v1.19
This release centers on the result viewer: text layout controls, custom
fonts, file manager integration, and a reorganized settings menu.
Result viewer
- The result viewer is now a chat transcript: user turns and answers share
one scrollable conversation. Bubbles size by turn length: one-liners keep
a compact shape, long selections take the full page, and caption-only
action labels hug the right edge instead of reading as a wide empty box. - Recursive queries: select text in any answer and run another prompt on it;
viewers stack so each follow-up keeps its own context. The acted-on
selection can move out of the caption into its own muted band above the
question (Response Settings > Show Highlighted Text), so the original is
easy to compare with the follow-up answer. - Turns show book identity (title/author) in the bubble; minimalist mode
keeps only the answer actions for a quieter layout. - Streaming dialog shows a reasoning/answer status row, and the reasoning
switch now applies to turns already on screen. - Mixed-direction replies render per block (code stays LTR), and a Response
Font picker lists installed KOReader fonts and registers the choice with
MuPDF.
Settings menu
- Response Settings is reorganized around display: a Text Appearance submenu
groups Text Size, Text Direction (Auto/RTL/LTR), and Response Font; the
old RTL checkbox in the language dialog is gone. - New Minimalist Mode master switch shows the answer alone and takes over
Reasoning Text and Follow-up Questions while active. - New Show Highlighted Text switch gives the acted-on selection its own
block above the question. - Dictionary settings gain Auto Add Word to Vocabulary Builder.
File browser
- Book Info (AI) and Recap (AI) run directly from the file manager
long-press menu on unopened books (FileManager, History, Collections,
File Searcher), no need to open the book first.
Dictionary and Vocabulary Builder
- Dictionary lookups can auto-add to the Vocabulary Builder; the builder
button is always offered and a failed auto-add is reported on both paths.
Safety
- An unrecognized web-search key now fails closed instead of silently
enabling search; each search tool stores only the credential it declares.
Fixes
- Default prompts include KOReader context.
- Web search is preserved on feature follow-ups.
- Recap reads last-read time from history (with sidecar mtime fallback)
instead of unreliable file atime. - Model picker no longer shows a false selection after paging; the staged
pick survives page changes. - API connection tests use a larger token budget so reasoning models return
a usable answer. - Text direction label refreshes without closing the options menu.
- Search keyword line is flush left.
l10n
- The localization pipeline no longer ships wrong-language or misplaced
strings (some locales carried text in the wrong language); wordless
markup-only templates are out of the msgids. Catalogues refreshed in this
cycle.