Repository navigation
·
274 commits
to release
since this release
What's Changed
Multi-Agent Orchestrator
- Added two built-in director presets: Default (Memory Graph + Search) and Default (No Memory Graph, No Search).
- Agent preset pickers can import the bundled agent presets in one click.
- Skills can now be bound to orchestration presets, and the binding survives renames, deletions, and exports.
- Skill tools are available only to orchestration agents, not the main chat.
- Custom tools can now be authored and test-run inside the iteration studio.
- The orchestration preset dropdown can directly select global or per-character agent presets.
- Each preset can define a world info filter that applies wherever agents read the book.
- Loop, spec, and agenda agents can force-activate dormant lorebook entries.
- Runtime agents gain world book browsing tools to list visible books and inspect entry indexes and contents.
- Spec and agenda agents now read Open Notes.
- User regex now applies to what agents read and what they output.
- Raised agent round limits and removed hidden upper bounds from configurable settings.
- Reworked the orchestrator settings drawer into Agents, Tools & Skills, and General tabs, and all changes save automatically.
- The run panel shows per-agent timers.
- Finished rounds collapse automatically when a run ends.
- Improved the simulation review experience.
- Concurrent sub-agent dispatches wait for the first streamed chunk to reuse the upstream prompt cache.
- Spec preset cards show which nodes bind them.
- Draft editing is now a per-agent tool permission, enabled on the main agent by default.
- Switching the director preset no longer overwrites the user's chat completion preset files.
- Deleting a chat completion preset clears references to it from orchestration presets.
- Custom tools embedded in an imported character card are reviewed before they are registered.
- Editing spec and agenda presets now keeps tools, skills, and default-tool state.
- The Open Notes panel refreshes as agents write and reports failed writes.
- The running indicator no longer covers the toolbar.
- Stopping a run now halts immediately instead of waiting for the current round to end.
- Fixed stale preset selections after creating or renaming presets; sub-agents deleted from a character card are no longer brought back by the global config.
Memory Graph
- Added a graph revision studio where the AI proposes changes as reviewable diff cards with rollback.
- Added thread nodes that track long-running plotlines.
- Extraction gains a crawl mode: it first explores the graph structure and starts from a rough outline, saving prompt tokens.
- Simplified recall to LLM and RAG modes, with optional rerank and query rewrite.
- Recall is now balanced across node types, so one dominant type no longer crowds out the rest.
- Added a recency horizon for always-injected nodes; nodes beyond it fall back to normal recall.
- Setting the recall query window to zero now uses only the last user message.
- Removed the length cap on RAG recall queries.
- Event extraction now separates what a character lived through from a mere mention, and merges symmetric relations into a single edge.
- Event extraction scales detail with hierarchy depth and classifies psychological state and NPC baselines.
- Graph editor saves and compressions anchor at the floor they actually cover; a failed batch edit no longer applies the rest of its batch.
- Vector search returns concrete events only, without rollup summaries.
- User regex now applies to the text extraction reads.
- Reorganized the settings drawer into Recall, Extract, Graph, and Advanced tabs, with field help throughout.
- Fixed empty replays wiping the graph, stale vectors leaking after a chat reset, and read failures going unreported; deleted event summaries no longer leak into the world info scan.
- Swipe regenerations reuse the recall snapshot from the same turn.
- Swiping the opening greeting no longer invalidates the memory graph.
Character Card Editor (CardApp Studio)
- Fixed a batch of card-editing reliability issues: applied edits no longer roll back on their own, the rollback button works again, and duplicate diff cards are gone.
- Replacing a character card now offers world book choices: import the new card's embedded book, keep the previous book, or merge the two in the editor.
Iteration Studio
- Sessions in all four studios now persist and replay tool calls and results across reloads.
- Tool calls render as diffs, and message cards show the assistant's thinking.
- Regenerating a step now undoes its already-applied edits first, and every user message gains an Edit button.
- The AI now reads only the fields it needs instead of reading everything up front, and failed edits come back with a clear diagnosis.
- Failed tool calls now return the real error and reason instead of silently doing nothing.
- Session histories moved to per-character storage, and closing a popup flushes pending changes immediately.
Presets & Prompt Manager
- A card can now bind multiple chat completion presets, with a default and a clear-all control.
