RIP ( Coding Harness )
rip is a small coding harness built around an explicit agent loop. You give it a prompt and a workspace. The model streams a reply, requests tools, and receives the tool output on the next turn, up to a fixed iteration limit. Tools can list a directory, search file names, grep file contents, read a file, read file metadata, and write a file. Each operation runs in a short-lived container with no network, a read-only root filesystem, dropped capabilities, and a bind mount limited to the selected workspace. Model access sits behind one provider interface, with adapters for the OpenAI Responses API and for Gemma on Ollama. The harness is meant to show the runtime around the model: streaming, tool routing, and a boundary on what the model is allowed to touch.
The constraint
A model that can write code still needs a loop that turns a request into tool calls, feeds results back, and stops. It also needs a place to run those tools that is not the host filesystem with full permissions. Provider APIs do not share a streaming or tool format, so swapping OpenAI for an Ollama model means rewriting the loop unless that translation lives in one adapter. The project is a small, readable runtime that keeps those decisions in the open.
How it was built
- –CLI that takes a prompt, a provider, a model, and a workspace, defaulting the workspace to the current directory
- –Agent loop that streams text, records tool calls, runs the matching tool, and sends the result back for the next turn, capped at 40 iterations
- –One provider interface with adapters for OpenAI and for Gemma through the Ollama cloud API
- –File tools for listing, name search, grep, reading, metadata, and writing
- –Docker sandbox per tool call: no network, read-only root filesystem, memory and CPU limits, a timeout, and path checks that reject escapes and symlinks leaving the workspace
- –Writes allowed only through the write tool, on a workspace mount, with a size limit
Stack
The repository is private. Email sidhantsinghrathoreprsnl@gmail.com with your GitHub username and I'll add you.