Prompt renderer, extglob and shopt: 8 findings from building a 30-theme config on 2.10.2

[LESdylan]

Everything below was found while rebuilding hellishrc_plugins against 2.10.2 — a config framework with a 30-theme prompt engine, so it leans hard on the prompt renderer, the hook arrays and shopt. Each item was reproduced in a real pty (not just -c) before being written down, and each says what it cost downstream.

hellish, version 2.10.2 (hellish-linux-x86_64)
Ubuntu 22.04.4, xterm-256color, COLORTERM=truecolor, LANG=en_US.UTF-8
user-mode install: ~/.local/bin/hellish

The prompt work in 2.9/2.10 is genuinely good — % and \ in one PS1, the self-spacing badges, HELLISH_*_FUNCS, right prompts. Most of what follows is the sharp edges around that.

---

1. print -rP and the prompt renderer disagree about variable content

The most consequential one, because it makes prompt preview lie.

The prompt renderer does not re-scan what a variable expanded to. print -rP does. So the same theme renders two different ways depending on which path you go through.

$ V='50%off'; print -rP "[${V}]"
[50ff]                                   # %o was eaten as a prompt escape

$ V='100%done'; print -rP "[${V}]"
[100/tmp/…/scratchpadone]                # %d expanded to the cwd

versus the real prompt, in a pty:

PS1='[${V}] '   with V='50%off'    →  [50%off]        # correct, left alone
PS1='[${V}] '   with V='%F{46}x%f' →  [%F{46}x%f]     # inert, not colour
PS1='[${V}] '   with V='x\g y'     →  [x\g y]         # inert, not the branch

Both behaviours are defensible; having both is not. share/hellish/rc.d/40-prompt-switch.hsh ships prompt preview built on print -rP, and it already carries a comment explaining that it doubles % "because print -P always enters through the zsh frontend" — that comment is documenting this bug rather than a design.

Why it matters: any prompt whose content comes from a variable — which is every prompt with a computed segment, since $( ) is unavailable (§2) — previews as something you will never actually see. A branch called feature/50%-done or a directory called 100%done is enough.

Suggested: make print -rP treat post-expansion text as literal, the way the renderer does. Failing that, document the split loudly, because the natural assumption is that the preview renderer is the renderer.

Worked around by escaping every segment on the way into a preview and restoring it after — _hx_preview_escape in lib/prompt.hsh.

---

2. $( ) is never expanded in a prompt, and nothing says so

PS1='[$(echo LIVE)] '   →   [$(echo LIVE)]      # literal, forever
PS1='[${VAR}] '         →   expands every render
PS1='[$((6*7))] '       →   [42]

hellishrc.example says "Live $VAR / ${VAR} expansion happens at every render, and so does $? and the other single-character specials" — true, and it never mentions command substitution, so the reader is left to infer it from an omission. Every other shell expands $( ) in PS1, so the inference goes the wrong way.

This is not a small limitation: it is the difference between "a theme is a string" and "a theme needs a hook, a variable, and an invalidation story". The whole segment architecture in my engine exists because of it.

Suggested: either expand it (behind setopt PROMPT_SUBST, as zsh does), or state the exclusion explicitly next to the $VAR line. A prompt that silently prints $(git branch) looks like a broken config, not an unsupported feature.

---

3. shopt -s extglob is accepted, reports on, and then fails to parse

$ hellish -c 'shopt -s extglob; shopt extglob'
extglob             	on

$ hellish -c 'shopt -s extglob; case ab in @(ab|cd)) echo MATCH;; esac'
hellish: line 1: syntax error near unexpected token `('     # exit 2

This is worse than an unimplemented option, for one specific reason: the failure is at parse time, so the guard everyone writes does not work.

if shopt -q extglob; then       # true
    case "$x" in @(a|b)) ... ;; esac   # the whole FILE fails to parse
fi

You cannot feature-detect your way out of it from inside the file — the branch never runs because the file never parses. The only safe probe is out-of-process (hellish -c '…'), which is what my capability probe ended up doing.