- Prompts can attach to an existing message instead of only relative positions.
- Preset-scoped world info now survives export and import.
- Prompt entries can be dragged into existing groups, and sub-group rendering is fixed.
- Plugin prompt structure is reworked for prompt-cache reuse, making repeated requests faster and cheaper.
- Saving an orchestration profile to a card now offers to embed the chat completion presets it references.
- Clearing card-bound presets now offers to save each snapshot to the global library first.
- Card-bound presets can be renamed and no longer depend on a stale global name.
- Third-party extensions reading the preset list now see the active card-bound preset.
- Fixed stale sampler and prompt values left behind after switching presets.
- Prompt Manager reports render failures and falls back to the identifier for entries without a name.
- Edits to a card-bound preset are saved back to the card instead of being lost.
- Deleting a preset now purges its associated state everywhere.
Connections & Models
- Each connection profile can now set its own request timeout and retry policy.
- Connection profiles can now auto-continue when a response is truncated, with a configurable attempt limit.
- Added an OpenAI Responses chat completion source.
- Custom models are now saved per connection profile, so switching profiles no longer mixes them.
- Kimi gains partial prefill and correct reasoning forwarding.
- Added support for the newer Claude 5 and Opus 4 models, and dropped obsolete beta flags that blocked Bedrock.
- Advanced request settings fold into a collapsible drawer.
- Google AI Studio now retries when a request is blocked by safety filters instead of failing, and Vertex model listing is fixed.
- OpenRouter adds a Gemini history cache, with a configurable number of recent turns left uncached.
- The Custom source now forwards reasoning effort and omits the parameter when set to Auto.
- The connection manager warns when Claude prompt caching is combined with squashing post-processing.
- Fixed the model picker selecting a wrong model.
- Fixed profile switching overwriting the model field.
- Fixed "Save and Update" skipping the profile write when the exclude list was unchanged, which dropped additional parameters.
- Fixed /model with no argument not reading the input box.
Generation & Streaming
- Unified the generation pipeline for chat completions, text completions, NovelAI, Kobold, and image generation, fixing generation under Bun.
- Claude, DeepSeek, and OpenRouter thinking content now persists and replays across turns.
- Stop requests reach providers immediately, so a ComfyUI image job can be interrupted mid-run.
- Token counting and token encoding are configured independently.
- Recovered generations reconnect to the live stream.
- Token counting moved off the main thread so long prompts no longer freeze the UI, and more tokenizers are bundled locally.
- Upstream errors now arrive as proper error responses instead of hanging streams, and error text is no longer truncated mid-message.
- Empty responses are retried automatically.
- Hiding a message now hides its tool call records as well.
- Unknown models fall back to client-side tiktoken tokenization.
- WebSocket connections reconnect automatically after a drop.
- Fixed WebSocket connections failing on https pages and inside iframes.
- Fixed aborted requests being rejected by request validation and leaving the underlying stream open.
- Fixed third-party response wrappers not applying to non-streaming requests.
- Fixed the retry prompt still appearing after a user aborts a request.
- Fixed Claude requests failing due to whitespace-only message blocks in converted history.
- Fixed DeepSeek requests failing when tool_choice was sent in thinking mode.
- Fixed Gemini force-merging consecutive same-role messages.
Chat & Characters
- Replacing a character card now preserves its local bindings, asking per category when the new card conflicts.
- Card replacement now shows a structured full-screen diff overview.
- Chats can be merged in a chosen order or split at chosen points.
- Streaming token usage is displayed and normalized across providers.
- Chat search now covers older messages that are not loaded yet.
- Pending chat saves now flush on page unload, save conflicts recover automatically, and save failures show the server error.
- Chat file names no longer accumulate duplicate extensions.
- Rapid character switching no longer writes to the wrong chat.
- Fixed the reasoning editor being left behind when committing or cancelling a message edit.
- Character saves surface real errors, and embedded lorebooks survive edits and world deletion.
- The open chat stays selected across card replacement, character reloads, and unshallowing a character.
- Opening a recent chat no longer creates a phantom empty chat file.
- The Manage chat files popup highlights the currently open chat.
- Replacing a card with the same file again now works after a failed replacement.
