Skip to content

CHANGELOG.md

Latest commit

 

History

History
82 lines (59 loc) · 5.17 KB

CHANGELOG.md

File metadata and controls

82 lines (59 loc) · 5.17 KB

CHANGELOG

All notable changes to KREASYS are documented in this file.

[0.4.0] - 2026-03-08

Added

  • AI-Controllable Browser Panel: A new Browser tab with a sandboxed iframe viewer, URL bar, and integrated JS Console. The AI can now autonomously open URLs, run JavaScript in page context, and take screenshots.
  • Three New Agentic XML Skills:
    • <browser_open url="..."/> — Opens a URL or inline HTML blob in the browser iframe. Navigates the Browser tab automatically.
    • <browser_js>...</browser_js> — Executes arbitrary JavaScript inside the loaded iframe and logs the return value to memory.
    • <browser_screenshot out="/workspace/screen.jpg" [tg_chat_id="ID"]/> — Captures the iframe using html2canvas, saves the image to the VFS, and optionally dispatches it directly to a Telegram user via sendDocument.
  • Page Scraping: The "Scrape" button extracts the full text content and title of any loaded page and writes it to /workspace/scrape_result.txt in the VFS for the AI to read.
  • js/ui/browser.js Module: All browser panel logic is decoupled into its own module.

Changed

  • WebLLM Fully Simplified: webllm.js was completely rewritten to mirror the minimal experimental.html pattern.
    • The hardcoded WLLM_MODELS array has been removed.
    • Model list is now dynamically fetched from webllm.prebuiltAppConfig.model_list on page load, giving users access to the complete and always-up-to-date catalog from the WebLLM CDN.
    • Removed the complex download library/cache tracking UI (renderWllmLibrary, wllmActivate, wllmDelete, wLib, wMarkDownloaded).
    • Simplified to: select model from dropdown → click Download & Load → engine.reload() with progress callback.
  • app.js Cleaned: Removed all dead WebLLM library management functions. wllmInit() now delegates to wllmPopulateDropdown().
  • System Prompt Updated: D_SKIL in state.js now includes documentation for <browser_open>, <browser_js>, and <browser_screenshot> with Telegram dispatch.
  • Script Cache-Buster bumped to ?v=4 across all core JS includes.

[0.3.2] - 2026-03-07

Added

  • Autonomous Telegram File Dispatch (<tg_doc>): The AI can now autonomously upload any VFS file (images, code, text) directly to a Telegram user's chat using the Telegram sendDocument API. The file is read from the VFS, converted from Base64 or plain text to a Blob, and uploaded as a proper file attachment.
  • Unified Models Tab: The "Local AI" sidebar tab was removed. The WebLLM local compute interface was fully integrated into the main "Models" tab under a "Local Browser Compute" card.
  • AI Activity Tracking Dashboard: A new "Tracking" tab (#p-activity) provides a live visual timeline of all system events, VFS writes, and AI execution plan steps (<plan> tags).

[0.3.1] - 2026-03-07

Fixed

  • <render_html> VFS Race Condition: Inline HTML in block-form <render_html> tags is now processed directly without a separate <file> write, eliminating the VFS write/read race condition.
  • WebGPU Hard Block Removed: The strict WebGPU check in webllm.js was converted from a hard blocker to a soft warning. Users can attempt any model download regardless of GPU probe result.
  • AI Streaming Enabled: llm() refactored to use fetch Body.getReader() for SSE streaming. Cloud APIs and local WebGPU now stream tokens to the UI in real-time.
  • Script Cache Busting: All core JS scripts include version query strings (?v=3) to force cache invalidation.

[0.3.0] - 2026-03-07

Added

  • HTML-to-Image Rendering (<render_html>): AI can generate visual content using html2canvas, rendering styled HTML directly to a Base64 JPEG in the VFS.
  • Sub-Agent Delegation (<delegate>): Main agent can spin up isolated sub-LLM calls with custom temperatures for pipeline-style parallel task execution.
  • <plan> Enforcement: System prompt now strictly requires <plan> tags for complex agentic workflows.

[0.2.1] - 2026-03-01

Added

  • Single-Instance Telegram Polling Lock: Uses a randomly generated Tab ID stored in localStorage to ensure only one browser tab polls the Telegram API at a time, eliminating 409 Conflict errors.
  • "Channeling" Dashboard UI: Redesigned integrations tab with a user directory and multi-service routing framework.
  • "Take Control" Override: Manual button to steal the Telegram polling lock from other open tabs.

[0.2.0] - 2026-02-28

Added

  • Dual-Layer Persistent Memory: memory.log (raw session) + memory.md (AI-summarized). Idle-time auto-compression prevents context window bloat.
  • Telegram Autonomous Hub: Bot auto-registers users; AI can cross-message other registered users on its own initiative.
  • Local AI Tab (WebGPU): First iteration of the local model management UI with Llama, Phi, Gemma model support.
  • Multi-Modal API Router: Intent-based routing across OpenRouter, Groq, NVIDIA, OpenAI, HuggingFace, Anthropic, and Ollama.

[0.1.0] - Alpha

Added

  • Virtual File System (VFS) via localforage.js backed by IndexedDB.
  • Telegram webhook polling, photo/document ingestion.
  • Basic <plan> flowchart rendering in the UI.