UI: UI refactor - #362
Merged
Merged
Conversation
Assisted-by: Trae + MiniMax-M3
Address three follow-up items from code review of the LLM Setting refactor (93a913d): 1. .gitignore: clarify the blanket *.png rule with a comment explaining the convention (screenshots / one-off artifacts go in out/, which is already gitignored at the top of the file). The rule itself is unchanged so this is documentation-only. 2. LLMSettingFragment: document the "heal on access" side effect of preference.setValue() inside onCreatePreferences(). The previous "6" default caused first-launch users to NPE in KANTVAIModelMgr.getModelName(); the silent overwrite is intentional but should be loud in the code so the next maintainer doesn't strip it as dead code. 3. KANTVAIModel: rewrite getModalityTag() using List + String.join. The previous StringBuilder+boolean-first dance was hard to extend to a fourth modality (e.g. video) without subtle ordering bugs. The new version scales by adding one line per modality. Assisted-by: Trae + MiniMax-M3
…rkdown
The old AI page was a single big TextView under a long row of spinners.
It was hard to follow long inference outputs (no clear turn boundaries)
and the settings surface was overwhelming on first launch.
Layout:
- RecyclerView of user / assistant bubbles (ChatAdapter, ChatMessage,
bubble drawables, item_chat_user.xml, item_chat_assistant.xml)
- Sticky input row (IMG / AUD / prompt / Send / CLR) at the bottom
- Bench button on the top bar opens a single settings dialog,
replacing the old 4-spinner row + 3-dot overflow
UX fixes bundled with the refactor:
- Default bench type is LLM on first launch (was ASR, which made
the app send the user to a broken state when no ASR model was
present)
- ASR mode auto-fills the prompt with "help to transcribe the audio
file"; switching back to LLM clears it back to "Ask AI..."
- Pre-flight check for ASR subsystem init: toast a clear error
instead of the cryptic "asr instance not initialized" string
- Backend status (NPU vs CPU) is colour-coded in the top info bar
and refreshed from Settings on onResume, not from a stale field
Streaming improvements:
- 80ms throttle on the event handler so we re-bind at most ~12.5 fps
- UI thread hop for native events (the EventListener callback was
firing on the workThread, racing the UI thread's onBindViewHolder
and causing visible flicker)
- Markwon rendering applies to STREAMING rows too, not just
COMPLETE, so **bold** / lists / headings appear as tokens arrive
instead of the whole row suddenly reformatting at the end
Docs:
- Move known issues out of the model table into a new "Known
issues" section in the top-level README, so GitHub users see
them at a glance
Cleanup:
- Remove llInfoLayout (the unused intermediate view) and the
associated displayImage() / ivInfo plumbing
- Drop the 3-dot overflow + PopupMenu that was over-engineered for
a single "Bench" item
Assisted-by: Trae + MiniMax-M3
- vendored markwon due to network access issue
- pre-strip `^#{1,6}\s+` in ChatAdapter before passing text to Markwon,
since Markwon renders `##` as 2x-sized bold which breaks chat layout;
the stripping is gated by a `debug` field (0=off, 1=on) for release
- when debug=1, ChatAdapter also writes streaming snapshots to
/sdcard/Android/data/.../chat_dumps/ (text+spans state.txt +
per-chunk PNG) for offline debugging of future regressions
Assisted-by: Trae + MiniMax-M3
Previously, AIResearchFragment.onSend() fell back to strDefaultPrompt
("introduce the movie Once Upon a Time in America") whenever the user
hit Send with an empty input box. That prompt is meaningful for LLM
inference but is nonsense for ASR (whisper can't act on a movie plot
request), so an empty ASR input silently used the wrong default and
then surfaced a transcription failure.
Match the ASR default to the string the Bench dialog already auto-fills
("help to transcribe the audio file") so empty ASR inputs are
self-consistent: bench dialog auto-fill = empty-input fallback.
Assisted-by: Trae + MiniMax-M3
During LLM streaming, the chat adapter's 5Hz markwon re-render was
the only path updating the assistant TextView. Between re-renders
(up to 200ms) the TextView did not refresh, so the user saw no
growth on chunks that arrived after the row was bound. The next
re-render then jumped to a freshly-formatted Spannable, which
looked like "no markdown until inference ends".
Split the per-chunk update into two stages:
1. appendRawChunk() runs on every chunk arrival and writes the
raw text straight into the TextView's underlying buffer
(SpannableStringBuilder / Spannable / plain CharSequence)
so new characters appear instantly.
2. maybePeriodicFullRender() is the existing 5Hz markwon
full re-render, which catches cross-chunk markdown
(e.g. `**` / `bold` / `**` split across three chunks).
Heuristic for the LLM's unclosed-bold output ("**总结:" with no
matching close): preprocessUnclosedBold() runs on the text that
goes to markwon (NOT the data model), splits on \n\n, and only
repairs paragraphs with an ODD `**` count. This preserves valid
commonmark cross-line bold like `**xxx\nyyy**` which a naive
line-level pass would break.
Applied at both bind(COMPLETE) and maybePeriodicFullRender().
Assisted-by: Trae + MiniMax-M3
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.
Overview
UI refactor:#360
AI assistance disclosure
AI assistance was utilized during development. However, this is not an unreviewed dump of AI-generated code or documentation. All testing was performed manually by JZ.
Assisted-by: Trae + MiniMax-M3