KOAssistant v0.23.0 Release Notes

Run any action with another model in one long-press, X-Ray marking and book search that no longer stall large books on e-readers, Perplexity on its new API, a new provider and a round of new models, and many fixes.

Highlights

• Run an action once with another model, or switch models from the highlight menu (#86). Long-press any action button for Quick answer, your favorite and recent models, No reasoning, web search on or off. The highlight menu's first KOAssistant button shows your model and switches it.
• X-Ray underlines no longer freeze the reader, and Chapter Appearances and Mentions open at once from a name index kept next to the book.
• Perplexity moves to its new Agent API, with Sonar and other companies' models on one key; the old API was ended on 2026-09-27.
• AI Book Tools search long books without stalling, read short texts whole, and search Japanese, Chinese, Korean, Arabic and other scripts properly.
• X-Ray builds are steadier: they survive the device sleeping, respect hidden sections, skip front matter, resume where they stopped, and you can set a book's X-Ray type (Fiction, Nonfiction, Academic).
• New provider A2Agent (#108) and new models on most providers: Claude Sonnet 5.5, Opus 5.5 and Fable 5.1, GPT-6 and GPT-6.1, Gemini 3.8 Flash, Grok 4.7, GLM-5.3, Kimi K3 and more, with new defaults on Anthropic, OpenRouter, Gemini and xAI.
• Fixes that mattered: backups on e-readers (#110), a Mac-copied install that failed every action (#112), an empty reply closing KOReader (#111), Kimi and regional servers, and updates that check the download before installing.

Models and Run Options

• Run an action once another way. Long-press an action button (highlight menu, dictionary popup, Quick Actions, input dialogs, file browser): the menu opens with Quick answer, up to three models (favorites, then recent picks, then your provider's fast model), No reasoning, with or without web search, and More models… for any model you have a key for. The action runs that one time with your pick; your settings do not change. An action's own settings still hold, but a model you pick beats even a model pinned on the action. Artifact actions (X-Ray, summaries, recaps, the quiz, AI Wiki) carry the pick through their popup to the run you start there; checkpoint builds and automatic updates keep your default model. The chat that opens stays on the picked model, also when you resume it later. Below a double line sit the rows that change the action for good: Add to or Remove from that menu, Edit… and More….
• Favorite models. Long-press a model in any model list to add it to your favorites; ♥ marks them (★ still marks a custom model). Favorites take the first model buttons of every long-press menu and lead the More models… list.
• Send with another model. Long-press Send in an input dialog, the reply window or when starting a chat about an artifact: your model buttons, No reasoning and More models…. The pick applies to that message and the rest of the chat.
• Switch your model from the highlight menu (#86). The highlight menu's first KOAssistant button shows your model; tap it for favorites, recent picks or More models…, and the model changes for everything after while the menu and your selection stay open. It is the Switch model entry of Settings ▸ Menus & Buttons ▸ Highlight Menu Actions; Dictionary Popup Actions can add it to the dictionary popup.
• Your model in the input dialog's gear menu, with the same list; a pick changes your model and the dialog reopens on it.
• Model lists open on the model you use instead of the first page, in Settings and in AI Quick Settings.
• The model list's tier tags follow your own tier picks ("· fast", "· ultrafast").

