Implement zcr_extended_text_input_v1 handlers and fix IME backspace holding - #27
Closed
kkimdev wants to merge 40 commits into
Closed
Implement zcr_extended_text_input_v1 handlers and fix IME backspace holding#27kkimdev wants to merge 40 commits into
kkimdev wants to merge 40 commits into
Conversation
…ion and confirm_preedit Map transactional host CJK/Korean IME preedit region changes and confirmation events to v3 delete_surrounding_text, preedit_string, commit_string, and done event sequences.
Track the host's event serials on preedit_string, commit_string, keysym, language, and text_direction, and propagate them in commit_state requests. This ensures bidirectional serialization sync and prevents the host compositor from ignoring text state updates, resolving the IME backspace ghosting/lock bug.
Pass the host-sent serial and time in wl_keyboard::key events generated from on_keysym instead of hardcoding them to 0. This ensures the guest client associates the synthetic key event with the correct user input transaction, fixing blocked/ignored backspace composition clearing bugs.
Track host activation state in TextInputState to avoid null-surface activation crashes. Bind update_host_activation to focus transitions and commit events. Forward host serials through commit_state. Implement CJK backspace holding test.
…, host_surface log
…_msg queue_host_msg/queue_client_msg took &mut Context, which prevented Rust's field-level borrow splitting. This forced the use of 3 helper structs (ActivationAction, CommitActions, PreeditAction) and the extract-then-send pattern throughout the IME bridge. Replace with push_msg(queue, sender, opcode, builder) that takes the specific queue reference. This allows all handlers to queue messages directly inside their get_mut/iter_mut blocks. Removed: - ActivationAction struct (update_host_activation) - CommitActions struct (on_commit) - PreeditAction struct (on_set_preedit_region) - Tuple extract pattern (on_confirm_preedit) Added: use std::os::unix::io::RawFd
Remove .direnv/, _local/, pr_body.local.md, test_e2e.local.ts from tracking — these were accidentally committed local artifacts. Apply rustfmt to changed source files.
…meter; add debug logging
… fix and logging - Removed commit-parameter → commit_string forwarding from on_preedit_string - Removed pending_implicit_commit field and all its uses - Reverted to original behavior where commit parameter is ignored - Kept done_serial as monotonic independent counter - Kept comprehensive log::info! traces for debugging
- Remove commit_serial (field was incremented but never read) - Remove thread_local XKB keymap caching; rebuild keymap per keysym event (keysym events are human-paced, caching adds unnecessary complexity) - Keep all functional fixes: done_serial, host_serial, delete_surrounding_text, activation state machine, keysym serial/time forwarding
The v1→v3 conversion loses the commit parameter from preedit_string. Korean IMEs rely on v1's implicit commit mechanism, which has no v3 equivalent and cannot be replicated without language-specific heuristics.
Logging added to all handlers: - on_keysym: sym char, guest_id, keycode lookup result, forwarding - on_delete_surrounding_text: index/length, before/after conversion - on_set_preedit_region: host_ext_id, index/length - on_confirm_preedit: host_ext_id, cached preedit text - on_language / on_text_direction: serial and values - v3 handlers: set_surrounding_text, set_content_type, set_cursor_rectangle, on_commit (enabled state, host_v1_id, forwarded ops, commit_state serial) - wl_keyboard.on_enter/on_leave: keyboard/surface IDs, seat info, which text_inputs get updated
Logging now covers every event handler in both text_input.rs and keyboard.rs: - on_modifiers_map, on_input_panel_state, on_preedit_styling, on_preedit_cursor, on_cursor_position (v1) - on_clear_grammar_fragments, on_add_grammar_fragment, on_set_autocorrect_range, on_set_virtual_keyboard_occluded_bounds (ext) - on_get_text_input (v3) - on_modifiers, on_release (keyboard)
…_text When the user holds backspace after clearing composition, the IME sends repeated confirm_preedit (ChromeOS extension) events with empty preedit instead of delete_surrounding_text. Our handler was treating these as no-ops, making backspace stop after the preedit is cleared. Now, when cached preedit is empty, confirm_preedit is forwarded as v3 delete_surrounding_text(before=1, after=0) + done, deleting one character per event.
When the IME pulls the last committed character into preedit and deletes it component by component, continued backspace (empty preedit) generates confirm_preedit events. The old fix used before_length=1 unconditionally, which is wrong for Korean UTF-8 (3 bytes per character). Track UTF-8 byte sizes of each committed character in a stack. On confirm_preedit(empty), pop the last tracked size and use it as before_length in delete_surrounding_text. Reset when the guest sends set_surrounding_text. This correctly handles mixed English+Korean text since each character's actual byte length is recorded.
