testsAndMisc/linux_configuration/zsh/claude-tty-fix.zsh
Krzysztof kuhy Rudnicki b57a130e9b claude-tty-fix.zsh: precmd mouse-heal, debug logging, on-demand MCP launchers
Diagnosed the "claude keeps crashing + mouse-command garble" as three stacked
bugs (remoteControlAtStartup end_session collision, SIGTTIN suspend on early
input, and an upstream REPL-exits-on-startup-input). This carries the shell-side
mitigations and the tooling that caught them:

- _claude_mouse_heal precmd hook: disables mouse tracking before every prompt,
  so garble left by a *suspended* claude clears on the next Enter (the on-exit
  reset in claude() can't fire while the function is blocked in a suspend).
- Always-on --debug-file logging in claude() (~/.claude/logs, keeps last 20;
  opt out with CLAUDE_DEBUG_LOG_OFF=1) -- this captured all three root causes.
- On-demand launchers claude-rag/-travel/-sports/-media/-extras load the niche
  MCP servers via --mcp-config, keeping them out of the always-on global fleet.
- claude-debug raw-binary probe (note: zsh printf has no %(...)T directive).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LZ15AWWYN8xGKdwXKn7foM
2026-07-18 15:55:28 +02:00

130 lines
7.1 KiB
Bash

# ============================================================================
# claude-tty-fix.zsh — make Claude Code survive early typing in a fresh terminal
# ----------------------------------------------------------------------------
# Symptom this fixes:
# Launch `claude` in a fresh terminal and type/interact during its ~5s startup
# (instead of waiting) and the process stops with
# [1] + <pid> suspended (tty input)
# while the terminal spams sequences like `35;44;41M` (SGR mouse-motion
# reports) and `997;1n` (the reply to Claude's dark-mode color-scheme probe).
#
# Why it happens (upstream, not our config):
# Claude Code enables mouse tracking (\e[?1003h) and fires terminal-capability
# probes the instant it starts. `suspended (tty input)` is a SIGTTIN job-control
# stop: if the process is stopped mid-init, those terminal modes are never torn
# down, so the shell — back in the foreground — echoes mouse movement and probe
# replies as garbage. Known Claude Code bug, closed "not planned":
# anthropics/claude-code #7807, #11898, #23581, #50032. No CLI flag / env var /
# setting disables the mouse tracking or the probes, so we wrap the command.
#
# The wrapper does two things:
# 1. PREVENT the stop — run claude with zsh job-control monitoring off, so zsh
# does not put it in its own background process group and it can never take
# SIGTTIN. This is the only lever available outside the binary; whether it
# fully prevents the stop is validated empirically (see the plan). Trade-off:
# you cannot Ctrl-Z-suspend a claude session while this is active.
# 2. SELF-HEAL — on exit (however claude ended), disable the terminal modes it
# may have left on and reset the line discipline, so a garble never persists.
#
# Opt-out: `export CLAUDE_TTY_FIX_OFF=1` (falls back to the bare binary).
# ============================================================================
# ----------------------------------------------------------------------------
# Safety net for the SIGTTIN-suspend case: when claude is *stopped* (not exited),
# the reset at the end of claude() below can NEVER run — the function is blocked
# inside `command claude`. So a suspended/killed claude leaves mouse tracking on
# and the terminal spews SGR mouse reports (e.g. `35;48;12M`). This precmd hook
# disables mouse tracking before every prompt, so the garble clears on your next
# Enter regardless of how claude died. Emits ONLY invisible mouse/cursor escapes
# (NOT bracketed-paste 2004, which zsh needs) so it can't disturb normal paste.
_claude_mouse_heal() {
[[ -t 2 ]] || return
printf '\e[?1000l\e[?1002l\e[?1003l\e[?1006l\e[?25h' > /dev/tty 2>/dev/null
}
autoload -Uz add-zsh-hook 2>/dev/null && add-zsh-hook precmd _claude_mouse_heal
claude() {