X-Ray

• Underlines no longer freeze the reader. Passive marking searched the whole book for each name on the reader's thread (seconds per name on an e-reader, minutes for an Arabic name), which could stall pages until Android closed KOReader. Marks are now found on the page you are on, after it settles, at the same cost on any size of book. They land on the right words with soft hyphens and Japanese furigana.
• Chapter Appearances and Mentions open at once. Where each name appears is found once, in the background, and kept next to the book (seconds for a novel, a minute or two for a collected works on an e-reader). New names and aliases are added on their own; a font or margin change indexes again for the new layout. Opened before the index is ready, Chapter Appearances searches right then, with a message you can tap to cancel. A section X-Ray opens on its own pages at once.
• "After 10/25 unseen pages" and "First appearance only" are exact now that they read the name index; until the index is ready, a name is marked once per page.
• One way of counting everywhere. Chapter Appearances counts whole names, as Mentions and the underlines always did ("Ann" no longer counts inside "Annual"). A bracketed part of a name counts as a name only when it is the name in another script ("Tokyo (東京)"), not a label like "(archetype)". A leading "The" is optional while two words remain. Whole-book Mentions no longer stops after the first 5 MB of text. Chinese, Japanese, Korean and Arabic names are underlined and counted like the rest.
• Arabic names no longer match inside other words. They match at the start of a word or after its prefixes (و, ف, ب, ك, س, ل, the article), with at most three letters after; a word's first alef is required, and a name needs three letters. Each book's index is rebuilt once for the rule. The X-Ray instructions also say aliases are names, never pronouns.
• Tapping an underlined name opens its card when the name has a part in brackets, and a name in another script inside the brackets finds it too.
• X-Ray cards say when an entry comes from a section X-Ray ("From the section X-Ray:" and the section's name). "Marking & lookup…" names what marks and lookups read (this book's X-Ray, its section X-Rays, the other books in its group) and is offered for books with only section or group X-Rays too.
• Choose an X-Ray's type. "New X-Ray type" in Book Settings ▸ X-Ray picks Fiction, Nonfiction or Academic (for every book: Settings ▸ Reading & Library ▸ X-Ray ▸ Type of New X-Rays; groups can set it). Auto, the default, lets the AI decide. A set type sends that schema alone, so the request is smaller. The creation form shows the type, and picking a type a book's X-Ray does not have offers a rebuild. Updates keep an X-Ray's type.
• Fiction is no longer typed as nonfiction because of its introduction. The AI judges the work itself (a known title and author, otherwise the main text), never an editor's introduction or a chronology. New X-Rays also see the table of contents of the text they receive. A very small model can still be misled when the first checkpoint ends inside the introduction; "New X-Ray type" sets it right with a rebuild.
• A checkpoint with nothing to build from is skipped. In an annotated edition the first checkpoint may hold only front matter; an empty step is skipped and the next reads further, at most three times and never past half the book, after which the build stops with "nothing to build from so far". Section X-Rays are built whatever the section holds.
• The checkpoint introduction is an introduction again. Since v0.22.1 it came out as a full X-Ray of the first stretch whenever the default categories applied, and the first checkpoint was built on top of it. It covers the premise, the setting and the first characters again.
• Checkpoints in books with hidden sections. Checkpoint positions and section targets now count the visible pages, as your reading position does, so builds no longer fall into text hidden with KOReader's hidden flows, and a step that would send no book text stops the build and says so.
• Checkpoints carry on after the device sleeps. On e-readers where KOReader turns Wi-Fi off for sleep, a request in flight stops, the popup says the build is waiting for Wi-Fi, and the checkpoint is built again seconds after Wi-Fi reconnects. Checkpoints already built keep installing meanwhile.
• A stopped build tells you, and resumes in one tap. The notice stays until you close it, with a Resume button; a step that fails for a passing reason is tried twice first. Resume continues the build that stopped (to your position, to a section's end, or as one request) rather than planning to the end of the book, and is not offered after a build that found nothing to build from.
• A rebuild works like a new X-Ray. It starts with the same introduction, and your current X-Ray stays until that first part is saved, so a rebuild cancelled or failed before then changes nothing. Each step counts only what it saved itself.
• Section X-Rays follow your X-Ray settings (the categories, depth and type of Book Settings ▸ X-Ray or the global X-Ray settings) instead of every category at standard depth.
• A cut-off answer is never passed off as done. An answer cut at its length limit or ended early by the provider is not saved, and you now see it marked as cut off instead of the old saved copy (an X-Ray update no longer opens your old X-Ray as if it had worked, nor a Recap update the old Recap). "X-Ray not saved" says what helps: a lighter depth or fewer categories, updating in checkpoints, or a model that allows longer answers. When the AI declines an X-Ray, you see its answer instead of the X-Ray already on the device.
• X-Ray popups update while open (build progress, the Wi-Fi wait, background updates), and the popup names the checkpoint being built ("The next checkpoint (20%) is being built and installs when you reach it.").
• Smaller fixes: the duplicate-entries question waits for a checkpoint build to finish; Extend and Rebuild stay in the popup after a cancelled automatic build; "Resume building checkpoints" no longer appears while Automatic X-Ray is on; a tap outside the large-request question counts as Cancel instead of leaving the build waiting.

