An example to sort a list of jobs depending on their depends.

sort_pipeline_example.py

from typing import Dict, List, Optional
import networkx as nx


def main():
    stages = {
        "Lint": [],
        "Test": [],
        "Coverage": ["Test"],
        "Docs": ["Coverage", "Lint"],
        "Benchmark": ["Coverage"],
    }

    print(list(_sort_jobs(stages)))
    # Output: [['Lint', 'Test'], ['Coverage'], ['Docs', 'Benchmark']]


def _sort_jobs(dependencies: Dict[str, List[str]]) -> List[List[str]]:
    """
    _sort_jobs is taking a dependency dict and sort the jobs to be able to have parallel run.
    :param dependencies: A dependency dict that contains every stages and their dependencies
                         something like { "1": [], "2": [], "3": ["2", "1", "4"], "4": ["2", "1"] }
    :return:  The order of the stages
              example: [["1", "2"], ["4"], ["3"]]
    """
    g = nx.DiGraph(dependencies)

    # detect cycling workflows
    cycles = list(nx.simple_cycles(g))
    if len(cycles) > 0:
        raise CyclingPipeline(cycles=cycles)

    # sort the stages
    out_degree_map = {v: d for v, d in g.out_degree() if d > 0}
    zero_out_degree = [v for v, d in g.out_degree() if d == 0]
    while zero_out_degree:
        yield zero_out_degree
        new_zero_out_degree = []
        for v in zero_out_degree:
            for child, _ in g.in_edges(v):
                out_degree_map[child] -= 1
                if not out_degree_map[child]:
                    new_zero_out_degree.append(child)
        zero_out_degree = new_zero_out_degree


class CyclingPipeline(Exception):
    """
    CyclingPipeline is an exception raised if we detect a cycle in the pipeline.
    """
    def __init__(self, cycles=Optional[List]):
        message = f"Invalid workflow you have a cycling dependencies: {cycles}"
        super().__init__(message)


if __name__ == "__main__":
    main()

Hi good stuff! Comparing the initial dependencies vs the output ([['Lint', 'Test'], ['Coverage'], ['Docs', 'Benchmark']]), they are not exactly the same. In the output, Coverage has a dependency on both Lint and Test. Or maybe there's no way to represent accurately your sample graph in a list.