- New group chats no longer hit a save conflict on the first message.
- First-run onboarding no longer blocks characters and group chats from loading.
World Info
- Reworked mobile World Info editing with a cleaner layout that survives viewport switches.
- Lorebooks embedded in a card stay bound across reloads, and deleted entries stop leaking into exports.
- World books with array-form entries import and edit correctly.
- Exporting a card whose bound world book cannot be embedded now shows a warning.
- Deleting a missing world book no longer errors.
Request Inspector
- Request and response views now display reasoning content, thinking blocks, and signatures.
- The upstream native finish reason is displayed alongside the normalized one.
- Vector embedding and rerank requests are recorded.
- Non-streaming requests and early failures are recorded, and streamed responses are inspected chunk by chunk.
- Added a configurable retention window, with expired records cleaned up automatically.
- Fixed HTTP 200 responses with an error body not being marked as failed.
- The inspector and streaming usage stats now cover OpenAI Responses instructions, input, function calls, images, and reasoning tokens.
Storage & Sync
- Added SQLite, MySQL, and PostgreSQL storage engines with an admin migration panel and resumable migrations.
- The Backup Manager now converts backups between storage modes.
- Added a storage inspector with subdirectory browsing.
- Restore archives upload in chunks with retry and resume.
- Added LAN Sync: two Luker instances on the same network can pair, sync incrementally by category, resolve conflicts file by file, and undo the last sync.
- LAN Sync saves peer credentials and warns when a pairing link belongs to a different account.
- Migrations between file-system and database backends preserve chat integrity and timestamps.
- Deleting a chat or character now offers to delete associated media as well; character asset folders are deleted with the card.
- Unsafe file names are now rejected before being written to storage.
- The Backup Manager groups its data-category selection into its own section that applies to download, restore, and migration links.
TTS
- NPC dialogue lines are now attributed to their speaking character, with per-quote playback.
Search Tools
- The pre-request agent no longer collides with world info operations.
- Injection depth and role fields now hide when the entry position is not At Chat Depth.
- Search snapshot write failures now report the reason instead of being silently dropped.
UI & Mobile
- Android gains a crash diagnostics bundle, and a crash loop disables all third-party extensions into safe mode.
- Android adds a debug recording toggle that writes native events and logs into crash reports, with a copyable diagnostics snapshot in the endpoint dialog.
- Android runtime and endpoint notifications gain a reload button.
- Android cold start is faster, and debug builds can upgrade in place over release builds.
- Chat export now uses native download on Android.
- The iOS keyboard no longer breaks the viewport layout.
- Fixed Android viewport-height and full-width layout quirks.
Authentication & Users
- New OAuth accounts start with the provider profile picture, OAuth-only accounts reject password login, and debug export details are admin-only.
- Fixed the default avatar pointing at a nonexistent file.
Internationalization
- The lazy-load media setting, log viewer time filters, and the Swipe picker now ship with Simplified and Traditional Chinese translations.
- Fixed the admin panel treating quota and OAuth status values as translatable text.
Platform
- Now tracks SillyTavern 1.19.0.
Completion Preset Assistant
- The iteration workbench can import an existing chat to work from.
- Edit conflicts and already-applied operations are now reported.
Background Keep-Alive
- Added a mobile keep-alive toggle using picture-in-picture or audio so off-screen generations keep running, with audio playing only while generating.
Extension API
- Added character state read, patch, and batch-get helpers, exposed saveChatDebounced, and added lookup APIs for recall results and inline memory-graph UI.
- Character replacement world book decisions now run in the core replace flow.
- Checkpoint creation now emits a branch-created event so plugins can copy chat-bound state.
- Fixed chat metadata assignment and chat-scoped variable persistence.
- Floor state now exposes a log size accessor.
New Contributors
- @hershalakenya519-arch made their first contribution in #17
- @KronosXup made their first contribution in #27
- @Illustar0 made their first contribution in #25
- @Bobpage-sys made their first contribution in #31
- @jojo552 made their first contribution in #32
- @liushuangls made their first contribution in #33
- @ZZZdragondYNGPHX made their first contribution in #37
- @chieftain4201 made their first contribution in #42
Full Changelog: v2.7.0...v2.8.0