Performance Regression Cluster: Typing lag, modal freezes, and PTY corruption

[wowlegend]

Summary

This is a cluster of 4 related performance/reliability issues that all stem from the same rendering pipeline. I've opened separate issues for each, but they should be tracked together since fixing the root cause (synchronous rendering overhead + missing PTY allocation) would resolve most of them.

---

The Issues

Issue	Severity	Root Cause	User Impact
#2032 — Extreme typing latency in inline modal inputs	Critical	render_to_ansi() creates a new rich.Console on every keystroke + 10 Hz forced redraws + completion engine fires on hidden input buffers	Inline feedback/"Other" input is completely unusable (1–3s per keystroke)
#2038 — Bottom toolbar git subprocess lag	High	_render_bottom_toolbar() polls git subprocesses on every prompt_toolkit redraw	Per-keystroke lag even during normal prompt typing (no active turn)
#2037 — Shell tool PTY allocation / terminal corruption	High	stdin.close() + no PTY allocation causes race between subprocess and prompt_toolkit over /dev/tty	Interactive commands (ssh-add, sudo, npm init) corrupt terminal input or hang
#2010 — Shift+Enter for newline	UX	Missing keybinding	Muscle memory mismatch with every modern chat UI

---

Common Thread: Rendering Pipeline

Issues #2032 and #2038 both trace back to synchronous, unthrottled rendering in the prompt_toolkit UI layer:

1. No memoization — _render_bottom_toolbar() and modal delegates recompute everything on every redraw
2. Expensive object creation — render_to_ansi() instantiates a new rich.Console per call
3. No state-aware throttling — Background refresh runs at 10 Hz even when UI is static
4. Completion engine not context-aware — start_completion() runs even when the buffer is a hidden text field

Issue #2037 is architectural: the Shell tool needs PTY allocation or terminal yielding to handle interactive subprocesses correctly (Claude Code, Gemini CLI, and Codex CLI all do this).

---

What I've Already Done

I've locally patched #2032 with three changes:

• Skip completions when modal hides the buffer
• Use idle refresh interval during modal hidden-input states
• Cache Rich→ANSI output in approval/question delegates

After patching, inline input feels instant in Ghostty. I've shared the full root cause analysis and suggested fixes in each issue above.

---

Ask

Could a maintainer acknowledge this cluster and indicate which fixes (if any) are on the roadmap? Happy to open PRs for the rendering patches if that would help.