Development loop¶

Skip the build with `--dev`¶

Every karotte run without --dev rebuilds the image first. podman and docker reuse cached layers, so an unchanged environment builds quickly. A changed one can take a while.

uv run karotte run --config run_config.json --dev

--dev skips the build and runs the existing image. It bind-mounts your src/environment/ over the installed copy at /root/.venv/lib/python3.12/site-packages/environment/. That way, changes to task code, judges and tools take effect on the next run.

Build the image once before your first --dev run. Rebuild it after you change any of these:

dependencies (pyproject.toml, uv.lock, the venvs in venvs/), including a Karotte update
the Containerfile
data in student_data/, shared_data/, root_data/ or intermediate_data/
the CLI agents in the environment's manifest

You can combine --dev with --mount.

Build without running¶

uv run karotte build

karotte build builds the same image without running a task. run always uses the image tagged karotte, which is the default tag for build. There's no option to make run use a different image. Use build --tag for images you want to use elsewhere.

--build-secret name=path passes in a file that the Containerfile can mount with RUN --mount=type=secret,id=<name>. run takes the same option.

Mount host paths¶

--mount bind-mounts a host path into the container. The format is source:target[:ro]. Mounts are read-write by default. Append :ro to make one read-only. The target must be an absolute path, and the source must exist.

uv run karotte run --config run_config.json --mount /path/on/host:/path/in/container
uv run karotte run --config run_config.json --mount /data:/data:ro

To add several mounts, pass --mount more than once, or read them from a file with @:

uv run karotte run --config run_config.json --mount @mounts.txt

The file has one spec per line. Karotte ignores blank lines and lines starting with #.

Warning

With -n (parallel runs), all containers share the same bind mounts. That makes a read-write mount shared state between them, and Karotte warns you about it. Use :ro where you can.

The VM runtimes handle mounts differently. firecracker supports read-only mounts only. apple-container copies each mount in when the run starts, and copies the read-write ones back when it ends. See Runtimes.

To change a task's behavior without a rebuild or a mount, pass values through extra_config in the run config. See Run config.

Keep containers¶

--keep-containers keeps the containers after the runs instead of removing them, so you can inspect them or copy data out. They're named karotte_run_<run_id>. When the next run starts, Karotte removes leftover containers from earlier runs whose names start with karotte_run_ followed by the longest common prefix of the new run ids.

When the run ends, Karotte prints how to copy the student's workdir out, for example:

docker cp karotte_run_<run_id>:/workdir/ ./out/

firecracker doesn't use a container. Instead, Karotte keeps the VM's drives and logs where they are.

Credentials in the image¶

karotte check runs as the last step of the image build. It runs inside the image and fails the build if the image contains credential material. It looks in the home directories (/root, the student's workdir, /home/*) for the places where credentials end up: uv's credential stores, ~/.netrc, ~/.git-credentials, the GitHub CLI's token, gcloud's credential files, ~/.aws/credentials, ~/.kube/config, Hugging Face tokens, keyring stores and SSH private keys. Config files such as .npmrc, .pypirc, .docker/config.json, pip.conf and uv.toml fail the check only if they hold a secret inline. If they read the secret from an environment variable, they pass.

A build secret mounted with --mount=type=secret stays out of the layers. But a tool you give the secret to can still save something it derives from it, usually a cached token under the home directory. That write ends up in a layer like any other file. Making the file root-only doesn't help, because anyone who can pull the image can read it.

If the check fires, delete the file in the same RUN step that created it, and rotate the credential it came from. It has to be the same step. The check sees only the final filesystem. If a credential is written in one layer and deleted in a later one, it passes the check but still ships in the layer that holds it.

`--no-containerized`¶

--no-containerized runs the task in the current process instead of starting a container. It only works inside a Karotte image, where the Containerfile sets KAROTTE_CONTAINERIZED. Anywhere else, karotte run refuses it. Karotte uses it inside the container it starts.

--dev, --mount, --keep-containers and -n greater than 1 all need a container, so you can't combine them with --no-containerized.

Lint and test¶

An environment's justfile has these recipes:

Recipe	What it does
`just lint`	Runs `ruff format --check`, `ruff check`, and a check of the uv supply-chain settings.
`just fix`	Runs `ruff check --fix`.
`just fmt`	Runs `ruff format`.
`just test`	Runs `pytest`.

uv sync --extra dev installs just into the venv, so run the recipes through uv:

uv run just lint
uv run just test

uv run just --list shows every recipe, including the ones that create tasks. The default template ships tests in tests/ for task discovery and its other helpers. Add your own tests there as you change things.

Past transcripts¶

uv run karotte dashboard <dir> shows the transcripts in a directory. By default, runs write to out/. See Artifacts and transcripts.