Suggested: make shopt -s extglob fail and report off until the lexer supports it. An option that lies about being on is strictly worse than one that refuses.

The other glob knobs are fine now, and worth noting as fixed: dotglob, nocaseglob, globstar and nullglob (arrays included) all behave correctly on 2.10.2, where they were accepted no-ops on 2.7.6.

---

4. PS2 is silently ignored

$ PS2='CUSTOM2> '
$ echo one \
> two            # still '> ', not CUSTOM2>

PS2 is settable, readable, and has no effect. hellishrc.example and several shipped themes set it, which suggests it is expected to work. Either honour it or drop it from the examples, so nobody else spends time wondering why their continuation prompt never changes.

---

5. COLUMNS and LINES are never set

$ shopt -s checkwinsize; pretty on resize-aware; /bin/true
$ echo "[${COLUMNS:-unset}] [${LINES:-unset}]"
[unset] [unset]

checkwinsize is on by default and pretty on resize-aware is advertised as "track the terminal size after every command" — so the shell clearly knows the size. It just never exports it, where bash maintains both in interactive shells.

Cost: anything that needs the terminal width has to fork tput cols. In a prompt that is a process per render, which is exactly the thing a fast prompt cannot afford. My full-width-rule theme has to declare itself expensive and cache the answer for the life of the shell, so it gets the wrong width after a resize.

Suggested: set COLUMNS/LINES when checkwinsize/resize-aware is on. The tracking already exists; this is publishing it.

---

6. Missing variables: EPOCHREALTIME, SRANDOM

$ echo "[${EPOCHSECONDS}] [${EPOCHREALTIME:-unset}] [${SRANDOM:-unset}]"
[1789416704] [unset] [unset]

EPOCHSECONDS is there, EPOCHREALTIME is not. For prompt timing, one-second resolution is not useful — date +%s%N is a fork, which is the cost you were trying to avoid by having EPOCHSECONDS at all. SRANDOM is minor by comparison.

---

7. No bind, so key bindings cannot come from a config

$ hellish -c 'bind -p'
hellish: bind: command not found

set -o emacs / set -o vi work and are the only line-editing control an rc has. With complete landed in 2.10, bind is the one remaining hole where a config cannot reach the line editor at all — no history-prefix search on Up, no completion-ignore-case, no rebinding C-w.

Not a bug, but it is the highest-value missing builtin from a configuration author's point of view, and worth a line in the docs saying it is unavailable rather than leaving people to discover it.

---

8. Docs and examples that no longer match the binary

Small things, but each one costs a reader real time because they read as authoritative.

• USER_DOC.md has no prompt section. Everything about the prompt — the bilingual PS1, the badges, the hook arrays — lives in hellishrc.example comments and wiki/interactive.md. grep -n '^#+ ' USER_DOC.md for prompt/PS1/theme returns nothing. The example file is doing the job of the manual.
• wiki/interactive.md says prompt lists 29 themes. prompt is not a builtin; it is a shell function seeded into ~/.config/hellish/rc.d/40-prompt-switch.hsh. On a machine where that seed did not run, prompt is command not found and the wiki looks wrong.
• \A shadows bash's 24-hour clock escape. Documented, and the right call given it shipped first — but a bash prompt pasted in gets an animation frame where it expected 14:30, with no warning at the point of use.
• The animation warning deserves the docs, not just a comment. hellishrc.example says \A "has been reported to corrupt the prompt on some terminals … not reproducible on demand". That is important and currently only visible to someone reading the example file.

