Layla

- Fixed some functions hanging during background execution in task.js. - Fixed a small OpenCL GPU performance regression.

Layla 7.5.0 is the full public release of the 7.5 line. NEW - Character workspaces: give each character a dedicated private workspace - a filesystem its agents can read, write, update, and delete files in. Attach the Character Workspace and Workspace Shell agents to enable it, and manage files from the Create Character screen or Chat Info, including importing files and ZIP bundles and downloading a backup. - Filesystem-based memory: characters can keep files like system_prompt.md, Memory.MD, and Soul.MD in their workspace. The agent reads system-prompt files at the start of a conversation to guide what it reads and updates. - Automatic agent triggering: enabled agents are added to the system prompt and the model decides when to activate them, instead of relying on trigger phrases. Layla also handles models that call two or more tools in one turn. - Slash commands for agents: type / in the message field to list the agents available to a character, then pick one to trigger it directly. - Layla Music: a new mini-app for the music you generate. Browse the latest releases from other users, generate music, collect tracks into albums, set cover art, create artists, and import your own audio. - Experimental Streaming MoE: run mixture-of-experts models larger than your device's RAM by streaming expert weights from storage. - Partial and complete backups: back up and restore individual sections, and optionally include custom mini-apps, model files, and inference engines. - Apple App Intents (experimental): Layla now exposes App Intents so Siri shortcuts can trigger Layla as part of a shortcut. This is an early implementation and is expected to change. - Layla SDK v7.5: the OpenAI-compatible endpoint now supports tool calls, with new notification-scheduling and Layla Cloud login APIs for mini-apps. IMPROVED - Layla now uses the chat template built into each model, so there is no need to set a system prompt format per model. Existing system prompts still work. - Multiple images can now be attached to a single message. - SillyTavern Presets support regexes and older preset formats from before SillyTavern 1.10. - The MCP Support mini-app adds automatic server detection and searchable, filterable tool lists. - The References mini-app no longer requires Long-term memory, and uses it to improve document ingestion when installed. - The ElevenLabs integration now uses their newest API and supports multilingual voices. - Faster loading of the Character Selection page for large character libraries. - Pagination now accepts a page number instead of stepping one page at a time. FIXED - Fixed backups becoming corrupted if the app was closed too soon after starting one. - Fixed some models failing to handle a greeting or assistant message as the first message. - Fixed message regeneration not working. - Fixed max_response_length = 0 not being respected. - Fixed image-generation context no longer working. - Fixed animated images (WebP, APNG) losing frames when cropped. - Fixed exporting an LLM not keeping its file extension. - Fixed Python error handling on iOS.

- fixed a crash after music generation

- fixed a crash using MobileVLM - fixed missing translations in the new settings page

- fixed a bug where Layla Cloud asks you to login repeatedly - fixed a bug where model benchmarks were not showing progress