v0.1.6

Core

• Agent mode rewritten from rigid generateApplication() to tool-calling loop (ToolLoopAgent) — model calls write_file/run_command tools directly, enabling iterative building and error recovery.
• New /model slash command — switch AI provider/model mid-session (Gemini, OpenRouter, NVIDIA NIM).
• Persistent stdin handler — prevents "terminal state may be corrupted" crashes.
• Stream cancellation via Escape key, with partial response preservation.
• Global unhandledRejection/uncaughtException handlers for crash resilience.

Tools

• Standalone tool definitions: write-file, run-command, read-instructions with path traversal protection, 1MB size limit, auto mkdir, and interactive stdin support.
• PermissionManager — granular tool permission system with wildcard matching, dangerous command detection (20 regex patterns), and [y] Once / [a] Always / [n] Deny prompts.
• ToolExecutor — manual tool-calling loop orchestrator for streaming providers.
• OpenRouter provider rewritten from SDK to raw fetch API with native tool call detection and multi-iteration tool loops.
• NVIDIA NIM provider migrated to @ai-sdk/openai-compatible SDK with proper tool-calling support.

CLI

• Model switching at runtime via /model command: Google Gemini, OpenRouter (GPT OSS, DeepSeek V4 Flash, MiniMax M3, GLM 5.1, Kimi K2.6), NVIDIA NIM (MiniMax M2.7, DeepSeek V4 Flash, Llama 3.3 70B).
• Default OpenRouter model changed to moonshotai/kimi-k2.6:free.
• Animated thinking spinner with live tool call and reasoning updates.
• Agent mode shows live tool call progress; supports exit and continue prompts.
• Chat loop resilience — catch-all error handling prevents crashes; terminal state restored on errors.