[eliotwrobson]

Changes

MessageType enum. The runner compared message types to raw strings, so a typo would silently match nothing. The runner, fixture and request types now use MessageType. The values sent over the socket don't change.

Server failures. When the sandbox server hits its own error, it sends {"status": "failure", "message": ...} and closes the connection. The grader never handled that shape, so it reached tests as a KeyError, and later queries failed with socket errors.

• The status is now ServerStatusCode.FAILURE, with a ServerFailureResponse type.
• query_raw, query, query_function_raw and query_function raise a RuntimeError: The sandbox stopped unexpectedly, so later queries won't work: .... Later queries on that sandbox raise the same error.
• This raises instead of returning a response, because the connection really is gone, unlike the cases in Return serialization failures from query_function_raw normally #29 and Return serialization failures from query_raw normally #30.

Not changed:

• query_setup_raw still doesn't handle a failure response, because Protect grader data from student code and load data.json in images #27 rewrites it.
• The start response keeps its existing "Unexpected status" handling.
• Value type tags in json_utils.py, pytest outcomes and TypedDict keys stay as strings.

Testing

• New test_server_failure scenario passes an argument the sandbox can't load, then checks the error, later queries and the message a student sees.
• uv run pytest tests -p no:prairielearn-grader on Python 3.14: all 73 passed on two runs after merging main. Before that merge, the first run in a fresh environment failed only test_matplotlib_plots, a first-run failure also seen without this change.
• mypy src/ is clean, and ruff check shows nothing new.
• Lite image built from this branch's docker/ files, with this branch's wheel installed over the PyPI package. I confirmed the installed package contains the enum. With PrairieLearn's square tests and a hand-written data.json, a correct submission scored 1.0, and x_sq = Thing() scored 0.0 with Return serialization failures from query_raw normally #30's message. The server failure path itself is covered by the scenario, not by an image run on this merged code.
• This branch merges cleanly into main after Return serialization failures from query_raw normally #30. All tests passed on a trial merge with Protect grader data from student code and load data.json in images #27 and Return serialization failures from query_raw normally #30 together, with one import line in fixture.py resolved by hand.

The full image wasn't built.

🤖 Generated with Claude Code