TIRx Harness

Use TIRx kernel-development skills in your own agent, or start an optimization run with a prepared task, isolated workspace, and benchmark.

The documentation covers the components, agent loop, and guides for adding workloads and contributing kernels and fixes. See build and preview instructions to run the site locally.

Start with Installation for packages and skills, then follow the Quick Start for an example agent goal.

See prerequisites before installing.

Repository map

Run preparation and execution code lives in evolution/. Generated workspaces and results live under kda_flow_runs/; each workspace stores candidate kernels in candidates/<workload>/scratch/ and frontier/.

Installation

For your own Python environment, install the harness with python -m pip install --group harness .. Copy the three local skill directories to your agent's skills directory and run tirx-wiki/scripts/fetch_references.py from the copied wiki directory. The Installation guide gives the complete commands.

The reference fetcher also downloads tirx-kernels into the wiki's references/repos/tirx-kernels/, including its root README and source docs. The installed tirx-kernels package is used for execution.

For an optimization run, follow Optimization Runs to install the setup dependencies and select a task and GPU/server. Setup creates the run's worktree and venv and installs its packages and skills. Review the generated prompt, then launch your agent with the printed command.