[tebriz91]

0.3.0 item 3 (ADR 0073).

What changes

• One table, src/shortcuts.rs: every user-facing binding as a row (label, context group Everywhere / Editor / Popup / Panel), each binding with its action, key context, keys per platform (mod- = ⌘ or Ctrl) and the bind_keys that registers it. Every bind_keys registers its rows from there, so each module keeps its place in the keymap (GPUI breaks a tie between equally deep bindings by registration order). The mdoc_editor and gpui_pdf crates keep their own keymaps for their tests and other hosts; the app registers their rows from the table (shortcuts::bind_crate).
• Plumbing stays put: Tab inside a control group, moving inside menus, choosers and the palette list, caret and deletion basics, the editor's own Esc, the OS character palette.
• Palette: shortcut labels (and the sidebar's ⌘K / Home hints and Home's Open hint) come from the table; new action "Keyboard shortcuts". Pseudonymize now shows ⇧⌘P, which it was missing.
• Sheet (src/workspace/shortcuts_ui.rs): read-only modal like the confirm dialog: filter field, rows grouped by context, this platform's keys only, list scrolls with arrows / Page Up/Down / Home / End, Tab stays inside, F6 blocked, Esc or click outside closes, focus returns. Opens with ? outside text fields, ⌘/ (Ctrl+/) everywhere (toggles), the palette, and Help → Keyboard Shortcuts on macOS.
• Headless renders: new keyboard-shortcuts screen in light and dark.
• main.rs::bind_all_keys is the one registration order, shared by launch and the headless renders.

Keymap unchanged

I dumped the full keymap before and after on Windows: the same 216 bindings plus the sheet's three (?, ctrl-/, Esc in the sheet). A few same-key pairs changed relative order inside pii, ui and shell; each is between contexts that never sit on the same element, so neither can win a tie. The ties that exist (Enter in the chooser, which is also PseudonymReview; Esc in the Replacements panel, which is also UiMenu) keep their order.

Tests

• src/shortcuts/tests.rs: a bound action that's neither in the table nor named plumbing fails; a table row nothing registers fails; the editor and PDF crates' keymaps must match the table on both platforms. Checked that a stray ctrl-j binding fails it.
• src/workspace/tests/shortcuts_sheet.rs: ? on Home opens it and every table row is drawn; ? types in the editor and the filter; ⌘/ toggles from the editor and focus returns; filter by word and by key; F6 stays inside; the palette opens it, and every key the palette shows is in the table.
• Rebased Windows validation: just lint, just test, just fmt-check and git diff --check pass. Full workspace: 820 passed, 20 ignored, zero failures; includes 433 app tests, the isolated headless-render contract (light/dark shortcuts sheet), DOCX worker checks and release-script regressions. Native macOS/Windows keyboard, accessibility and presentation acceptance was not run in this takeover.

🤖 Generated with Claude Code

Codex review

Round 1 — gpt-6.1-sol, high

The first review reported three P2 findings; all were accepted and fixed in the review-fix commit:

1. Commands could move focus behind the sheet. New could focus an obscured editor while the sheet stayed open. The sheet now closes when a command moves focus outside it, while switching apps keeps it open. Palette, Settings and Open Case are blocked while the sheet is open. Regression a_command_that_moves_focus_closes_the_sheet checks New, typing into the resulting editor, palette blocking and window deactivation.
2. PDF and confirm-dialog shortcuts escaped the table and invariant. PDF Page Up/Down, Home/End and Go to Page, plus Discard in the save prompt, now use bound actions registered from the table. The crate-keymap invariant checks both platforms, and the sheet test draws every listed row.
3. Comparison windows had no ShowShortcuts handler. App-level routing now activates the workspace and toggles its sheet. the_sheet_opens_from_another_window checks dispatch from another window; comparison-window availability includes ShowShortcuts. Native macOS Help-menu acceptance remains pending.

The minor missing changelog fragment is fixed by changes/shortcuts-panel.md. A separate render check found no focus ring on the filter; the filter now has an accent border while focused.

Integration and documentation

Rebased onto origin/main 6a6267f, preserving #201 restore wording, #202 save retry, #203 tab numbering and #205 fleet documentation. ROADMAP items 1, 2, 3 and 6 are absent. The former WIP is now a proper finishing commit: ADR 0073 is Accepted in ADR 0052 current-rules style, with the changelog fragment and acceptance row H23.

Round 2

Independent second round: provider codex, model gpt-6.1-sol, reasoningEffort: high, reviewing the rebased diff against origin/main and the earlier replies. One initial attempt failed at model capacity; the retry completed.

• P2: Settings' restrictions could be bypassed by opening shortcuts over it. Reproduced Settings → Mod+/ → Mod+N → typing: New created a second tab behind visible Settings. Fixed by extracting Settings' existing document-command blockers into settings_ui::block_document_commands and applying the same helper to the shortcuts sheet while Settings is open. The regression shortcuts_over_settings_keep_document_commands_blocked failed before the fix (two tabs rather than one), then passed. It checks unchanged document text, no new tab, sheet focus retained, and focus returned inside the still-open Settings on Esc. ADR 0073 states this rule.
• The review otherwise found the PDF/Discard fixes, cross-window routing, keymap registration order, theme-render coverage and documentation consistent, with Restore bar: say restored names are proposals and what Apply does #201–Number new untitled tabs #203/docs: Fleet section links the fleet skill for models and reviews #205 preserved.

This is static review plus Windows headless tests. Native macOS Help/keyboard and native accessibility/presentation acceptance remain unverified. Final validation at head 94fde9a: just lint, just test, just fmt-check and git diff --check pass. Workspace: 820 passed, 20 ignored, zero failures (433 app tests passed); isolated headless renders and release-script regressions passed. All four P2 findings across both rounds are fixed; no finding was deferred. Hosted checks for this pushed head are separate from these local results.