v0.2.0 — Debug running servers without restart

Debug running servers — no restart, no code changes

Note: This release supports Python only. Support for JavaScript, Go, Rust, and other languages is planned for future releases.

The biggest limitation of v0.1 was that you had to launch your program through the debugger. That doesn't work for servers — nobody wants to restart their running uvicorn/flask/django process just to debug it.

v0.2.0 fixes this. Just point at a running process and go:

agent-debugger attach --pid <PID> --break app/routes.py:42
agent-debugger continue   # trigger a request, hit the breakpoint
agent-debugger vars        # inspect state

That's it. The server keeps running. No restart, no code changes, no debugpy setup required.

What happens under the hood

1. Auto-detects the Python runtime and venv from the running process
2. Auto-installs debugpy if it's not already in the environment (via pip, using the process's own sys.executable)
3. Injects debugpy.listen() into the live process using the native debugger (lldb on macOS, gdb on Linux) — calling Python C API functions directly
4. Connects the DAP client to the spawned adapter
5. Sets breakpoints and hands control back to you

Also in this release

• Linux support — gdb-based injection alongside lldb (macOS)
• Port-based attach — agent-debugger attach [host:]port for processes already running with debugpy.listen()
• macOS ARM64 fix — debugpy's built-in --pid inject ships only x86_64 dylibs; the new lldb/gdb approach works natively on Apple Silicon

Requirements

• Python processes only (other languages coming soon)
• macOS: Xcode Command Line Tools (for lldb) — already installed if you use git
• Linux: gdb (apt install gdb / yum install gdb)
• Target process must have pip available (for auto-install of debugpy)

v0.1.1

• Add npx fallback to skill for zero-install usage
• Add MIT license

v0.1.0

Initial release — CLI debugger for AI agents via DAP