---

What works well, for balance

Worth saying, since a bug report is a biased sample. All verified on 2.10.2, all fixed since 2.7.6:

complete / compgen (with shopt -s progcomp) · printf %q · printf '\e' · declare -F and alias -p · subscripted array writes from a function defined in a sourced file · local -n / declare -n namerefs · [ -v VAR ] · read -n N · globstar, dotglob, nocaseglob, nullglob in arrays · process substitution <( ) · dirs as a builtin · mapfile / readarray -d · FUNCNAME with a real call stack · BASH_SOURCE · trap DEBUG

And specifically in the prompt: RPROMPT right-aligns correctly and is erased as the line grows; %F{n} %K{n} %B %~ %N<…< %(?.a.b) %(!.a.b) all match zsh; $? is correctly preserved across HELLISH_PRECMD_FUNCS, so a status badge shows your command's result and not the hook's:

PS1='[rc=%?]\S> '  with a precmd hook installed
    false  →  [rc=1] ✘1>      # correct

That last one is easy to get wrong and you got it right.

---

Priority, if it helps

1	§1 preview/renderer split	makes the shipped prompt preview lie
2	§3 extglob reporting on	cannot be guarded against from inside a file
3	§2 $( ) undocumented	changes how every dynamic prompt must be built
4	§5 COLUMNS unset	forces a fork per prompt for width
5	§4 PS2 ignored	silent, and the shipped examples set it

Happy to open these as separate issues, supply the pty harness, or test patches — the harness that produced every transcript above is a ~40-line pty.fork() driver and I can attach it if useful.

[LESdylan]

Status of each section, checked on develop (9beab62) with a pty where the behaviour is interactive:

§	Finding	Status
1	prompt preview disagrees with the live prompt	Open PR #159. ${var@P} now renders through the PS1 renderer itself, and prompt preview uses "${PS1@P}" for PS1 themes. A 100%done or 50%off segment previews exactly as the prompt shows it.
2	$( ) in a prompt is undocumented	Merged in #154. USER_DOC.md has a prompt section with a table of what is live, and it and hellishrc.example both show the hook pattern for a computed segment. tests/prompt_doc_example_test.py runs that example straight from the doc.
3	shopt -s extglob reports on, then fails to parse	Already matches bash; details below.
4	PS2 ignored	Merged in #155. PS2 is the prompt of every continuation line.
5	COLUMNS / LINES unset	Already works in 3.0.0 and later. In a pty, /bin/true; echo "[$COLUMNS] [$LINES]" prints [97] [31], then [120] [40] after a resize.
6	EPOCHREALTIME, SRANDOM	Merged in #157.
7	no bind	Merged in #152. bind works like bash's, including bind -x, and zle -N / bindkey exist too. Up/Down history prefix search built on it is Univers42/hellishrc_plugins#1 (#136).
8	docs out of date	Merged in #154. It covers the prompt section, prompt being a function seeded into rc.d, the \A clock clash (with \D{%H:%M} / %T as the replacement), and the animation warning moved into the docs.

§1, why @P rather than changing print -P. print -P scanning the substituted value for % is zsh's documented behaviour, and zsh configs depend on it. The bug was that the preview used the zsh path for a bash-rules prompt, because hellish had no way to render a PS1 exactly as the prompt does. bash's ${var@P} is that way. It is now implemented on the same zsh_to_ps1 + ps1_render pair that draws the live PS1, and a pty test checks that the live prompt and ${PS1@P} are byte-for-byte equal. With it, _hx_preview_escape in lib/prompt.hsh is no longer needed on a hellish that has #159.

§3, extglob. On develop, shopt extglob reports off, and it takes effect for lines parsed after it, which is the same rule bash follows:

# a file: shopt on line 1, the pattern on line 2
shopt -s extglob
case ab in @(ab|cd)) echo MATCH;; esac      # MATCH, in both shells

hellish -c 'shopt -s extglob; case … @(…)' is a syntax error in bash too: the whole line is parsed before shopt runs. So is the if shopt -q extglob; then case … @(a|b)) …; fi guard, since the if is parsed as one command. The out-of-process probe is also what bash users need.

This check turned up one real difference: the line number in the error. For a syntax error inside an if, hellish reports the line after the end of the file, where bash reports the line of the bad token. It is tracked on its own and is unrelated to extglob.

---

Generated by Claude Code