diff --git a/.dockerignore b/.dockerignore index e06167f4..3ea5b1ba 100644 --- a/.dockerignore +++ b/.dockerignore @@ -3,3 +3,6 @@ __pycache__ *.pyc .pytest_cache +results +.jetskicli + diff --git a/evalbench/evalbench.py b/evalbench/evalbench.py index 943a3240..c80910f3 100644 --- a/evalbench/evalbench.py +++ b/evalbench/evalbench.py @@ -34,6 +34,7 @@ None, "Path to a suite configuration file to run multiple experiments.", ) +flags.declare_key_flag('experiment_config') def eval(experiment_config: str): @@ -208,6 +209,18 @@ def main(argv: Sequence[str]): def run(): """Starting function for the uvx package entrypoint.""" + # Fix absl help output when run via uvx/launcher + if '__main__' in sys.modules: + main_module = sys.modules['__main__'] + if main_module.__doc__ and 'exec' in main_module.__doc__: + main_module.__doc__ = sys.modules[__name__].__doc__ + # Clean up sys.argv[0] to hide the full temporary path + sys.argv[0] = os.path.basename(sys.argv[0]) + # Register key flags for __main__ and sys.argv[0] so they show up in launcher's short help + flags.FLAGS.register_key_flag_for_module('__main__', flags.FLAGS['experiment_config']) + flags.FLAGS.register_key_flag_for_module('__main__', flags.FLAGS['suite_config']) + flags.FLAGS.register_key_flag_for_module(sys.argv[0], flags.FLAGS['experiment_config']) + flags.FLAGS.register_key_flag_for_module(sys.argv[0], flags.FLAGS['suite_config']) app.run(main) diff --git a/evalbench_service/Dockerfile b/evalbench_service/Dockerfile index bf0a42db..5eb43485 100644 --- a/evalbench_service/Dockerfile +++ b/evalbench_service/Dockerfile @@ -24,8 +24,9 @@ RUN curl https://packages.cloud.google.com/apt/doc/apt-key.gpg | gpg --dearmor - WORKDIR /evalbench COPY pyproject.toml uv.lock ./ COPY viewer/pyproject.toml viewer/ -RUN uv sync --frozen +RUN uv sync --frozen --all-packages COPY . . +RUN uv sync --frozen --all-packages # Create a non-root user for Claude Code. It refuses # --dangerously-skip-permissions when running as root. diff --git a/viewer/version.txt b/viewer/version.txt index 5ef1b5f8..cbf16756 100644 --- a/viewer/version.txt +++ b/viewer/version.txt @@ -1 +1 @@ -023d150 +dbfbafa