OpenRath Examples

Runnable production workflows built with OpenRath.

English | 简体中文

---

OpenRath-Example turns the OpenRath durable runtime into three complete, auditable workflows.

Each example includes fixed inputs, a short Google Colab notebook, versioned workflow source, durable execution state, explicit QA, and a committed reference output. You can inspect the result before providing an API key, then run the same workflow yourself.

Choose an Example

Example	Open in Colab	Durable workflow	Reference output
Literature research	Run notebook	View source	Read report
Presentation creation	Run notebook	View source	Download deck
Excel visualization	Run notebook	View source	Download workbook

OpenRath in These Examples

fixed input
    |
    v
@step / @router workflow
    |
    v
SQLite Run -> Events -> Checkpoints
    |                       |
    |                       +-> Interrupt / resume
    |
    +-> evaluation and QA
    |
    v
content-addressed Artifact + final deliverable

OpenRath capability	Literature	PowerPoint	Excel
Explicit @step workflow	✓	✓	✓
Durable SQLite Run and Checkpoints	✓	✓	✓
Worker restart and lease recovery	✓
Human approval Interrupt and resume		✓
Dynamic quality routing			✓
Evaluation or structural QA gate	✓	✓	✓
Content-addressed Artifact	✓	✓	✓
Masked API-key prompt	✓	✓	✓

These are scenario examples, not replacements for the small API ladder in OpenRath/example/. Use the core ladder to learn one API at a time; use this repository to see the pieces working together.

1. Literature Research

The research workflow reads a fixed paper catalogue and request, screens sources against an allowed topic set, asks DeepSeek for a constrained synthesis, writes a Markdown report, evaluates citation coverage, and registers the result as an Artifact.

It intentionally stops after two steps, closes the SQLite store, expires the worker lease, and resumes from the durable checkpoint with a new runtime. The final assertions verify:

• five committed checkpoints;
• a successful terminal Run;
• an evaluation gate of PASS;
• source identifiers constrained to the supplied paper set;
• a registered artifact:// URI;
• no API key in durable state or the report.

2. Presentation Creation

The presentation workflow drafts a six-slide outline, pauses on a durable human approval Interrupt, resumes after the decision, renders the deck, previews every slide, validates its structure and notes, and registers the .pptx as an Artifact.

The final assertions verify the waiting and queued transitions, six rendered slides, PowerPoint package structure, source notes, a successful Run, and an empty pending-interrupt inbox.

3. Excel Visualization

The Excel workflow profiles a fixed sales dataset, routes invalid data to a quality stop, plans a dashboard, builds the workbook, renders every Sheet for inspection, audits formulas and KPI values, and registers the .xlsx as an Artifact.

The final assertions verify 24 input rows, the passing quality route, four previews, zero formula errors, five checkpoints, a successful Run, and the final Artifact.

Run in Google Colab

1. Open one of the notebooks above.
2. Choose Runtime → Run all.
3. Let the notebook install its pinned public dependencies.
4. Enter a DeepSeek API key in the masked prompt.
5. Inspect the Run status, checkpoints, QA result, and final download link.

The key is collected through a masked prompt and is not stored in the notebook or Git repository. The literature workflow additionally asserts that the key is absent from durable Run state and the final report.

Reproducible Inputs, Variable Language

The notebooks pin:

• the OpenRath package version;
• the support-code commit;
• the input commit and SHA-256;
• output filenames and directory boundaries;
• citation, metric, formula, and structural QA rules.

Live DeepSeek calls use temperature=0.2, so prose can vary between successful runs. The reference outputs show the expected format and quality boundary; they are not byte-for-byte golden files. Offline fixtures remain deterministic.

Repository Layout

OpenRath-Example/
├── examples/
│   ├── 01_literature_research/
│   ├── 02_ppt_creation/
│   └── 03_excel_visualization/
│       ├── input/          # pinned source material
│       ├── output/         # reference deliverable and notes
│       ├── src/            # durable workflow and optional builder
│       └── notebook.ipynb  # short Colab entry point
└── src/openrath_colab/     # shared Colab, DeepSeek, and runtime helpers

Each run gets an append-only directory under the ignored .workspace/ path. Temporary databases, checkpoints, artifacts, previews, and caches stay there; only the named final deliverable is written to output/.

Update Discipline

When changing an example:

1. update and commit its input/ or src/;
2. run the workflow end to end and pass all assertions;
3. inspect the generated report, deck, or workbook;
4. update the committed reference output;
5. pin the new support and input commits in the notebook;
6. verify the notebook contains no output cells or credentials;
7. run the notebook from a clean Colab session.

Never make a notebook depend on a floating support-code commit.

Links

• OpenRath: github.com/Rath-Team/OpenRath
• Documentation: docs.openrath.com
• Core API examples: OpenRath/example/
• Issues: OpenRath issue tracker