Read two AskUserQuestion shapes that fall through to the raw mirror

[konpyl]

Two AskUserQuestion shapes currently fall through to the raw mirror. On a phone that means a wall of ~180-column terminal text with nothing to tap and no way to answer — both were hit in normal use today, in different repos.

Both come from the same root mistake: a grammar asserting a layout instead of reading it.

1. A wrapped option label kills the preview dialog

preview-select required its numbered rows to be on consecutive lines. The assumption is written down as fact:

// 3. Option rows: numbered rows 1..k in CONSECUTIVE lines (the left column
//    never wraps its 30-col labels)

That gutter is only ~30 columns, so Claude Code wraps any longer label onto continuation rows and the numbered rows stop being adjacent — the whole dialog bails to raw. Observed live: an option labelled Rename + keep old as hidden aliases (Recommended) occupied three rows, putting the options on lines 183/186/187.

Reproduced by running the real parseAnsi → splitLines → buildBlocks pipeline over the capture, and flipped by the single change of making those rows adjacent — nothing else about the dialog mattered.

Adjacency is replaced by a contiguity walk keyed on the label column (the width of the ❯ 1. prefix): left of it is a new option, at it is label text, right of it is not label text at all. That one rule also prevents a label ending …and 3. Backfill later from minting a phantom third button, and stops preview-pane box-drawing being folded onto a button in layouts where the Notes: column and the preview pane don't coincide.

2. A multiSelect question inside a multi-question wizard is owned by nobody

wizard refuses it correctly — a wizard digit selects and advances, a checkbox digit only toggles, so claiming it would mis-drive the dialog — and multi-select only ever handled the single-question form. Two gates encoded that: a stepper of exactly two chips, and an advance row spelled Submit.

multi-select now carries the wizard's chips (via the wizard's own parseStepperLine, so the chip grammar stays in one place) and navigates with the wizard's own Left/Right keys, following the pattern preview-select already established for this exact shape. The advance row's label is read, not assumed: it is Next on every step but the last.

The row is located by position — last non-blank line above the rule separating the menu from the escape row — never by scanning for the word. Pane text is model-authored; matching on text alone let an option description reading Next rename the button to an action it doesn't perform, satisfy the "row absent → fail closed" check on a pane with no advance row, and forge the pointer so Enter landed on whatever row the terminal's ❯ was really on (on Chat about this, that aborts the whole tool call).

The advance macro stops at the keystroke rather than predicting what follows — it verifies the pointer in the same iteration it presses Enter, then lets the next poll re-detect whatever appeared. The single-question review stays where it was; a multi-question dialog reviews once at the end, and that screen belongs to wizard.

Evidence

The claude--wizard-multiselect-* fixtures are sandbox captures of a real two-question dialog driven end to end, because nothing in the repo described this shape and guessing feeds a keystroke into a live terminal. They settled four things worth knowing:

• the question chip flips ☐ → ☒ on the first tick — "answered" means touched, not complete;
• digits toggle pointer-independently inside a wizard step, exactly as standalone;
• the last step's row reads Submit, earlier ones Next;
• the footer is not invariant — it gains ctrl+g to edit in Vim when the pointer reaches the advance row, which is why the signature already stops before it.

claude--wizard-preview-wrapped-label.txt is derived, not captured: its live specimen was real work and this repo is public, so the gutter of claude--wizard-preview-q1.txt was rewritten and everything from the Notes: column rightward carried over untouched. Flagged as derived in the fixtures README.

Existing single-question dialogs are unchanged (steps: null, review phase untouched), and every non-dialog fixture still yields zero detections.

Also in here

• The race guard now compares the wizard steps. signature normalises ☒/☑ → ☐ across the whole chip line — it has to, since the current chip flips on the first tick — but that also erases which step you are on, so two questions of one wizard sharing a question text and option labels were byte-identical to the guard.
• The stepper strip is one shared component; wizard, preview-select and multi-select all render it from the same place instead of three copies.
• multi-select-action.ts carried a literal NUL byte as a map-key separator, which made git classify the file as binary and hide every change to the race guard from git diff. It's now a source escape, plus a .gitattributes rule so a future stray byte can't cost reviewability.

Notes for review

• One commit, code only. No version bump and no CHANGELOG entry — this comes from a fork with its own numbering lane, so slotting it into your release is yours to do.
• Four new fixtures are pinned in prompt-binding-regions.json; the completeness gate and the bridge-side binding test both pass.
• Gate on top of main: 1271 web tests pass, tsc clean.

🤖 Generated with Claude Code

[konpyl]

Rebased onto main (ae8f562, 0.20.2) — it was cut from c674af5, so it had 33 commits of drift. Still one commit, no conflicts, and re-verified against the current tree: tsc clean, 1271 web tests and 479 bridge tests pass.

Worth confirming the change is still needed rather than just still applying: preview-select.ts on main today still carries the adjacency check —

if (rows.some((r, k) => r.n !== k + 1 || r.index !== rows[0]!.index + k)) return null;

— so a wrapped option label still drops the whole dialog to the raw mirror.

Also, thank you for 4841e37. You were right, and it's a better fix than mine: command -v reporting a shell function as a bare word makes dirname return ., and my version would then have put the working directory on the PATH used by every later git/systemctl/tailscale in that script. I've cherry-picked it into my fork and reproduced both behaviours to confirm. The point about the lifecycle suite stubbing BUN=/bin/true — and that being exactly why a missing Bun failed four deploys unnoticed — landed too.