Remove mammoth dependency and implement lightweight DOCX text extraction using fflate + DOMParser. Extracts text from body, headers/footers, and footnotes/endnotes by parsing w:p/w:t elements. Make readDocument() synchronous. Add PDF.js type definitions to eliminate unsafe any access.
Switch PDF extraction to Obsidian's bundled PDF.js via loadPdfJs() instead of unpdf. Replace officeparser with mammoth for DOCX and custom fflate-based ZIP parsers for PPTX/XLSX/ODF formats. Eliminates heavy dependencies, removes all dynamic eval and .wasm references, and simplifies the esbuild plugin configuration.
Adjust EXPECTED_NEW_FUNCTION_COUNT from 6 to 9 to account for unpdf >=1.6
vendoring PDF.js twice (main + worker/legacy), duplicating the 5 feature-
detection probes and 2 PostScript JIT sites (Hogan pair unchanged).
Extend WASM neutralisation to handle 4 total .wasm literals (qcms_bg.wasm ×2,
jbig2.wasm, openjpeg.wasm) instead of just qcms_bg.wasm, moving from single-
reference constant to EXPECTED_WASM_REFS map with per-filename counts and
unknown-literal detection.
Harden officeparser CJS shim regex to capture all minified identifiers (outer
var, arrow param, getter params, function param) via backreferences rather than
hardcoding param names, fixing breakage across esbuild releases (v7.1.0→v7.2.1
changed closure params r/t/e→n/e/t).
Remove runtime asset loading from AssetsService by importing assets directly and configuring esbuild loaders for SVG (text) and PNG (dataurl). This ensures assets work in release builds where only main.js is shipped without the Assets folder.
Extends the build neutralisation plugin to rewrite the dead qcms_bg.wasm
reference from unpdf's vendored PDF.js. The reference is unreachable (code
path never entered, loader gated, file not shipped) but trips Obsidian's
.wasm scanner. Build now fails if .wasm reference count drifts, preventing
undocumented changes from slipping through.
Add explicit instructions across all agent prompts explaining that binary
files (PDFs, images, documents) return content as attachments in the
message following the tool result, not as text in the result itself.
Update tool response messages to clearly state the attachment delivery
mechanism and prevent agents from re-reading the same file expecting
different output.
Improve build safety by neutralizing dynamic eval constructs and
hardening the officeparser plugin against bundle shape changes.
- Change ChatMode enum values from numeric (0,1,2) to string literals ("read_only", "edit", "planning")
- Move chat mode state management from ChatWindow to SettingsService with persistence
- Add chat mode awareness to system prompts and tool validation
- Replace Map with WeakMap for settings subscribers to prevent memory leaks
- Add type-safe event handlers to EventService
- Update esbuild and TypeScript targets from ES2018/ES2020 to ES2022
- Add requiresEditModeEnabled() validation to prevent tool hallucination in read-only mode
- Update all test fixtures to include chatMode in mock settings
- Calculate height for consecutive assistant messages instead of just last message
- Track message role in ChatArea for proper height calculation
- Update MultiAgentIntegration test to use ChatMode enum
- Update officeparser browser bundle path and shim for v6
Prevent esbuild from resolving dynamic require() calls in officeparser's
bundled code by replacing its require shim with a stub that throws and
stripping eval("require")("stream") calls. This ensures the bundle remains
self-contained and prevents "fs" module errors on Obsidian mobile.
Replace hardcoded "The contents of the file" strings across all AI provider classes (Claude, Gemini, Mistral, OpenAI) with centralized Copy.AttachedFile enum value. Update SystemPrompt to document attached file behavior and clarify files are not in vault. Fix officeparser import to use named export and add esbuild plugin to handle browser bundle's IIFE format for mobile compatibility.
Alias officeparser to its browser bundle to avoid Node.js "fs" module dependency that crashes on Obsidian mobile. The browser bundle provides identical functionality using web APIs.
Introduce a new AbortService to centralize cancellation logic across all async operations, replacing scattered AbortSignal parameters with a unified singleton service. This improves maintainability and provides consistent cancellation behavior throughout the application.
Key changes:
- Add AbortService for centralized abort signal management with automatic cleanup
- Refactor all AI providers (Claude, Gemini, OpenAI) to use AbortService instead of passing AbortSignal parameters
- Update streaming operations to use centralized abort handling
- Add CancellationIndicator component to show visual feedback during operation cancellation
- Rename ChatAreaThought to ThoughtIndicator for better semantic clarity
- Add Environment enum for consistent environment detection
- Enhance ChatService lifecycle with proper cancellation state management
- Remove scattered abort-related UI selectors and error messages in favor of dedicated indicator
- Add safeContinue() factory method to ConversationContent for internal continuations
- Update all tests to reflect new abort handling architecture
This change simplifies the API surface by removing AbortSignal parameters from method signatures while improving the user experience with clearer cancellation feedback.
Implement a comprehensive diff viewing system that allows users to review and approve/reject file changes before they're applied. The system includes event-driven architecture for managing diff lifecycle and integrates diff2html for rich visual diffs.
Key changes:
- Add DiffService for managing diff approval workflow with accept/reject/suggest actions
- Create EventService for type-safe event handling (DiffOpened/DiffClosed)
- Add DiffView component with diff2html integration for visual diff rendering
- Modify VaultService to propose changes and require confirmation before file operations
- Update FileSystemService to support optional confirmation for write operations
- Add Event enum for centralized event type definitions
- Import custom styles and diff2html styles for proper diff rendering
- Update ConversationContent validation to support optional toolId field
- Remove FileManager from dependency injection (now accessed directly from app)
- Update .gitignore to track styles.css instead of main.css
- Add mobile-specific height constraints for input buttons
- Fix message alignment to left for better readability
- Change click to mousedown for better mobile responsiveness
- Adjust top bar grid columns for mobile layout
- Update view type identifier to 'ai-agent-main-view'
- Standardize quote style to double quotes across config files
- Deleted highlight-default.min.css, katex.min.css, and markdown.css as they are no longer needed.
- Implemented StreamingMarkdownService for improved markdown processing with support for math and syntax highlighting.
- Added StreamingService for handling streaming requests to the Gemini API with error handling and chunk parsing.
- Introduced styles_old.css for enhanced code block styling and better readability.
- Updated package.json to include dependencies for rehype and remark plugins for markdown processing.
- Added default highlight.js CSS for code highlighting.
- Included KaTeX CSS for rendering mathematical expressions.
- Created markdown.css for styling markdown content in the Obsidian plugin, incorporating styles for headings, paragraphs, lists, tables, and code blocks.
- Customized highlight.js colors to match Obsidian theme.
