Pad/doc/spec.md
Greg Pomerantz 6968f6a284 Restore scroll by logical line, not pixel offset (wrap-aware)
On device: scroll far down a wrapped file, relaunch, and the app lands
further DOWN than where the user left off — the deeper the scroll, the
further off.

Root cause: the persisted Scroll is a pixel offset in VISUAL-line space.
Restoring maps it through the WrapIndex (scrollDecompose -> LineForVisual),
but on relaunch every count is the estimate (1) until the line is shaped,
and shaping covers the visible window only — the lines ABOVE the restored
viewport are never shaped. With all-ones counts LineForVisual maps the
offset 1:1, landing a logical line deeper by every wrapped continuation
above the viewport, and the state is stable (the under-counted lines never
re-enter the window), so it never self-corrects.

The snapshot now persists wrap-independent coordinates: the logical line
at the viewport top (derived with the same mapping the layout uses,
against the current index, so it is exactly the shown line) plus the
sub-line remainder. The restore re-derives the offset as line*lh + sub,
which maps to the saved line under any wrap state (all-ones or populated).
The raw Dp offset is kept for pre-line-coordinate session files
(loadSession defaults the missing key to -1; BeginRestore rejects the
ambiguous zero value: a genuine line-0 snapshot always has Scroll < lh).

TestRestore_ScrollSurvivesWrapState reproduces it: 150 lines recorded as
wrapped x3, viewport at logical line 200 (visual 500); a relaunch with a
fresh WrapIndex must land the window on line 200. Fails pre-fix (window at
line 254, i.e. deeper) and passes with the fix.

Docs: spec §2.4 (line-based scroll persist + current save policy),
architecture §6.7 (why the offset is unrestorable by re-mapping).
2026-08-20 19:31:51 -04:00

13 KiB
Raw Blame History

Pad — Text Editor Specification

This is the specification of what Pad actually is (v1, current code). Behavior that was once spec'd but never built is listed as deferred in §7, not described as a requirement. If code and this document disagree, the code wins — and this document should be fixed.

1. Overview

Pad is a minimal plain-text editor for Android, written in Go with the Gio UI framework. It presents a single root directory of text files (a browser) and a full-screen editor with the Android soft keyboard. The design goal is a fast, honest editor for a directory that is synced across devices — instant open, autosave on every edit, and large files that stay smooth.

Design philosophy:

  • Plain text only — no formatting, no syntax highlighting.
  • Automatic everything — autosave with a 1 s debounce; there is no save button.
  • Minimal surface area — a browser and an editor, nothing else.
  • Bounded memory — a measured file-size limit with an explicit "too large" state, instead of silently degrading or OOMing.

Platform: Android-first. Window 390×844 dp. Default root directory is /storage/emulated/0/Notes on Android (overridable with -root); . elsewhere.

2. Implemented behavior

2.1 File browser

  • Lists files and directories under the current directory.
  • Sort: four modes — name asc/desc, date asc/desc — cycled by the sort button. Default: date descending (newest first).
  • Search: incremental, case-insensitive filter over entry names via the search bar (a Gio widget.Editor).
  • Pagination: directory entries are loaded asynchronously in pages through the worker pool; the first paint does not wait for a full directory read.
  • Navigation: tap a directory to enter it, back to return. Single tap opens a file — no long-press menus, no file management actions.

