3 min readfeature
The keyboard round: the shortcut that never worked, and five more finds
The planning round's headline shortcut never worked on a fresh install. The keyboard round found it and five more, fixed all six, verified live.
- #planning
- #keyboard
- #tui
- #release
The planning round shipped Shift+Tab mode cycling — the Claude Code lane — and documented it in two places. One of them was wrong about what the key did, and the honest answer is: both were, because the shortcut never worked. Not on a fresh install, not once. This round audited every interactive surface in plan mode and fixed what the audit found.
The shortcut that never worked
unown's plan-mode cycle registers as an extension shortcut on shift+tab.
The pi coding agent underneath drops any extension shortcut that collides
with a reserved built-in default — and app.thinking.cycle defaults to
exactly that key. On every fresh install, Shift+Tab cycled thinking levels,
the permission cycle never ran, and the docs contradicted themselves about
which one you'd get.
The fix moved the vendored thinking-cycle default to Alt+T (free,
terminal-safe, mnemonic). Your explicit keybindings.json still wins, and
the runner test now pins that shift+tab registers under default
keybindings — the exact failure can't come back silently.
Five more finds
- Questionnaire arrows pointed the wrong way in half the terminals. The ask-user-question tab switcher matched raw escape bytes, listed SS3 right-arrow in both the left and right branches, and never listed SS3 left at all. In application-cursor mode (DECCKM), right moved you back. It now matches logical keys through the same matcher everything else uses — CSI, SS3, Kitty CSI-u, modifyOtherKeys and Windows encodings all route identically.
- Typing an emoji corrupted the input line. The input buffer split non-escape input per UTF-16 unit, so a directly-typed astral character (emoji, rare CJK) became two lone surrogates the editor happily inserted. Pastes and Kitty-protocol terminals were safe; legacy terminals were not. The buffer now emits whole code points and the editor rejects unpaired surrogates as defense in depth.
- The approval gate ignored half the keyboard. PageUp, PageDown, Home,
End — bindings that exist and every other selector honors — did nothing.
Options are now numbered: press
1-4to pick an approval lane directly, in plain, Kitty and modifyOtherKeys encodings. Lists longer than nine stay unnumbered rather than display an unreachable number. - A stray Esc destroyed your plan edits. The plan editor cancelled on the first Esc with no dirty check. Now the first press arms a "press Esc again to discard" guard, any other key disarms it, and the second consecutive press discards. No timer — deterministic, and testable without a clock.
- The plan-mode bash gate had a curl write primitive.
curl -o,--output,-O,--remote-name,-Dand clusters like-sOwrote files inside "read-only" plan mode. All of those forms now sit on the destructive side; read-only fetching still works.
Verified like a user, not like a test suite
Every fix has a regression pin, but the headline got more than that: a live
PTY session drove the real TUI against the real model — Shift+Tab stepped
default → acceptEdits → plan → default, the gate rendered its numbered
options, digit 1 approved, and the agent executed the approved plan (it
created the file and ran the verification command). The questionnaire
rendered live too, numbered options and all.
The receipts — findings, fixes, milestones and the resumable log — live in the session doc.
Spotted something wrong or worth adding? The record is the repo — open an issue.