v0.1.1 — Provider Override & Sandbox Isolation
Highlights
🔀 Model-agnostic — Supports Google Gemini, Anthropic Claude, and any OpenAI-compatible provider out of the box. Switch with --model or --provider.
🛡️ OS-level sandbox isolation — Shell commands run inside sandbox-exec (macOS) or bwrap (Linux) by default. Network access denied; writes restricted to project root.
🧩 Extensible skill system — Compatible with the Agent Skills open standard. Built-in skills: /review, /explain, /debug, /test, /commit.
📋 Plan mode — Review changes before they're applied (/plan <task>). Read-only tool access during planning, full execution after approval.
⚡ Lightweight — No heavy framework. Just npm install -g @zjshen/opencli and go.
What's New in v0.1.1
- Explicit
--providerand--base-urlflags for connecting to any OpenAI-compatible API - Sandbox isolation for bash tool (macOS sandbox-exec, Linux bwrap)
- Stuck-loop detection (aborts after 3 identical consecutive tool calls)
- Session management (resume conversations with
--resume) - MCP client integration (load external tool servers from
mcp.json)
Install
npm install -g @zjshen/opencli