Pipeline Guide

autob run executes five stages and writes every intermediate artifact to the output directory.

The deterministic stages are knowledge-aware. The package includes data/synonyms.json, data/domain-patterns.json, data/parameter-constraints.json, and data/risk-catalog.json; a full run loads them once and reuses the same knowledge base across atomization, hypergraph construction, and requirements generation.

Stage 1: Atomize

Command:

autob atomize <sop-file> -o <output-dir>

Output: 01-ops.json

The atomizer splits SOP text into operation records with action, inputs, target, tools, parameters, risks, human judgment, and output state.

Stage 2: Hypergraph

Command:

autob hypergraph <output-dir>/01-ops.json -o <output-dir>

Outputs: 02-nodes.json, 03-hyperedges.json

This stage normalizes operation fields into reusable nodes and creates one operation hyperedge per SOP operation.

Stage 3: Requirements

Command:

autob requirements <output-dir>/02-nodes.json <output-dir>/03-hyperedges.json -o <output-dir>

Output: 04-requirements.json

The generator maps hypergraph evidence into deterministic R1-R10 requirements and deduplicates by fingerprint.

Stage 4: Infer

Command:

autob infer <output-dir>/04-requirements.json -o <output-dir>

Output: updated 04-requirements.json

When JSON config contains a complete LLM setup, this stage asks an OpenAI-compatible provider for candidate requirements. Without config, it leaves deterministic requirements intact and records a clarification.

Stage 5: Review

Command:

autob review <output-dir>/04-requirements.json -o <output-dir>

Outputs: 05-coverage.json, report.md, diagrams/*.mmd

The review stage builds coverage matrices, warnings, diagrams, and an optional interactive candidate review flow. In a full autob run, the pipeline also writes 06-clarifications.json before review artifacts.

Debugging

• If output is missing, rerun the last stage command directly with the previous stage artifact.
• If LLM candidates are absent, run autob config show and verify apiKey, baseUrl, and model are set.
• To confirm whether Stage 4 called an LLM, inspect run-meta.json, 06-clarifications.json, and any LLM-Candidate records in 04-requirements.json.
• If coverage is low, inspect 03-hyperedges.json and 04-requirements.json together.
• For paper or reproducibility work, use the checked-in publication/ sample artifacts and diagrams as a known baseline.
• For codebase navigation, inspect graphify-out/GRAPH_REPORT.md or open graphify-out/graph.html.