DeepRhapsody

AI-powered debugging for software (8 languages) and LLM/transformer reasoning. Debug programs with natural language via real debuggers (GDB, LLDB, CDB, JDB, Delve, Node Inspector, rdbg). Debug LLM internals with Logit Lens, Attention Analysis, Probing, Activation Patching, and LoRA fine-tuning. Client-server architecture works with any AI agent.

Debug C/C++ programs at the assembly and machine-code level. Disassemble instructions, step through individual CPU instructions, inspect registers and raw memory, set breakpoints on addresses, patch live process memory, NOP out instructions, and modify binaries on disk. Works with GDB, LLDB, and CDB. All standard source-level debugging commands remain available.

Debug Python, C/C++, C#, Rust, Java, Go, Node.js/TypeScript, and Ruby programs interactively (persistent session with step/continue/inspect commands) or non-interactively (one-shot). Python uses stdlib bdb; C/C++ drives GDB (MI mode), LLDB, or CDB; C# uses netcoredbg (MI mode); Rust uses rust-gdb/rust-lldb/GDB/LLDB; Java uses JDB; Go uses Delve; Node.js uses built-in inspector; Ruby uses rdbg. Supports auto-detection of platform, compilers, and debuggers.

Debug reasoning of hosted LLMs (GPT-4, Claude, Gemini) through API-based probes — logprob analysis, consistency testing, chain-of-thought extraction, counterfactual probing, and calibration checks. No model weights required.

Debug LLM/transformer model reasoning interactively using PyTorch hooks and interpretability techniques. Step through forward passes layer by layer, inspect activations and attention weights, run Logit Lens to see how predictions evolve, Activation Patching to find causal layers, Probing to test what information is encoded, and Attention Analysis to understand head behavior. Supports GPT-2 family, Llama, Mistral, Qwen, DeepSeek, and custom architectures via model adapters. Same TCP/JSON protocol as the traditional debugger skill.

Fine-tune GPT-2 family models with LoRA to inject missing knowledge. Diagnose knowledge gaps, generate paraphrased training data, run lightweight LoRA training, verify predictions before/after, and persist merged weights to disk. Supports distilgpt2, gpt2, gpt2-medium, gpt2-large, gpt2-xl. Uses the same TCP/JSON server as the LLM debugger skill.

Advanced LLM interpretability toolkit — Sparse Autoencoders, neuron-level analysis, hallucination detection, and attention head surgery. Decomposes model internals into interpretable features, finds important neurons, detects hallucinations, and enables targeted interventions on attention heads. Uses the same TCP/JSON server as the LLM debugger skill.

Static binary analysis for reverse engineering. Parse PE/ELF headers, extract strings, discover functions in stripped binaries, build cross-reference maps, generate control flow graphs, analyze entropy for packing detection, and perform static disassembly. Works with .exe, .dll, .so, .elf, and raw binaries. Pure Python — no external dependencies required.