2.4.0
·
2568 commits
to main
since this release
Release 2.4
New Features
- Provider model caching + persistence
- New provider models database caching system (chat/image/embedding models cached per endpoint, with
modelTypesupport and updated indexes) - Provider models added to backup/restore, including manifest counts, archive export (
data/provider-models.json), restore warnings/error handling, and restore summary tracking
- New provider models database caching system (chat/image/embedding models cached per endpoint, with
- NPC system upgrades
- New ad-hoc NPC character system (
npcflag, quick-create NPC dialog from chat picker, Settings → NPCs management tab, characters list filtering, and?npc=true|falseAPI filtering) - New “Convert to NPC / Convert to Character” toggle on character view
- New ad-hoc NPC character system (
- System prompts & templates
- Multiple system prompts per character (new edit/view tabs, default prompt selection, prompt templates, sample prompts, and new API endpoints)
- Per-user post-login migration to move legacy system prompt data to the new structure
- Enhanced template display (TemplateDisplay with highlighting for
{{char}}/{{user}}, plus warnings for hard-coded names) - Pseudo-tool support for models without native function calling (memory search / image generation / web search via text markers)
- Chat & turn-system functionality
- Persisted turn state for multi-character chats (+ new PATCH endpoint)
- Stop streaming response button (desktop + mobile) with clean abort handling
- OpenRouter custom model ID support (toggle + custom input + saved preference)
- UI / navigation
- Mobile-responsive upgrades across chat conversation UI, chat footer/tool palette, participant controls, and dashboard
- Navbar restructuring (controls consolidated into user menu + content-width toggle + responsive collapse behavior)
- New About page with project links/tech stack and a simplified footer
Improved Features
- Documentation
- Major refresh across development docs and API docs (expanded to v2.4; new endpoints; all 8 LLM providers; removed outdated JSON-storage references; updated About page notes)
- Streaming, tooling, and provider behavior
- SSE error handling now produces clearer, fuller client-visible errors; reduced noisy logs; better continue-mode handling
- Anthropic streaming now detects tool-use blocks correctly; prompt caching expanded (tool caching, cache strategies, TTL options, cache usage stats surfaced to client)
- OpenAI plugin updated with newer GPT-Image models and deprecation guidance for older ones
- Capabilities report now prefers cached provider models and reads plugin versions from
package.json
- Backup/restore completeness
- Cloud backup/restore now includes user-created prompt templates and roleplay templates, with backward compatibility and correct remapping in new-account restores
- Theming & UI consistency
- Significant enhancements to Earl Grey and Rains themes with broad qt-* utility coverage and improved tab styling
- Tab styling consolidated into qt-tab utility classes; text styling consolidated into qt-text utility classes; reduced redundant button flex utilities
- Avatars refactored into reusable components (Avatar / AvatarStack) and now respect the global avatar style preference
- Codebase maintainability & test coverage
- Major component breakup and TSX refactors to reduce duplication, introduce reusable hooks/components, standardize patterns, and improve logging/error utilities
- Expanded unit + integration test coverage for templates and pseudo-tools (with the suite remaining green)
Bugs Fixed
- Chat streaming / SSE
- Cleaner SSE error handling (including Zod validation formatting) and suppressed benign parse-noise
- Fixed OpenRouter streaming failures involving tool calls (SDK patch for null tool-call IDs)
- Reduced SSE parse errors further (filter empty chunks, skip
[DONE]markers for OpenAI-compatible streams)
- Avatar styling
- Avatar style changes now apply immediately without refresh (context sync)
- Avatar display style now consistently applies across all chat avatars and uses a CSS variable for radius
- Removed duplicate avatar display during “waiting for response”
- Settings / pages / navigation
- Fixed Physical Descriptions tab infinite API loop
- Fixed empty default Connection Profile dropdown on character edit page (mapping corrected)
- Fixed setState-during-render character editing error
- Added missing “Back to NPCs” breadcrumb behavior and correct Settings → NPCs navigation
- Settings page now supports
?tab=URL parameter - Fixed user menu mobile responsiveness issues (layout, truncation, submenu placement)
- Chat composer & message rendering
- Restored sendMessage behavior after refactor (enter key / send button)
- Textarea auto-resize now shrinks correctly when deleting
- Roleplay bracket segments now render correctly with markdown inside
- Removed duplicate attach-file button
- Fixed desktop ToolPalette toggle behavior (no immediate reopen) and centered toolbar icons
- Balanced desktop composer padding
- Multi-character & persona correctness
- Participant sidebar now highlights the correct character during streaming
- Skip button now advances turn correctly and resets the cycle
- Fixed duplicate name prefixes in multi-character messages
- Chat gradual renaming now replaces generic titles correctly
- Persona display-name disambiguation now works with the API’s actual response shape
- System prompts / templates
- Template variables (
{{char}},{{user}}) now reliably process in LLM context building - SystemPromptsEditor buttons no longer accidentally submit parent forms
- Template variables (
- Export / cleanup
- SillyTavern chat export now produces proper JSONL format (and importer can auto-detect JSON vs JSONL)
- Removed obsolete Prisma/PostgreSQL artifacts (project is MongoDB-native)