bpmn-generator

BPMN Generator Skill v2.0 — Enterprise Edition

Converts natural language process descriptions into OMG BPMN 2.0.2 compliant XML files and SVG previews via a 4-phase pipeline. All visual rendering follows the bpmn-js reference implementation.

Pipeline Overview

User Text
   ↓  [Phase 1] Intent Extraction  (Claude LLM)
JSON Logic-Core
   ↓  [Phase 2] Validation          (rules + deadlock detection + structural soundness)
Validated JSON
   ↓  [Phase 3] Auto-Layout         (ElkJS Sugiyama layered algorithm)
JSON + Coordinates (edge endpoints clipped to shape boundaries)
   ↓  [Phase 4] Serialization       (pipeline.js)
BPMN 2.0 XML + SVG

The LLM never handles coordinates. Layout is 100% algorithmic.

---

Modes: Document (IST) vs. Optimize (Soll)

Two distinct intents — keep them separate:

• Document mode (default, IST / as-is): the user describes a process and wants it captured faithfully as BPMN. No judgment, no improvement suggestions. This is the default for every entry point (CLI, runPipeline, HTTP, MCP).
• Optimize mode (Soll / to-be): the user wants a better process. Enables the opt-in Optimization Advisory layer, which flags graph-detectable redesign opportunities (Reijers 2005 heuristics + BABOK Lean metrics) as non-blocking advisories — never auto-applied.

Select the mode consistently across entry points:

• CLI: node bpmn/pipeline.js in.json out --optimize
• Programmatic: runPipeline(lc, { mode: 'optimize' })
• HTTP: { "logicCore": {...}, "mode": "optimize" } on /api/v1/generate|validate|orchestrate
• MCP: mode: "optimize" on generate_bpmn / validate_bpmn / orchestrate_bpmn

Advisories are review suggestions with trade-off tags (time/cost/quality/flexibility); they are heuristics, not proofs — present them as options, never silently apply them. Each advisory is an object { id, transform, targets, message, tradeoff, ref, judgment } (see references/api-reference.md); message is the human-readable line, transform names the matching intervention in the toolbox below.