Changelog

中文 · Project

Recorded version changes. Public availability and downloads follow Releases.

What's new in 2.2.7

Fix TXT translation matching and importing in EPUB folder projects, including nested books with the same name and older review pages.

What's new in 2.2.6

Edit AI glossary candidate translations before importing them into the formal glossary.
Keep the chosen warning-review action throughout a multi-file translation workflow, avoiding repeated prompts.
Apply proofreading reports and jump to existing review pages in EPUB folder projects.

What's new in 2.2.5

Fix EPUB review page initialization and multi-file translation startup, including projects that keep source EPUBs and generated files in the same folder.

What's new in 2.2.4

Final translation review now offers a Continue/Skip prompt, so you can choose directly instead of describing the next step in natural language.
Applying review suggestions in single-file mode automatically imports the complete translation into the review page and applies the changes, matching multi-file mode.

What's new in 2.2.3

Keep folder review pages usable after applying proofreading, preserving manual edits and the selected file.
Show unresolved alignment failures and keep repair reads bound to the correct document.

What's new in 2.2.2

Fix proofreading navigation to reuse imported translations in the original HTML and preserve EPUB export when rebuilding a missing page.

What's new in 2.2.1

Improved project path resolution keeps proofreading changes, navigation and exports bound to the same document, including after moving a project.

What's new in 2.2.0

Add text-layer PDF translation: extract prose, formulas, code and figure assets, organize and audit the document structure, then generate review HTML and reflow translated prose with preserved assets into a new PDF. Support complete paragraphs, page continuations and paragraph-local formula reordering while preserving code, formulas and references. See the PDF guide.
Add the official background Agent API, reusing project preparation, translation, proofreading and reference workflows, with status queries, Stop, resume and user replies.
Improve large-project handling: load large histories and validate/import translation candidates in the background with progress and cancellation; optimize session recovery and cross-chunk cache reuse.
Improve editing: activate generated HTML automatically and streamline page opening, translation imports, saving and closing while preserving manual edits.
Reduce false untranslated and terminology findings from English identifiers and character names; let the parent Agent confirm specific untranslated false positives.

What's new in 2.1.9

Fix the single-file mode regression introduced in 2.1.4.

What's new in 2.1.7–2.1.8

Exclude actual regex-preserved matches and correctly aligned glossary pairs from untranslated checks. Record explicitly preserved words as formal source-to-identical-target terms; validate the remaining prose with the same rules for candidate imports and AI review. Glossary matching, terminology conflicts, and recovery rescans also exclude regex-preserved spans.
Automatically reconcile stale review bindings in Host, retain unchanged accepted evidence, and resume work. Fix batch settlement after Stop.

What's new in 2.1.6

Stop when bound required files disappear, preserving drafts and reporting the missing path. Unbound glossary and character tables remain optional.
Fix Stop and project shutdown deadlocks.
Fix Electron HTML closing and task cancellation bindings; failed saves report the error and preserve edits.
Use relative bindings for project settings, HTML and recovery artifacts. Copying a complete project retains parent/child Agent history, drafts and review evidence. External source/reference files must be transferred and selected again.

What's new in 2.1.5

Fix large-file translation freezes with range-scoped chunk checks and background final validation, preserving structural checks, review and recovery.
Apply translation preservation rules and mechanical checks to proofreading suggestions to protect control codes, tags and placeholders.
Fix concurrent recovery: Host verifies and reconciles translations, staging and review evidence before starting Workers. Failed commits preserve recoverable artifacts; unretryable internal errors stop the workflow.

What's new in 2.1.4

Create a project from reference preparation before translating. Clear homepage entries and restore task defaults in one click.
Reference preparation uses editable drafts, checks them, then commits both formal assets together. Interrupted or failed tasks retain their drafts.
One-click setup checks parameters against the source and trials control-token rules, then uses the shared HTML parameters and translation prompt. Generate line-by-line HTML first for hands-on control.
Removed the unused chunking toggle. Folder HTML initializes the file order inside braces and preserves explicit selections.
Configure providers and models at the top of the homepage before opening a project or HTML; settings are shared globally.
Built-in task setup opens HTML and starts translation or proofreading after settings confirmation; reference materials can be organized directly. Defaults: 500 lines and up to 3 agents, with preservation-rule scanning before translation.
Editable HTML character tables and optional proofreading auto-apply to HTML; TXT saving remains manual.
Fixed canceled tasks and compaction starting after Stop.

What's new in 2.1.3

Fixed Grok 4.7 requests incorrectly sending reasoning_effort=none: Auto and legacy Off selections use the official high default; models without known effort controls omit the parameter.
HTML glossary management separates approved terms and AI candidates. Preview, search and delete candidates before import, or delete approved entries individually. Deletion affects only the selected table; older HTML upgrades when opened.
Upgrade the core to 0.99.1 with compatibility updates to prevent regressions.

What's new in 2.1.2

Standard Provider model catalogs can refresh automatically from upstream Pi, with local cache and bundled-catalog fallback when the network is unavailable.
Fixed inconsistent state when resuming parked translation-reuse workflows, retrying proofreading glossary candidates, and restoring typed Workflow intent.
Aligned structured repair capacity with the 500-line model page and merged duplicate length-risk signals for the same line.

What's new from 2.0.5 to 2.1.0

Very large review pages open now: the 80 MiB hard reject is gone, and jumping from a suggestion to the source dropped from about 4s to 1–2s in testing.
Large-file translation and proofreading are steadier: memory no longer climbs without bound, and prescans no longer freeze the main window.
The main window is responsive, and a character can have multiple dialogue lines.
Opening Agent or sending a short prompt no longer yanks the review page to the top.
Mobile remote is a full-screen workspace, with Agent in the tools drawer.

2.0.1 – 2.0.5: Workflow reliability and Providers

Proofreading and EPUB jobs consistently follow the translation explicitly selected on the page. When no file was selected manually, completed translation output becomes the authoritative translated binding instead of a temporary snapshot.
Added Grok (OAuth), model-specific thinking levels, and long-reasoning waits. Expired authorization refreshes and resumes the current work instead of stopping the entire batch.
Exhausted translation review hands exact failed lines back to the parent Agent. The Host refreshes proofreading findings from current bound rows, and accepted results cannot be erased by a later empty write or rejected item.
Longer glossary names cover shorter names, cross-file terminology differences move to final warning review, and disabling glossary candidates also disables their construction, updates, and validation.
Batch apply preserves human edits and TXT-written status. Workflow recovery explicitly distinguishes translation from proofreading to avoid resuming the wrong job.