When backspace clears the preedit (non-empty to empty transition tracked via preedit_cleared_for_backspace flag), subsequent confirm_preedit events are the IME asking us to delete committed text. Instead of unreliable delete_surrounding_text (wrong byte counts, guest may ignore it), synthesize wl_keyboard.on_key(KEY_BACKSPACE=14) press+release to the guest keyboard. This works universally: correct byte count for all encodings, works with pre-existing text, and the guest natively handles backspace. During initialization (no preedit clear) confirm_preedit is treated as a no-op (just done), preventing spurious deletions.
Previous approach consumed the flag on first confirm_preedit, so only one KEY_BACKSPACE was synthesized per backspace burst. The IME sends one confirm_preedit per character to delete during hold; we need to synthesize KEY_BACKSPACE for each one. Make the flag sticky: set when preedit is cleared via backspace, and only clear when: - A new non-empty preedit starts (new composition) - on_enter/on_leave (v1 input focus change) - on_enable/on_disable (v3 input activation change) - on_commit_string (commit normalizes state) This way the entire burst of confirm_preedit events during backspace hold correctly synthesizes one KEY_BACKSPACE per event.
The on_commit_string handler sends preedit_string('') + commit_string +
done to properly clear preedit on the guest. The confirm_preedit handler
(non-empty branch) was only sending commit_string + done — missing the
preedit_string('') clear. This caused stale IME tooltips on Enter because
the guest never received an explicit preedit clear before commit_string.
On Space, the IME uses commit_string (v1) which goes through the correct
handler. On Enter, the IME uses confirm_preedit (extension) which goes
through the now-fixed handler.
This stack was maintained by commit_string and cleared by set_surrounding_text, but never consumed — it was dead code from a prior approach (delete_surrounding_text with byte tracking). All backspace handling now uses KEY_BACKSPACE synthesis on the wl_keyboard path, which doesn't need byte-level tracking.
Replace nested if-let blocks with let-else early-exit pattern in on_commit_string, on_delete_surrounding_text, on_set_preedit_region, on_confirm_preedit, on_commit (v3), on_enter, and on_leave. Reduces indentation by one level in each function without changing behavior.
… trace - Extract store_host_serial() helper for the 3 repeated nested if-let patterns in on_keysym, on_language, on_text_direction. - Simplify on_commit trace logging to a single ctx.text_inputs.get(). - Downgrade on_modifiers from info! to trace! (fires per keystroke).
InternetOfTofu
self-requested a review
July 6, 2026 19:19
Author
|
I've split this PR into a stacked series for easier review:
The stack builds on #28, so each diff is much smaller and focused. |
This was referenced Jul 8, 2026
Author
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Reproducible bugs fixed:
Hold Backspace during Korean IME composition → only the last syllable is deleted. Type "가나다라마" and hold Backspace. "마" (the preedit) disappears, but "가나다라" stays and the repeated delete never happens. Root cause: v3 has no event for "delete one character on repeat". This PR adds a heuristic (preedit_cleared_for_backspace flag) that detects backspace bursts and synthesizes wl_keyboard.key events to bridge the gap.
Type "가나다" and press Enter → only "가나" stays, "다" disappears. Root cause: zcr_extended_text_input_v1::set_preedit_region was a no-op. ChromeOS IME calls this on every keystroke to position the preedit within surrounding text. Without it, the cursor is wrong when committing, so the last syllable writes to the wrong offset.
Other issues fixed:
zcr_extended_text_input_v1::confirm_preedit was a no-op. Confirming preedit silently did nothing. Now translates into v3 preedit_string("") + commit_string + done.
done_serial always 0. on_commit_string sent hardcoded 0 for the v3 done serial. Guests that validate serials reject updates.
commit_state always sent serial 0. host_serial from v1 IME events was never propagated to the host's commit_state.
Activation tied only to enable/disable commit. Keyboard enter/leave had no effect — IME could appear active on an unfocused surface or inactive on a focused one. Replaced with update_host_activation state machine.
surrounding_text consumed on every commit. The old code used .take(), losing surrounding text after one forward.
Missing preedit_string("") before commit_string. Without it, some guests keep a stale composition underline after committing.
Implementation: replaced manual wire-format building (~150 lines of extend_from_slice and len<<16) with MessageBuilder + push_msg helper, applied early-exit pattern, added 48 tests.
Testing: