Launch Stage 1: generate sulcal graphs with Morphologist from raw T1 MRI data.

Launch Stage 2: extract 28 sulcal region crops with cortical_tiles.

Launch Stage 3: generate Champollion dataset YAML configuration files.

The configs root defaults to //derivatives/champollion_V1/configs, where is the parent of the directory in crop_path; region YAMLs land at {configs root}/dataset//. output overrides the configs root; external_config overrides where the dataset_localization YAML is written (default {configs root}/dataset_localization/). Each is forwarded only when supplied, so the pipeline defaults apply otherwise. When supplied, output and external_config must be absolute paths and, when the MCP client declares roots, inside a declared root. masks is the mask version tag forwarded as --masks; it must match the stage-2 masks (script default canonical_25).

Launch encoder training: train a champollion_V1 self-supervised encoder for one sulcal region.

Dataset configs must exist before calling this tool: run start_config first. When the MCP client declares roots, config_dir defaults to <roots[0]>/data//derivatives/champollion_V1/configs (roots[0] = the first declared root; same base as the output_dir default). Otherwise config_dir defaults to /data//derivatives/champollion_V1/configs ( = the champollion_pipeline root). Pass config_dir when the configs live elsewhere (use //derivatives/champollion_V1/configs) or when start_config was given an explicit output (use that configs root).

Launch Stage 4: compute sulcal embeddings across all 56 model folds (28 regions × 2 hemispheres).

profiling: run the embeddings script under cProfile. no_cache: re-extract a models archive, ignoring the cache. legacy: read crops from derivatives/deep_folding-2025 (overrides cortical_version). use_last_checkpoint: evaluate with the native Lightning checkpoint in version_0/checkpoints/ instead of best_model_weights.pt, for regions that have one.

Launch Stage 5: collect all per-region embedding CSVs into a single output directory.

Launch Stage 6: render sulcal graph meshes, cortical tile masks, and UMAP scatter plots.

UMAP plots need both embeddings_dir and reference_data_dir; there is no default for reference_data_dir (not shipped with champollion_pipeline).

Launch the full Champollion pipeline (all 6 stages sequentially). Returns a pipeline job_id immediately.

reference_data_dir is optional and forwarded to stage 6 (snapshots) for UMAP plots.

A non-empty masks is forwarded to stages 2, 3 and 4, and names the <datasets_root>/derivatives/champollion_V1/<masks or canonical_25>/ directory that stages 5 and 6 read and write (region_embeddings, embeddings, snapshots). overwrite, run_cka, regions, masks_version and cortical_version apply to stage 4 (embeddings) only. regions are model names (e.g. SCsylv_left); cortical_version must name where the stage-2 crops actually are.

Remove all cortical_tiles derivatives for a single subject.

Deletes per-subject NIfTI files (crops, labels, extremities, distbottom), per-subject subdirectories (skeletons/, foldlabels/, transforms/, distmaps/), and filters the subject's row from aggregated .npy arrays and their subject CSVs.

Use dry_run=True to preview what would be deleted without modifying anything.

Remove cortical_tiles outputs for all subjects that failed QC.

Reads a QC TSV/CSV file with 'participant_id' and 'qc' columns and deletes all files belonging to subjects with qc==0 or absent from the QC file. Equivalent to having run cortical_tiles with --sk_qc_path from the start.

Use dry_run=True to preview what would be deleted without modifying anything.

Get the current status and progress of a running or completed job.

List all jobs for a given output directory, optionally filtered by status.

Cancel a running job by sending SIGTERM to its process.

Retrieve the last N lines of a job's log output.

Get metadata about the Champollion pipeline and available MCP tools.

Check whether the Champollion pipeline is correctly configured and accessible.