Book Groups and Cross-Book X-Ray (#90)

• One carried entry per character, even when earlier books name them differently (a short name in one volume, the full name in the next): the carried list shows one row, with the other volume's description as a "From earlier books" line. A carried entry's full view opens with its description; the "Not seen in this book yet" note closes it.
• Japanese names keep their middle dot. Models that write "_" or "-" for "・" split one character into two entries; names, aliases and carried entries are repaired when read, and the duplicate review and search treat the spellings as one.
• Underlines no longer vanish when a character is known under two names across a series; each name paints once.
• Look up in X-Ray ignores a stop or bracket the selection dragged along in Japanese or Chinese text.
• Fixed a crash when a search hit a carried entry from a book with a long title; the title is shortened beside the name.

AI Book Tools

• Long books no longer freeze the reader. Search asks KOReader's own document search for the pages worth reading, in a background process a tap on Stop cancels, instead of copying the whole text on every question (minutes, and more memory than an e-reader has, on a 6,000-page collected works). Typo-tolerant matching is gone; the AI retries a spelling instead.
• Short texts are read whole. When the readable text fits (about 64,000 characters at Quick and Standard effort, 128,000 at Thorough) the AI gets it in one piece instead of search snippets, and the usage line reads "Book text read in full". Toggle "AI Book Tools: Read Short Texts Whole" in Settings ▸ Advanced, on by default.
• Better hits. The 12 best hits per query (up to 40 on request), spread over pages, with exact totals; whole-word hits rank above hits inside longer words; sentences holding most words of a longer query are listed below full matches with the missing words named; very common words (decided from the book itself) are not required; a small word index built at the first lookup serves the rest of the request.
• Every script. Japanese, Chinese and Korean queries match literally; Arabic, Hebrew, Cyrillic, Greek and accented words are words (any order, the partial rule, case folded); Chinese, Japanese and Arabic sentence marks end sentences.
• Book text language. New setting "Tell the AI the Book's Language" (Settings ▸ AI Language Settings, off by default): set to "From the book's metadata", book-tool sessions name the language the text is in, so an English question about an Arabic novel gets Arabic searches. Book Settings ▸ Languages and group settings gain a "Book text language" row (Follow global, Off, From metadata, a listed or typed language). The model is also told to write queries in the text's language.
• A smarter table of contents tool. Exact entry counts, narrowing to the top level or to titles containing a phrase (each with its parent chapters), and when the list is too long the chapter levels that fit are kept rather than the first entries. Before its first lookup the model sees a short outline of the contents and its lookup budget in lookups and rounds.
• Honest results. Tool results say what they left out (caps, cut passages, pages out of reach because of spoiler protection or hidden sections), and the model is told to pass that on. An answer given without a lookup says the book was not consulted. Sources and Lookups lists the pages where each query hit, and the filters behind a "toc: 0 entries" line.
• Fewer caveats. Answers no longer end with a paragraph about hit counts and partial matches; a limit is mentioned in one sentence, only when part of the book was out of reach.

