each frame. A rewrite that tracks damage and renders viewport‑only content is required to restore responsiveness.

Claude Code 2.1.14 reduced the combined Qt1/get load from ~48% to ~22%, yet flickering reappears as the cell‑based diffing toggle flips. Developers now face a trade‑off between visual stability and performance, underscoring the need for smarter state management in terminal‑based LLM tools.