Overview

KWDagger turns parameterized definitions of existing command-line programs into static, inspectable graphs of shell commands and hashed result directories. It builds on cmd_queue and kwconf to provide:

• Reusable kwdagger.pipeline.Pipeline and kwdagger.pipeline.ProcessNode abstractions for constructing commands and wiring produced artifacts.

• Parameter-matrix expansion with operational deduplication of equivalent requested work.

• Per-process invoke.sh files and a navigable .pred / .succ symlink graph, so generated work can be inspected, rerun, or invalidated without remaining inside the kwdagger runtime.

• A scheduling CLI (kwdagger.schedule) that hands the static command graph to serial, tmux, or Slurm cmd_queue backends. The tmux backend is the most commonly used interactive runner.

• An optional aggregation CLI (kwdagger.aggregate) that loads completed results, requested and resolved parameters, and metrics into analytical tables and reports.

• A self-contained demo pipeline in kwdagger.demo.demodata that is used in CI and serves as a reference implementation.

• examples/yaml_introspection_example.py is a single-file modal CLI with three tiny pipeline stages, an introspect command that queries their declarative and compiled graphs, and a run command that executes the same YAML pipeline locally with the serial backend.

Kwdagger wraps ordinary scripts rather than replacing them. A node may use the default named-argument command convention or subclass ProcessNode to support an existing positional or otherwise specialized CLI.

Repository layout

• kwdagger/pipeline/ – core pipeline and process node definitions, networkx graph construction, and configuration utilities. Import from kwdagger.pipeline; the submodules inside it are private.

• kwdagger/schedule.py – ScheduleEvaluationConfig CLI for expanding parameter grids into runnable jobs and dispatching them through cmd_queue backends.

• kwdagger/aggregate.py – AggregateEvaluationConfig CLI for loading job outputs, computing parameter hash IDs, and generating text/plot reports.

• kwdagger/demo/demodata.py – end-to-end demo pipeline with prediction and evaluation stages plus CLI entry points for each node.

• docs/ – Sphinx sources, including an example user module under docs/source/manual/tutorials/twostage_pipeline.

• tests/ – unit and functional coverage for pipeline wiring, scheduler behavior, aggregation, and import sanity checks.

Quickstart

Run the demo pipeline locally to see the CLI workflow end-to-end:

TMP_DPATH=$(mktemp -d --suffix "-kwdagger-demo")
cd "$TMP_DPATH"
echo "demo" > input.txt

EVAL_DPATH=$PWD/pipeline_output
python -m kwdagger.schedule \
    --params="
        pipeline: 'kwdagger.demo.demodata.my_demo_pipeline()'
        matrix:
            stage1_predict.src_fpath:
                - input.txt
            stage1_predict.param1:
                - 123
            stage1_evaluate.workers: 2
    " \
    --root_dpath="${EVAL_DPATH}" \
    --backend=serial --skip_existing=1 --run=1

python -m kwdagger.aggregate \
    --pipeline='kwdagger.demo.demodata.my_demo_pipeline()' \
    --target "
        - $EVAL_DPATH
    " \
    --output_dpath="$EVAL_DPATH/full_aggregate" \
    --eval_nodes="
        - stage1_evaluate
    " \
    --stdout_report="
        top_k: 10
        concise: 1
    "

The scheduler generates per-node job directories with invoke.sh and job_config.json metadata. Each invoke.sh is intended to be a usable recomputation command even when kwdagger is not involved in the rerun. The aggregator is then one optional way to consolidate results and print a report.

The hashed result tree contains a graph-based symlink structure for navigating produced-artifact dependencies. The .succ folder links to results that depend on the current result, and .pred links to results consumed by the current process. This makes downstream inspection and invalidation possible after the original scheduling command has finished.

For more in-depth information see tutorials:

• Tutorial #1: twostage_pipeline

• Tutorial #2: ollama_benchmark

Command line entry points

• python -m kwdagger.schedule or kwdagger schedule – build and run a pipeline over a parameter matrix (see kwdagger.schedule.ScheduleEvaluationConfig).

• python -m kwdagger.aggregate or kwdagger aggregate – load completed runs and generate tabular and plotted summaries (kwdagger.aggregate.AggregateEvaluationConfig).

• python -m kwdagger – modal CLI that exposes the schedule and aggregate commands via kwdagger.__main__.KWDaggerModal.