Providers and Models

• Perplexity moves to its new Agent API. Perplexity ended the API older versions use (Sonar chat completions) on 2026-09-27; once it is switched off, every Perplexity request on an older version fails. Perplexity now serves Sonar and other companies' models on the same key: the list holds Sonar (the default), GPT-6 Luna, GPT-6 Sol, Claude Sonnet 5.5, Gemini 3.8 Flash and Grok 4.7, and Fetch models lists the rest. With web search on, the model answers through Perplexity's search with numbered citations, as in Perplexity's app; off, it answers without a search. Old Sonar model names move to Sonar with a one-time note. Sources shows only the cited sources; Show Sources lists everything found. Test provider is not offered for Perplexity; Settings ▸ Advanced ▸ Test Connection checks it.
• New provider: A2Agent (#108). A paid gateway to DeepSeek, GLM, Kimi, Qwen and MiniMax behind one key (Tools ▸ KOAssistant ▸ API Keys & Auth, or a2agent in apikeys.lua). Reasoning follows each model's settings; GLM models always reason. Book tools work on its DeepSeek, GLM, Kimi and MiniMax models.
• New models, each tested against its provider's API: Claude Sonnet 5.5, Opus 5.5 and Fable 5.1; GPT-6 Sol, Luna and Astra, and GPT-6.1 Sol; Gemini 3.8 Flash; Grok 4.7; GLM-5.3, 5.3 Flash and 5.3 FlashX; Kimi K3; Qwen 3.8 Max, Flash and 27B; MiniMax M3; DeepSeek V4.1 Flash on Fireworks; and the same models on OpenRouter, Requesty, Perplexity, Groq and Fireworks where they serve them. GPT-6 models reason by default; GPT-6 Astra, GPT-6.1 Sol, Claude Opus 5.5 and GLM-5.3 cannot turn reasoning off, so Minimal and the Quick chip give them their lowest level; Claude Sonnet 5.5's Off sends Anthropic's lowest setting, which does not think on ordinary questions. GPT-6.1 Sol is OpenAI's Flagship-tier model.
• New default models. Anthropic and OpenRouter start on Claude Sonnet 5.5 (same price as Sonnet 5), Gemini on Gemini 3.8 Flash (same price as 3.7 Flash), xAI on Grok 4.7 (cheaper than 4.6), DeepSeek on deepseek-flash (V4 Pro stays in the list and the flagship tier), Together on DeepSeek V4 Pro 0813 (from its catalog, untested). If you never picked a model you move at the next start, with a one-time note; a picked model stays. The tiers point at the new models: Claude Opus 5.5, GPT-6.1 Sol, Grok 4.7 and GLM-5.3 flagship; Claude Fable 5.1 and GPT-6 Astra frontier; Claude Sonnet 5.5 and Gemini 3.8 Flash standard.
• Retired models are replaced. NVIDIA (Nemotron 3 Nano, gpt-oss-120b, MiniMax M3), Groq (Compound, Compound Mini), Fireworks (DeepSeek V4 Pro and V4 Flash 0731, Kimi K2.6, gpt-oss-20b, GLM-5.2), OpenCode Go (GLM-5.1, GLM-5, MiniMax M2.5, Qwen 3.7 Max, 3.6 Plus, 3.5 Plus, Kimi K2.6) and OpenCode Zen (GLM-5, Kimi K2.6 and K2.5, DeepSeek V4 Flash Vision, the free Nemotron) stopped serving these; a saved pick moves to a replacement with a one-time note, and tiered actions use models that answer. No NVIDIA model offers book tools any more.
• Kimi chats work again. Since v0.21.0 kimi-k2.6 refused every chat over its temperature; it now gets temperature 1 while it thinks and none when reasoning is off. Forced values under constraints in custom_models.lua now apply to every provider.
• Z.AI, Qwen and Kimi use the region you set. The Region setting had no effect, so keys from the China or US platforms were refused. If one of these stops working after this update, set its region to the platform that issued your key.
• X-Rays work on Cohere, and so do Command R7B and Command A: KOAssistant knows each model's output limit and retries once at the limit Cohere names.
• Reasoning fixes. Grok 4.5 and Grok 4.20 Reasoning at Minimal work (the lowest level, or no setting on a model that always reasons); turning reasoning off works on Fireworks (GLM-5.3 and gpt-oss get their lowest level there); Claude Fable 5, Grok 4.6 and gpt-oss through OpenRouter, and Claude Fable 5, Sonnet 5.5 and Opus 5.5 through Requesty, get their lowest level at Minimal instead of a refused off; reasoning from models that end their thinking with a bare </think> (Qwen3 Thinking, DeepSeek R1 on some hosts) goes to the reasoning view instead of the answer.
• Book tools work on Claude Sonnet 5.5, Opus 5.5 and Fable 5.1, on Anthropic and through OpenRouter: they refuse a forced lookup, so they are asked without the force.
• Test provider tests the requests KOAssistant really sends. Each step is built like a real request (route, region, thinking off for tool steps), so it no longer fails where nothing is wrong, and a failed step shows the provider's explanation. It is offered for OpenAI-style providers; the others have Settings ▸ Advanced ▸ Test Connection.
• Also: Qwen at temperature 2.0 works (1.9 is sent); longer answers on NVIDIA (16,384 tokens) and on OpenRouter's GPT-6 (32,768); custom providers and hand-typed model ids ignore stray spaces once saved again.

