You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Web Crawl Tool – New crawl_website tool for comprehensive site-wide extraction (docs, knowledge bases, etc.) with depth control, path filtering, and natural language instructions.
GPT‑5.3 Codex Spark – Ultra-fast coding model now available via OpenAI Codex.
Claude Sonnet 5 – Added across Anthropic, CommandCode, and GitHub Copilot providers.
Ad‑hoc Skills Loading – Skills can now be discovered and registered at runtime without requiring pre‑existing skill files.
OAuth Lock for MCP – Prevents concurrent OAuth flows from conflicting on the same callback port during streaming server session creation.
Image Extraction in Web Fetch – The extract_webpage tool now returns up to 3 images as base64 data URIs alongside text content.
Cosine Distance for Memory – Switched to cosine distance metric in ChromaDB for improved memory retrieval quality.
Shrunk Context Argument Preview – Truncated tool call arguments now include a preview (e.g., tool_name(key=value...)) for better debugging.
Memory Similarity Score – Memory recall now shows similarity score in the header for transparency.
🔧 Changed
GitHub Copilot Models – Updated to latest model lineup (Claude Opus 4.8, Claude Fable 5, Gemini 3.5 Flash, MAI-Code-1-Flash, Kimi-K2.7-Code, GPT‑5.4 nano, etc.).
Web Search Tool Description – Simplified and clarified tool descriptions for better LLM understanding.
Tavily Extract Depth – Upgraded to advanced extraction depth for richer content.
Dependency Updates – tavily-python bumped to 0.7.26, chromadb to 1.5.9.
Memory Thresholds – Adjusted relevance thresholds for both default and Voyage embedding functions to improve recall precision.
Skills Service – SkillsService now always initializes even when no skills are found, enabling dynamic skill registration later.
🐛 Fixed
Context Consistency Between Modes – Fixed assistant message not being appended to history in certain agent loop exit paths, ensuring consistent context across different execution modes.
ACP Session Race Conditions – Added per‑session asyncio.Lock to serialize prompt, close_session, and cancel calls, preventing concurrent state corruption.
ChromaDB Collection Space – Automatically recreates collection with cosine space if it was previously using a different metric (e.g., l2), avoiding compatibility issues.
Task History Append – Corrected method call in A2A task execution to use the proper history reference.
📦 Version & Build
Version bumped from 0.17.6 → 0.18.0 in __init__.py, pyproject.toml, and Docker configuration.
🙏 Contributors
Quy Truong (@daltonnyx) – for all commits in this release.
Special thanks to the open‑source community for ongoing feedback and contributions.