Shell snapshot generation fails on extglob-dependent functions (e.g. bash-completion realpath)

[DnaMes]

Description

Codex CLI's shell snapshot bootstrap (used to give the sandboxed shell your interactive functions/aliases) crashes when the login shell has functions that rely on extglob patterns, because the snapshot generator dumps declare -f output without first running shopt -s extglob in the context that re-sources it.

Repro

Any system with bash-completion installed defines a realpath-style helper using extglob patterns like:

case $REPLY in
    */./*) ... ;;
    */.) ... ;;
    /..?(/*))
        REPLY=${REPLY#/..}
    ;;
    */+([^/])/../*) ... ;;
esac

This is valid bash only when extglob is enabled at parse time (interactive bash-completion enables it). When Codex captures the shell state via declare -f and replays it into a snapshot script without enabling extglob first, parsing the snapshot fails:

codex_core::shell_snapshot: Shell snapshot validation failed: Snapshot command exited with status exit status: 2: /home/<user>/.codex/shell_snapshots/<uuid>.tmp-<pid>: line 900: syntax error near unexpected token `('

Minimal local reproduction of the underlying parse failure:

bash -n <(bash -lc 'declare -f')
# /dev/fd/63: line 897: syntax error near unexpected token `('
# /dev/fd/63: line 897: `            /..?(/*))'

Impact

Any tool that shells out to codex exec (e.g. custom delegate/routing wrappers) fails outright with the above error whenever the invoking shell has bash-completion (or any other extglob-dependent function) loaded — which is the Linux distro default in many environments (confirmed on Fedora).

Expected

The shell snapshot generator should either:

• shopt -s extglob (and any other relevant shell option) before re-parsing captured function definitions, or
• Skip/guard functions it can't safely re-parse instead of failing the whole snapshot, or
• Capture/replay shopt -p state alongside declare -f so parsing context matches the original shell.

Environment

• codex-cli 0.146.0
• OS: Fedora Linux (bash-completion package installed)
• Shell: bash

Repro snippet to confirm root cause

shopt -s extglob
bash -n <(bash -lc 'declare -f')   # succeeds

[weshouman]

Confirmed on Debian 12 with codex-cli 0.145.0, bash 5.2.15, and bash-completion 2.11.

In this setup, every codex exec run logs:

ERROR codex_core::shell_snapshot: Shell snapshot validation failed

but the command itself continues and exits 0. The snapshot is discarded, so this is currently a degraded-but-nonfatal failure rather than a hard failure.

The underlying mechanism matches this issue exactly: the interactive shell has extglob enabled, declare -f dumps functions containing extglob syntax, but the snapshot is later sourced in a shell where that option is not enabled yet.

Minimal standalone reproduction:

cat >/tmp/min.sh <<'EOF'
shopt -s extglob
frob() { case "$1" in --!(no-*)dir*) echo dir ;; esac; }
EOF

bash -c 'source /tmp/min.sh; declare -f frob' >/tmp/snap.txt

bash -n /tmp/snap.txt
# syntax error near unexpected token `('

bash -O extglob -n /tmp/snap.txt
# clean

bash-completion triggers this without any custom shell code. Its _longopt function contains patterns such as:

--!(no-*)dir*)
--!(no-*)@(file|path)*)

One additional debugging annoyance: after validation fails, the temporary .tmp-* snapshot is deleted, so the path printed in the error is no longer inspectable.

A robust fix seems to be preserving the relevant shell option state alongside the function dump, rather than replaying only declare -f output. For this specific failure, enabling extglob before sourcing the dumped functions is sufficient.

It may also be worth making the log message explicit that snapshot validation failure means Codex is continuing without the captured shell snapshot.