Show HN: Ldbg: a Python library to ask LLMs while debugging (github.com)

Ldbg is a new Python library that lets you query an LLM directly from debuggers (pdb, ipdb, Jupyter, VS Code Debug Console, etc.), automatically enriching prompts with your call stack, local/global variable previews, current function source and surrounding code. From natural-language prompts like “describe my numpy arrays” or “plot example_numbers as a bar chart,” ldbg will generate Python debug commands (e.g., pandas.describe(), matplotlib code), show the suggested code, and ask permission before executing it in your session. It’s installable via pip/uv/pixi, MIT-licensed, and examples show it working with modern models (the demo uses “gpt-5-mini-2025-08-07”). This is significant because it brings on-demand, context-aware code synthesis into interactive debugging—speeding exploration, reducing boilerplate, and helping less experienced users craft analysis or visualization code quickly. Key technical notes: ldbg defaults to OpenAI (reads OPENAI_API_KEY) but supports multiple providers (Anthropic, DeepSeek, Groq, Together, OpenRouter, Ollama) via LDBG_API and provider-specific keys; you can override the model per call. Important caveats: the project prominently warns (“DO NOT USE THIS LIBRARY”) — risks include accidental execution of unsafe code, leaking sensitive runtime state to third-party APIs, costs/emissions, and possible overreliance on models. Using local providers like Ollama can reduce telemetry, but exercise caution and review any model-suggested code before running.