Prompt-pack inspection
See system, developer, user, retrieval, tools, runtime constraints, and raw payload fields without sending them to a model provider.
What it does
ModelMirror lets a human step into the model's seat before an agent runtime calls an LLM. It loads a prompt bundle, validates the structure, shows messages, tools, retrieved context, and constraints, then captures whether the expected next move is a response, tool call, clarification, or refusal.
Human emulation loop
Record what a reasonable model should do next: respond, call a tool, clarify, or refuse, with structured artifacts.
Local by design
No telemetry, no network runtime calls, no tool execution. The base debugger stays deterministic and non-executing.
Run it locally
Clone the repo, install editable, validate a bundle, then open the debugger UI.
git clone https://github.com/Roop-World/ModelMirror.git
cd ModelMirror
python3 -m pip install -e .
python3 -m model_mirror run --bundle templates/packed_prompt.example.json --out outputs --operator demo_operator