← field notes

qor harness: Strap in!

Strap in!
pipx install qor-harness

Core strength for your agents. Your tests. Your models.

I've been building qor harness, pronounced “core.” It puts a coding agent right where I want to work: in the terminal, with the project in front of it.

Not another pile of code to copy out of a chat window. Give it tools to try the change, and a test suite that doesn't care how confident the answer sounds.

Read the code. Make the change. Run the tests. Keep moving.

Three lines to the starting line

With Python 3.12 or newer and pipx installed:

pipx install qor-harness
cd /path/to/your/project
qor

First launch walks you through connecting a model and choosing a default. Bring an existing model server or a supported account; qor doesn't download a model or install a server behind your back.

Then give it a real job:

Find out why this test is failing. Read the relevant code, make the smallest fix, and run the test again. Tell me what changed and what you checked.

The model can inspect the project, edit the file, run the test, and use the result to decide what to try next. Its calls and results are visible in the terminal. Less shuttling between windows. More work happening where you can check it.

The model is the jetengine. This is the cockpit.

The model is only part of a coding agent. The harness gives it tools, feeds back their results, and manages the conversation around the work. The prompt, tool design and context handling all shape how that agent behaves.

qor gives the model tools to read files, search the project, edit code, and run shell commands. A failed test comes back into the conversation, where the model can inspect it and try another fix. A confident answer isn't the finish line. Check the diff. Run the test.

It's not limited to writing code, either. Ask it to trace a confusing module, explain a repository, review a change, or untangle a build. The tools work on your project, not a copy you assembled for a chat window.

Stay involved. Send a follow-up while it's working. Cancel when it's heading toward a refactor you never asked for. Come back to a saved session with /resume instead of explaining the project all over again.

For a quick job without the interactive interface:

qor run "Read this project and explain how to run its tests."

Bring your model

qor isn't tied to one model vendor.

  • Ollama or llama.cpp: connect to a server you already run. Discovery helps set up model choices and supported settings.
  • OpenAI-compatible endpoints: point qor at your endpoint and register your models. Compatibility varies, so check the endpoint you plan to use.
  • ChatGPT account: qor login signs in with an account that has the required agent-preview access. This is separate from OpenAI Platform API-key billing.

Use /model to switch models without throwing away the conversation. A smaller model for a quick question, something larger for the hard part. Keep the work, change the engine. Set a favorite as the default for next time.

Less log. More room to work.

Coding sessions collect baggage. Test output. File listings. The same source file read three times. Useful when it happens; expensive clutter when it occupies every later request.

qor's Deterministic Context Management (DCM) keeps older tool output from sitting in the live conversation forever. Large results are saved; older output can become a compact pointer, and older turns can fold into an index. The model can use recall to recover the captured text instead of relying only on a summary.

You can dig through the session yourself with /search, open a saved result with /recall, and see usage with /stats.

Spend the context on the work in front of you. Keep earlier observations within reach.

Your terminal. Real permissions.

qor runs with your user account's permissions. By default it can edit files and execute commands without asking for routine approval. For a more deliberate first run:

qor --review

Review mode lets you approve proposed changes and commands before they run. Shell access is not a sandbox. Start in a project you're comfortable experimenting with, keep your work in version control, and read the diff.

Take it for a spin

qor harness 0.2.0 is the version I'm sharing now: a keyboard-first interface, real project tools, model choice, saved sessions, DCM, and an opening sequence with no interest in being a gray loading spinner.

Get qor-harness on PyPI, install it, and give it something useful to do.

pipx install qor-harness

Strap in!

← all field notes