2.2 Editor

  • Input via the Android IME: soft-keyboard typing, composition, backspace, and autocorrect replacements all arrive through one IME replace-range path and are converted from rune indices to byte offsets. (Swipe-typing support follows from the same IME path; final sign-off on a physical device is the one open validation item.)
  • Word wrap: on by default; wrapped lines are virtual — the buffer stores only real newlines. Scrolling wrapped content is continuous: the viewport moves pixel-by-pixel with the finger, including across the bottom of a wrapped line (no jumping over wrapped remainders) — the scroll offset maps to content through the per-line visual-line counts corrected by the renderer every frame (architecture.md §6.2).
  • Virtualized viewport: only the visible byte range is shaped and drawn each frame (typically ~4 KB of a large file), keeping frame cost and shaper memory constant regardless of file size.
  • Cursor + scroll: tap to place the cursor, drag/scroll to pan; with a hardware keyboard, arrow keys, Home/End, and Page Up/Down move the cursor (verified on the Android emulator; Gio's mobile focus-navigation default for arrow keys is overridden — see architecture.md §2.1).
  • Text selection: two input paths. Touch (the Android-native model): long-press selects the word under the finger (on a blank spot it places the caret and offers a paste-only menu); double-tap selects the word; the selection has drag handles at both ends (resize) and the highlighted body can be dragged to move it (length preserved); a floating menu offers copy/cut/paste when a selection is active and paste alone for a bare caret. Any menu tap closes the menu; copy keeps the selection, cut removes it, and paste inserts at the caret or replaces the selection. A plain tap still places the caret and clears any selection; taps inside the visible menu are ignored by the editor. Hardware keyboard: shift+arrow extends a selection. In both cases insert, backspace, and delete replace the selected text, and the IME replaces a selection when the user types over it. The selection is highlighted in the editor and pushed to the IME. (Shift state is tracked by the app, since Gio's Android bridge drops modifier keys — architecture.md §2.1.)
  • In-file search: a find icon on the top bar (right side) toggles a find bar directly below it, with a text input, an "N / M" counter, next/prev navigation, and a clear (X) button. The X empties the query and the results but leaves the bar open; the top-bar icon closes it (and re-opens it). Opening the bar moves key focus to the search input and raises the soft keyboard. The query is a plain substring, case-insensitive, no regex; each change is scanned over the entire in-memory content by a worker (Search task), so typing never blocks the UI and stale scans are dropped by generation. Next/prev select and scroll to the surrounding matches (wrapping around); the first discovery selects and scrolls to the first match. Edits do not invalidate the result set: match offsets before the edit are unchanged, matches after it shift by the length delta, and matches overlapping the edited range are dropped (finding text an edit creates requires re-typing the query).
  • Autosave: every edit restarts a 1 s debounce; on expiry the full content is written to disk by a worker. At most one write per file is in flight at any time; saves requested during a write are deferred and re-issued with the newer content when it completes ("latest state wins"). Writes stage to a unique per-call temp file and rename into place, so a crash or a concurrent reader never observes a partial file. Failed writes are retried. This is the only document persistence mechanism (the app's own state is a separate tiny session file — §2.4).

2.3 Large files (measured)

  • Editable limit: 50 MB (MaxEditableFileSize), measured on-device: a 10 MB file (130,954 lines) opens in ~120 ms (stat ~76 ms, read ~27 ms, line-index ~18 ms) and idles at ~150 MB PSS / ~230 MB RSS, flat under scroll. 50 MB extrapolates to a few hundred MB — acceptable on a modern phone.
  • Files above the limit open into a "too large to edit" state: a notice is shown, edit operations are no-ops, and the browser still lists the file.
  • The buffer is chunked (64 KB chunks, prefix-sum offsets, full load for in-range files); edits splice only affected chunks. Details: architecture.md §6.

2.4 State restoration on relaunch

  • On launch, Pad re-opens the file from the last session and lands straight on the editor page, restoring the cursor, scroll position, live selection, and the find bar (query, open/closed, current match). A relaunch therefore never requires re-browsing to the last file.
  • The scroll position is persisted as the LOGICAL line at the viewport top plus the sub-line remainder (the raw pixel offset is kept too, for pre-line-coordinate session files). The pixel offset alone is not restorable: it lives in visual-line space, and the wrap counts that map it to a line are rebuilt from estimates on relaunch — lines above the restored viewport are never shaped, so the offset would land a line deeper by every wrapped continuation above it.
  • The snapshot is a tiny JSON file (a few hundred bytes) written by the cmd layer: $HOME/.pad/session.json off-Android, and /storage/emulated/0/Pad/session.json on Android (the dir that already carries the browser's .pad index caches). Search results are NOT stored: they are regenerated by re-scanning the restored query, and the current match is re-selected by byte offset.
  • The snapshot is written rate-limited (250 ms, only on change) while the app runs; a file change or a selection appearing/vanishing saves immediately; on Android the activity's onStop (recents-wipe, app switch) flushes it one last time; and it is written unconditionally at shutdown. A kill shortly after a change therefore loses at most a fraction of a second of state.
  • If the restored file no longer exists (deleted/moved, e.g. by Syncthing), the app falls back to the browser page; positions that exceed a shrunk file are clamped to its new end.