# Escape sequences that undo the terminal state Claude may leave enabled if it
# is stopped/killed before its own teardown runs:
# 1000/1002/1003/1006 mouse tracking, 1004 focus reporting,
# 2004 bracketed paste, 25 cursor visibility.
local -r _claude_tty_reset=$'\e[?1000l\e[?1002l\e[?1003l\e[?1006l\e[?1004l\e[?2004l\e[?25h'
if [[ -n "${CLAUDE_TTY_FIX_OFF:-}" ]]; then
command claude "$@"
return $?
fi
# `localoptions` confines the option change to this function; `nomonitor`
# disables job control so claude shares the shell's (foreground) process group.
setopt localoptions nomonitor
# Always-on debug capture (turn off: `export CLAUDE_DEBUG_LOG_OFF=1`). The
# trace goes to a FILE, not the screen, so the next real startup-race failure
# is recorded faithfully — same wrapper, same env, same nomonitor, no
# `script` surrogate to mask the job-control layer. Keeps the last ~20 logs.
local _dbg=""
if [[ -z "${CLAUDE_DEBUG_LOG_OFF:-}" ]]; then
mkdir -p ~/.claude/logs 2>/dev/null
_dbg=~/.claude/logs/claude-$(date +%Y%m%d-%H%M%S)-$$.log
for f in ~/.claude/logs/claude-*.log(NOm[21,-1]); do rm -f -- "$f"; done
fi
if [[ -n "$_dbg" ]]; then
command claude --debug-file "$_dbg" "$@"
else
command claude "$@"
fi
local -r rc=$?
# Only heal a real terminal, and write the reset to /dev/tty (the controlling
# terminal) rather than stdout — otherwise headless/piped runs like
# `claude -p '...' | jq` would get escape codes injected into their output.
if [[ -t 1 || -t 2 ]]; then
printf '%s' "$_claude_tty_reset" > /dev/tty 2>/dev/null
stty sane < /dev/tty 2>/dev/null
fi
return $rc
}
# ============================================================================
# On-demand MCP servers — the niche fleet is NOT loaded every session.
# ----------------------------------------------------------------------------
# Only ~9 broadly-useful servers live in ~/.claude.json and start every time.
# The niche ones live in ~/.claude/mcp-optional/*.json and load only when you
# ask for them via `claude --mcp-config <file>` (the flag MERGES them into the
# global set for that one session). These wrappers call the `claude` FUNCTION
# above, so an on-demand session still gets the TTY fix + self-heal.
# claude-rag knowledge-rag + memstack-skills (data dirs pinned, no cwd litter)
# claude-travel wander-agent + airbnb
# claude-sports sports-hub + sportscore + rundida
# claude-media video-analyzer + nakkas + open-museum + go-docs
# claude-extras ALL of the above at once
# ============================================================================
claude-rag() { claude --mcp-config ~/.claude/mcp-optional/rag.json "$@"; }
claude-travel() { claude --mcp-config ~/.claude/mcp-optional/travel.json "$@"; }
claude-sports() { claude --mcp-config ~/.claude/mcp-optional/sports.json "$@"; }
claude-media() { claude --mcp-config ~/.claude/mcp-optional/media.json "$@"; }
claude-extras() { claude --mcp-config ~/.claude/mcp-optional/*.json "$@"; }
# ============================================================================
# claude-debug — RAW-binary A/B probe for the startup-interaction failure.
# ----------------------------------------------------------------------------
# Runs the real binary via `command` (bypassing the claude() wrapper, so NO
# nomonitor) in your real interactive shell, with --debug-file capturing Claude's
# timestamped startup trace to a file (off-screen). This tests whether the
# failure needs the wrapper's nomonitor or happens on the bare binary too.
# NOTE: do NOT wrap this in script(1) — script runs Claude without zsh job
# control, which MASKS the SIGTTIN suspend we are hunting.
# Usage: run `claude-debug`, TYPE during startup to trigger it, then IMMEDIATELY
# run `jobs -l` and tell me the exact text the shell printed, plus paste the log.
# ============================================================================
claude-debug() {
mkdir -p ~/.claude/logs
local dbg=~/.claude/logs/claude-RAW-$(date +%Y%m%d-%H%M%S)-$$.log
print -r -- "→ RAW binary (no nomonitor), trace: $dbg"
print -r -- "→ TYPE during startup to trigger it, then run: jobs -l"
command claude --debug-file "$dbg" "$@"
print -r -- "captured: $dbg (exit=$?)"
}