scripts/ only keeps the currently maintained entry scripts and a few reusable manifests; new temporary experiment manifests should not be committed here.

scripts/ no longer keeps Python evaluator wrappers. When adding a task, place the system-side evaluator at tasks/<category>/<suite>/<task>/evaluator.py or tasks/<category>/<task>/evaluator.py, and use the task_package backend in the manifest.

Historical opt-solver command manifests are kept only as archived records; legacy evaluator wrappers are no longer maintained. To reproduce an experiment, migrate the old manifest to a task_package evaluator first. Data preparation runs through the data-prep agent of scienceflow prep, and task-shape rules live in .scienceflow/skills/.

uv run python -m scienceflow.cli parallel -m scripts/prep.yaml -j 1

# For tasks that need prep output, set later LNR input_data_dir to:
# ./workspaces/manual_data_prep/<run_id>/<exp_id>/dataset

uv run python -m scienceflow.cli parallel -m scripts/lnr.yaml -j 1

uv run python -m scienceflow.cli parallel -m scripts/lnr_nomad2018_3seed_cpu.yaml -j 3

./scripts/monitor_lnr.sh scripts/lnr.yaml 5

./scripts/lnr_kill_resume.sh /path/to/<run_id>/<exp_id> \
  --config scienceflow/config/default.yaml \
  --input-data-dir ./data/mlebench_all_data/<exp_id>/prepared/public

uv run python -m scienceflow.cli parallel -m scripts/lnr_two_tasks.yaml -j 2

monitor_lnr.sh is the monitoring entry for LNR run status, not a task execution entry. It takes a manifest path, invokes the current scienceflow.cli monitor, reads the LNR logs from the task directory the manifest points to, and refreshes the display.

./scripts/monitor_lnr.sh <manifest.yaml> [refresh_sec]

• <manifest.yaml>: an LNR manifest that has already started or is about to be monitored.
• [refresh_sec]: optional refresh interval, default 5 seconds.
• The script only reads task logs/resource events/stage CSVs; it never starts, stops, or modifies tasks.
• Experiments are still launched with uv run python -m scienceflow.cli parallel -m ....

• A scripts/ manifest only records run deltas: tasks, workspace, resource isolation, budget, and the LNR/REPL parameters you actually need to override. Stable values such as REPL profile, bash tool, workspace git, stage capture, estra, resource monitor, and log level come from scienceflow/config/default.yaml.
• Resource management in canonical manifests only exposes lnr.resource_control_mode; the default is resource_smart_llm. Do not expand low-level switches such as arbiter/advisory/GPU share at the script top level.
• The default REPL profile is lite; do not use the old codex_like naming.
• For multi-task experiments, put temporary manifests under mlebench_workspaces/<date>_run/<run_name>/; do not commit them to scripts/.
• parallel writes outer subprocess logs to each task workspace's task_logs/parallel_subprocess.log by default; do not use the repo-root parallel_logs/ anymore.
• Evaluators for new tasks do not belong in scripts/. With the task package backend, the evaluator code is visible to the system but not to the agent; the agent only sees the task description, the data view, and the artifact contract.
• Data preparation is no longer exposed to top-level runs through a standalone deterministic split script; use the data-prep agent of scienceflow prep, and keep task-shape rules in .scienceflow/skills/.
• prep.yaml uses phase: prep, and the output directory is organized as <workspace_base>/<run_id>/<exp_id>/dataset. A later LNR run can point input_data_dir at this dataset/, or directly at an already trusted prepared/public.
• CPU isolation is declared per task with cpu_list, and lnr.omp_threads_cap should match the cores available to a single worker. For example, with 2 workers at 8 cores each: task-level cpu_list: "0-15", lnr.num_workers: 2, lnr.omp_threads_cap: 8.
• GPU tasks declare gpu_list: "0" explicitly and keep the task-level lnr.resource_gpu_pool consistent, e.g. ["0"]. CPU-only tasks use gpu_list: "cpu" and set lnr.resource_gpu_pool: [].
• The default nomad2018 validation setup uses prepared/dataset_split/Deep, wall_clock_budget_sec: 3600, 2 workers, 8 CPUs per worker, CPU-only, and no skills.
• When two tasks run in parallel, -j should equal the number of concurrently running tasks; the worker count inside a task is controlled by task-level lnr.num_workers.

• New temporary experiment manifests are not committed to scripts/; put them under mlebench_workspaces/<date>_run/.
• Prefer placing plotting and analysis scripts in .codex/skills/, an external experiment workspace, or a task-specific directory.
• If a script does not serve the current LNR tracking, resume, REPL, monitor, or two-task parallel entry points, do not put it back into scripts/.