3. Code organization (actual)

cmd/pad/                  # Gio entry point, main loop, frameReceiver,
                          # Android defaults (impl_android.go)
internal/
  browser/                # BrowserState, BrowserManager, sort, search,
                          # pagination, layout, handlers
  editor/                 # Logic goroutine, State, ChunkedBuffer, LineIndex,
                          # IME handling, autosave, relaunch session (session.go),
                          # Frame handoff (frame.go)
  io/pool/                # Worker pool (priority lanes), task types,
                          #   real/  — real filesystem (rooted at /)
                          #   mock/  — in-memory FS for tests
                          #   types/ — shared types (LineIndex, ...)
  ui/                     # Element model, Renderer, units (Dp/Px), theme
  ui/icons/               # Vector icons
  test/e2e/               # Logic-level e2e harness + tests
doc/                      # spec.md (this file), architecture.md,
                          # development_plan.md

4. Architecture (summary)

The logic goroutine is the sole owner of mutable state; the main (Gio) goroutine reads only a Frame snapshot and writes only via channels; a frame-receiver goroutine stores frames and invalidates the window; a priority-aware worker pool does all file I/O. No mutex guards application state. The IME snippet, search box, and gesture state live on the main-goroutine side because Gio mutates them during draw.

Full details, channel topology, and ownership rules: architecture.md.

5. Invariants to preserve (future development must not break these)

  1. Single owner, no locks on state (architecture.md §1). Any new feature adds a channel or an owner-executed callback — never a direct state touch from another goroutine.
  2. Main never reads logic State — only the Frame snapshot.
  3. Never shape the whole file. The text shaper's internal line storage retains the largest layout ever performed; one whole-file layout permanently inflates memory (this caused a 1 GB leak, fixed in Phase 3).
  4. The IME snippet is the visible window, and rune↔byte conversion happens in one place (HandleReplaceRange / RuneIndexToByte).
  5. Persistence goes through the owner. Autosave dispatches WriteFile tasks for document content; the relaunch session (last file, cursor, scroll, selection, find state) is marshaled by the owner and written by the cmd layer's saver callback (tiny JSON, synchronous, torn writes rejected on load). If session state ever grows beyond a few KB, move the write to a worker-pool task (autosave pattern).
  6. Logic work stays < 16 ms. Anything that can block or scan more than the viewport goes to the worker pool.
  7. Scroll offset is always clamped to [0, maxScroll], where maxScroll = contentHeight viewportHeight floored at 0 and contentHeight counts VISUAL lines (a wrapped line is taller than one line pitch). A file whose content fits the viewport has maxScroll = 0 and cannot scroll. Verified on-device across 2→130,955 lines (development_plan.md Phase 6); the wrap-aware clamp lands exactly on the file end for wrapped content (Phase 13).

6. Performance expectations (validated on-device)

Operation Target Measured (10 MB file, Android 35 emulator)
Open file instant feel ~120 ms (stat + read + line index)
Scroll 60 fps logic-frame cadence flat across offsets 0.02→1.0 — no large-offset degradation; present fps is emulator-limited (~1427 on the software-rendered emulator, not a Pad quality metric)
Type responsive single IME path; rapid commits land cleanly
Memory bounded PSS plateaus ~250 MB at a 10 MB file (bounded high-water mark, no growth under sustained scroll; no OOM)

7. Deferred / not implemented (explicit non-goals for v1)

These were in the original v1 spec but are not in the code. They are recorded here so future rounds don't mistake doc text for behavior:

Feature Status Notes
Undo (any) not implemented No undo stack exists; the old SaveUndoTask is dead code.
External change detection not implemented No mtime compare on open/resume, no watcher, no Keep/Reload prompt.
Syncthing conflict handling not implemented No .sync-conflict-* file detection or merging.
File-system watcher not implemented Browser does not live-refresh; it re-scans on navigation.
Alphabetical index sidebar not implemented AlphaIndex element exists but is unused.
Tabs, split view not implemented In-file search IS implemented (spec §2.2); only tabs/split remain deferred.
Files > 50 MB not supported TooLarge state instead.
Desktop / other platforms not supported Android-first.

8. Product requirements (standing)

Requirement Detail
Instant open Files open without visible lag; content and line index load async.
Auto-save Every edit persisted with a 1 s debounce; no save button.
External change awareness Deferred (§7) — the sync-awareness story is explicitly out of v1.
Plain text only No formatting, no highlighting, no file management UI.
Bounded memory 50 MB editable limit with an explicit "too large" state.