Use the supported Python boundary

Import mapf.api to compose a study without depending on the GUI or private persistence methods. API version 1 supports the existing planning, execution, storage and export contracts.

from mapf.api import (
    API_VERSION, MAPFInstance, Point, SimulationConfig,
    check_solver, get_solver,
)

instance = MAPFInstance(
    grid_width=3, grid_height=3,
    starts={"a": Point(0, 0)}, goals={"a": Point(2, 0)},
)
receipt = check_solver(get_solver("CBS"), instance,
                       SimulationConfig(grid_width=3, grid_height=3))
assert receipt["status"] == "valid_solution"

This check executes one tiny fixture synchronously. A legitimate unsuccessful result with no candidate paths is not_checked, not a valid prefix or proof of infeasibility. It checks input mutation, solver identity, physical trajectories, successful claims and canonical costs. It does not impose a process deadline or prove optimality/completeness. For expensive or untrusted extensions, use ExperimentService and its process supervisor with declared budgets.

Public surface	Responsibility
MAPFInstance, MAPFSolution, MAPFSolverProtocol, Point, Path, SimulationConfig, SimulationSetting, CommitmentType	Domain inputs and results
register_solver, get_solver	Named factory registration and construction
JobSubmissionRequest, compile_experiment, verify_manifest	Validated requests, frozen plans and identity checks
RunRepository, ExperimentService	Workspace reads, supervised execution and explicit resume
create_study, experiment_card	Safe project initialization and all-outcome metadata cards
export_bundle, check_run_bundle, check_solver	Portable evidence and finite conformance witnesses

Register top-level factories in an importable module; multiprocessing spawn must execute that registration in each worker. Protect execution with if __name__ == "__main__". Never define a new algorithm merely to test the extension boundary. The registered CBS example wraps an existing solver unchanged and demonstrates worker composition, validation and export.

check_run_bundle(bundle) uses the production versioned importer in a disposable workspace, then reads the saved result. It rejects unknown formats, changed checksums and inconsistent trajectories/frames/metrics. Integrity is not source authentication. Keep the original bundle; an import receives a new local run ID with its ancestry retained.

Only documented members are stable. Underscore-prefixed methods, internal SQLite tables and implementation classes remain internal. Additive releases preserve this public surface; breaking changes require a new API version and migration notes. Existing module imports remain available in this release.

Next: run the installed extension walkthrough, then consult version compatibility.