OpenSciFlow Plugin Manifest

Draft standard for describing AI for Science tools, models, environments, execution commands, HPC support, validation tests, licenses, citations, safety notes, and known limitations.

Recommended manifest filename:

opensciflow.yaml

Status

Draft v0.1. This repository is intended for public review and correction, not production deployment.

Design goals

• Human-readable.
• Lightweight v0.1 fields.
• Local/HPC execution aware.
• Explicit license and citation.
• Validation before execution.
• No arbitrary LLM-generated shell execution.
• Compatible with Conda, Docker, Apptainer, Slurm, and existing model/tool ecosystems.

Required fields

Field	Purpose
schema_version	Manifest schema version, currently opensciflow.plugin/v0.1
name	Stable plugin identifier
version	Manifest version
domain	Scientific domain tags
task_types	Supported task types
description	Short human-readable summary
authors	Manifest/tool authors or maintainers
license	Software, model weight, and data license metadata
citation	Preferred citation, DOI, or URL
inputs	Required/optional inputs and formats
outputs	Expected outputs and formats
environment	Conda/Docker/Apptainer/system requirements
hardware	CPU, memory, GPU, and accelerator requirements
model_weights	Whether weights are required and where to obtain them
execution	Local and optional Slurm execution commands
expected_runtime	Small example and typical runtime estimates
examples	Example inputs and expected outputs
validation	Schema, dry-run, and smoke-test rules
safety_notes	Execution and scientific safety notes
limitations	Known scientific or technical limitations

Minimal example

schema_version: "opensciflow.plugin/v0.1"
name: "example-tool"
version: "0.1.0"
domain: ["computational-biology"]
task_types: ["trajectory-analysis"]
description: "Example plugin manifest for an AI for Science command-line tool."
authors:
  - name: "Tool Maintainers"
license:
  software: "MIT"
  model_weights: "not-applicable"
  data: "CC-BY-4.0"
citation:
  preferred: "Tool Maintainers. Example Tool. 2026."
inputs:
  - name: "structure"
    type: "file"
    formats: ["pdb"]
    required: true
outputs:
  - name: "metrics"
    type: "file"
    formats: ["csv"]
environment:
  conda:
    channels: ["conda-forge"]
    packages: ["python>=3.10"]
hardware:
  cpu_cores_min: 2
  memory_gb_min: 4
  gpu: false
model_weights:
  required: false
  sources: []
execution:
  local:
    command: "example-tool analyze --structure {inputs.structure} --out {outputs_dir}"
expected_runtime:
  small_example: "1-5 min"
examples:
  - name: "small example"
validation:
  dry_run_command: "example-tool --help"
safety_notes:
  - "Review commands before execution."
limitations:
  - "Accuracy depends on input quality."

Validation levels

Validation levels are documented in:

docs/validation-levels.md

Short version:

• V1: schema validation.
• V2: metadata validation.
• V3: command-template validation.
• V4: dry-run validation.
• V5: smoke-test validation.
• V6: run-record validation.
• V7: workflow validation.

Required and optional manifest fields are documented in:

docs/required-vs-optional-fields.md

Review process

Use docs/manifest-review-checklist.md before moving a manifest beyond schema validation.

The short rule is:

Famous project is not enough. A manifest advances only when the required evidence exists.

For protocol-level changes, use docs/protocol-change-process.md.

Command template rules:

• docs/command-template-rules.md

Installation policy

v0.1 should not auto-install arbitrary code without explicit user approval.

Recommended flow:

1. Read manifest.
2. Show dependencies, license, hardware, model weights, and command templates.
3. Let user choose Conda, Docker, Apptainer, or existing system install.
4. Run environment dry-run.
5. Mark status as ready, missing, partial, or unsupported.
6. Never download model weights without showing source, license, and checksum.

Contents

• schema/opensciflow-plugin.schema.json
• examples/mdanalysis-trajectory-analysis/opensciflow.yaml
• examples/mdanalysis-trajectory-analysis/readiness.md
• examples/mdanalysis-trajectory-analysis/dry-run-attempt-2026-07-06.md
• examples/gromacs-md-stability/opensciflow.yaml
• examples/proteinflux-dynamics/opensciflow.yaml
• examples/diffdock-docking/opensciflow.yaml
• examples/boltz-structure-prediction/opensciflow.yaml
• examples/proteinmpnn-sequence-design/opensciflow.yaml
• examples/mace-interatomic-potential/opensciflow.yaml
• docs/diffdock-review-notes.md
• docs/model-manifest-review-notes.md
• docs/model-manifest-backlog.md
• docs/manifest-review-checklist.md
• docs/protocol-change-process.md
• docs/command-template-rules.md
• docs/hpc-slurm-metadata.md
• docs/reviewed-wrapper-fields.md
• docs/required-vs-optional-fields.md
• docs/r3-evidence-template.md
• docs/r4-smoke-test-evidence-template.md
• docs/validation-levels.md
• docs/protocol-roadmap.md
• docs/readiness-levels.md
• docs/agent-contract.md
• docs/design-principles.md
• docs/license-and-citation.md
• docs/security-model.md
• tests/command-rendering-fixtures.json
• ISSUES.md
• MILESTONES.md

Non-goals

• Replacing upstream documentation.
• Replacing package managers such as Conda, Spack, or Bioconda.
• Replacing workflow engines such as Nextflow, Snakemake, CWL, AiiDA, or Parsl.
• Certifying scientific correctness of a model/tool.