OpenSciFlow is an early open initiative for building verified execution capsules for AI for Science tools.

OpenSciFlow is not a write-once-run-anywhere system. It does not promise that scientific tools will run across all machines, CUDA versions, Conda environments, HPC modules, Slurm policies, filesystem layouts, or model-weight locations.

Instead, OpenSciFlow makes tool requirements, environment assumptions, verification status, run records, and known failures explicit and machine-readable.

Not write once, run anywhere.
Write once, check before run, record after run.

中文定位：

OpenSciFlow 不承诺让科学工具跨环境自动运行。它要做的是把“工具需要什么、能不能跑、为什么不能跑、在哪里跑通过、这次运行如何复现”结构化、可检查、可诊断、可记录。

OpenSciFlow is not asking the community to adopt a standard immediately.

At this stage, we are asking for corrections:

• Are the capsule fields sufficient?
• Are license, citation, model-weight, and dataset metadata handled correctly?
• Are environment assumptions explicit enough?
• Are local-agent execution rules strict enough?
• What do real HPC / Slurm workflows require?
• Which failures should be recorded as known failure cases?
• What would users misinterpret from current workflow outputs?

If you maintain an AI for Science model, workflow engine, molecular simulation tool, HPC pipeline, package/container system, or reproducibility project, we especially welcome field-level corrections and missing-case reports.

OpenSciFlow is an early protocol-layer effort for check-before-run and record-after-run AI for Science execution.

It currently includes:

• A draft opensciflow.yaml manifest format for describing scientific tools and models.
• A draft verified execution capsule structure that combines manifests, environment specs, reviewed command templates, smoke tests, test inputs, expected outputs, run records, verified environment matrices, and known failure cases.
• Reusable workflow templates that define task structure, artifact handoff, report boundaries, and reproducibility requirements.
• A local-agent execution contract that restricts agents to reviewed command templates or reviewed wrappers.
• An OpenSciFlow Skill draft that teaches agents to inspect capsules, check environment readiness, run smoke tests when available, request approval, and write run records.
• A BioPilot reference prototype for local protein-computing workflows.
• A curated landscape of related AI for Science tools, workflow systems, model hubs, local/HPC execution tools, and reproducibility tools.

• Not a universal AI Scientist.
• Not a guarantee that tools run across all environments.
• Not a replacement for Docker, Conda, Apptainer, Slurm, Nextflow, Snakemake, Galaxy, CWL, AiiDA, Parsl, package managers, model hubs, or domain tools.
• Not merely a README-to-YAML summarizer.
• Not a claim that listed projects are partners or officially affiliated with OpenSciFlow.
• Not a clinical, drug-discovery, or scientific truth-validation system.

A local agent must not execute arbitrary shell commands generated by an LLM.

It should execute only reviewed command templates or reviewed wrapper submissions declared in a capsule, after checking inputs, environment readiness, model-weight availability, license/citation metadata, known failure cases, user approval requirements, and the verified environment matrix.

It must not claim reproducibility beyond the environments where evidence exists.

Early manifests are useful for inspection, but they are not proof that a tool will run.

• R1/R2: may reduce documentation-understanding cost.
• R4/R5: may cautiously reduce trial-and-error cost in the verified environment.
• R6/R7: provide stronger evidence for cross-environment migration or external reproduction.

OpenSciFlow does not eliminate scientific computing failures. It makes them explicit, checkable, diagnosable, and recordable.

As of 2026-07-08, OpenSciFlow is an early public draft.

• 83 projects mapped and assessed across AI for Science agents, workflow engines, model hubs, package/container systems, local/HPC execution, and reproducibility tools.
• 7 example plugin manifests drafted, including Boltz, ProteinMPNN, MACE, DiffDock, MDAnalysis, GROMACS, and ProteinFlux.
• Manifest and wrapper guardrails drafted, including required-field boundaries, license/citation propagation, model-weight metadata checks, placeholder validation, normalized scheduler fields, reviewed-wrapper metadata validation, and disallowed shell-fragment checks.
• 7 workflow templates drafted, with DAG, artifact-handoff, reproducibility-policy validation, and a protein-template review matrix.
• OpenSciFlow Skill drafted, with schemas, prompts, refusal tests, structured examples, Slurm workflow/execution alignment checks, BioPilot run-record crosswalk, and coding-agent behavior review.
• R0-R7 readiness levels proposed to distinguish indexing, draft metadata, schema validation, command/environment availability, smoke tests, example runs, multi-environment verification, and external reproduction.
• BioPilot prototype defined with review-only planning, read-only manifest/workflow artifact resolution, and run-record validation paths.
• Verified Capsules v0.1 alpha productized with an installable opensciflow-capsule CLI, quickstart, release checklist, CLI tests, one narrow mdanalysis-rmsd R6 capsule covering a tiny example across local Windows and GitHub Actions Ubuntu, and one narrow gromacs-rmsd R5 capsule covering a tiny GitHub Actions Ubuntu example. No HPC/GPU/large-trajectory portability claim is made.

• verified-capsules: alpha capsule registry with installable opensciflow-capsule CLI, narrow mdanalysis-rmsd R6 evidence, and narrow gromacs-rmsd R5 evidence.
• docs/reference/standards-crosswalk: how OpenSciFlow should align with CWL, RO-Crate, workflow engines, and MCP instead of replacing them.
• plugin-manifest: opensciflow.yaml manifest draft.
• workflow-templates: reusable AI for Science workflow templates.
• opensciflow-skill: agent-facing skill and protocol adapter layer.
• docs: user and contributor documentation.
• biopilot-prototype: first protein-computing reference prototype.
• awesome-ai4s-workflows: curated project landscape.
• whitepaper: position paper drafts.
• community: governance, outreach, and community process.

• Review the verified execution capsule structure.
• Add a known failure case for a tool or environment.
• Review whether a capsule overclaims its readiness level.
• Correct license, citation, dataset, or model-weight metadata.
• Add smoke-test evidence for one environment.
• Point out missing HPC/Slurm scheduler assumptions.
• Review local-agent refusal behavior.
• Tell us which workflow outputs users are likely to misinterpret.

OpenSciFlow is currently maintained as an early student-led open initiative. We welcome corrections before adoption claims.

We do not claim partnership with listed projects unless explicitly agreed by their maintainers.