Changelog | YuvaDev

The agent loop

June 2026

▸Model-driven agent loop: the model picks tools (read, write, edit, search, run) and iterates until the task is done — replacing the fixed Planner → Coder → Reviewer pipeline.
▸Native tool calling across Anthropic, OpenAI-compatible, DeepSeek, and Ollama (local and cloud), with a text fallback for models without native support.
▸MCP client: connect servers via a Claude Code-compatible mcp.json (stdio and HTTP); their tools join the agent’s toolbox.
▸yuvadev CLI: the same agent in your terminal — streamed tool calls, colored unified diffs, approval prompts, and an interactive mode.
▸Project memory: YUVADEV.md, AGENTS.md, and CLAUDE.md from your workspace are injected into agent prompts.
▸Refreshed model defaults to current-generation models (Claude Sonnet 4.6, GPT-4.1).

May 2026

▸Autonomous edit validation with placeholder and stub detection before changes are proposed.
▸Self-healing CI loop: the agent runs tests, detects failures, and iterates fixes with convergence detection and budget guards.
▸“Edit with AI” CodeLens for Python, TypeScript, JavaScript, Java, Go, and Rust.
▸Provider quick-switch (Cmd+Shift+L) and one-command Ollama onboarding.
▸Web Mission Control: progress, file change summaries, healing telemetry, elapsed time.
▸Security hardening: strict CSP on all webviews, explicit CORS allowlists, host validation on the embedded backend.