Chat Windows

• Dictionary lookups can show the word in its sentence. Settings ▸ Dictionary Settings ▸ Show Excerpt (off by default, also on the gear menu of dictionary windows) opens the answer with a short excerpt around the word, the word in bold. It needs Ctx on; nothing extra is sent.
• The compact dictionary popup grows to fit a long answer, from 60% of the screen up to 80%, and rotating the device no longer snaps it to full height. Its Expand button is now → Chat, one tap to the full chat; actions set to the full-size dictionary view (Deep Analysis) still open in it.
• Translate and artifact windows fit their text as laid out, without a band of empty space at the bottom.
• A long Quick Explain answer opens the normal chat window instead of the compact dictionary window.
• The dictionary bypass works in artifact windows (a book's About from the file browser, AI Wiki entries, pinned answers, X-Ray versions).
• Smaller fixes: no reasoning line in dictionary windows (Show Reasoning still has it); the response window names the action while an answer streams; the reply window's ⚡ chip reads ON or OFF instead of "SET".

Actions and Menus

• Book passages and an attached notebook go only where your privacy settings allow. Both are now checked against the provider that receives the request (the Quick preset's model, or a one-run pick), and left out with a notice where book text or notebook sharing is not allowed for it.
• X-Ray (Simple) is a button in the input dialog of a book that is not open (a long-press in the file browser); it was the one action behind Show More Actions… there.
• Quick Settings and Quick Actions are the first rows of Tools ▸ KOAssistant, for readers without a gesture for them.
• The "(KOA)" mark can be turned off per surface in Settings ▸ Menus & Buttons (highlight menu, dictionary popup, file browser).
• Fixes: switching actions in the compact window no longer keeps the last action's model; the action wizard no longer leaves a second settings screen behind; "+ Dict. Popup" on an action's details screen says "Added"; Generate Image no longer shows the web search icon; dictionary lookups in PDF and DjVu get the plain word (no trailing punctuation or "'s"), as KOReader's dictionary does.

Backups and Updates

• Backups work on e-readers again (#110). "Create Backup" reported success with 0 B and an empty folder on every e-reader; backups are written for real, a failing backup says so, and a restore makes its restore point again.
• A failed restore can undo itself again; the rollback no longer waits on a lock the restore holds.
• Safer updates. Update Now checks the download against the release's size and checksum, so a cut or damaged download never replaces your plugin; slow Wi-Fi gets up to 10 minutes; your files are checked before anything is deleted, and a file that cannot be carried over keeps the old folder and tells you where it is; a linked domains/ or behaviors/ folder keeps its files; a git checkout is never replaced. These protect the update after this one: an update always runs the updater you already have.
• No more "Don't warn again" on the large-text and cut-text warnings; they offer Cancel or Continue, and Settings ▸ Privacy & Data ▸ Text Extraction turns one off.

Other Fixes

• Actions no longer fail with "there was an error parsing the body" (#112). A plugin copied from a Mac carries hidden companion files next to every file; one next to a behavior or domain file could be loaded as the AI's instructions, and the provider then refused every request (only Dictionary, Quick Definition and Translate worked). Those files are ignored, a damaged behavior or domain falls back to the built-in one, and every request is checked and repaired before it is sent. The same fix covers text cut mid-character in Japanese, Chinese, Korean and other multi-byte books, for X-Ray chapters, book-tool results and the long-text notice.
• An empty reply from a provider no longer closes KOReader (#111). You get "Empty response from . Please try again."; a reply the plugin cannot read ends in a message too; a checkpoint run tries that checkpoint again, up to twice, before stopping.
• A tap outside a step goes back one step, everywhere a screen closes to open the next: the X-Ray creation form's lists, checkpoint and version cards, section X-Ray options, Marking & lookup, Merge from another book, the duplicates list, the alias picker, a group's "Move to position", a book's Groups screen, the Quiz numbers, the action wizard and editors. A tap outside does what Back or Cancel does.
• The note about a changed model names KOAssistant, so it no longer appears out of nowhere.
• Translations. This release's 94 new strings are machine-translated in all 26 languages (marked for review on Weblate), and labels keep their edge spaces: prefixes such as "Provider:" and suffixes such as "(book)" had lost the space that joins them to a value, so a German screen read "Anbieter:Anthropic". About 50 labels per language are repaired, and the translation tooling keeps the spaces from now on.

How You Can Help

• Device reports.
• Translations: review passes on Weblate.
• Bug reports and feature requests.