---
description: LangGraph AI workflows — LLM nodes, prompt versioning, safety filters, grounding
globs: ["apps/api/**/workflows/**", "packages/agents/**"]
alwaysApply: false
---

# AI Workflows

Reference: `docs/02-architecture/01-SYSTEM_ARCHITECTURE.md`, `apps/api/src/api/workflows/`.

## Rules

- Every AI flow change includes **prompt/version metadata** (version id, model, changelog note).
- Nodes fail fast with partial user-safe output — no silent failures.
- Run safety post-filters on all patient-facing text; never diagnose or prescribe.
- Ground outputs in uploaded report data; no internet-generic reference ranges.
- LLM access only through provider interface (`providers/llm/`); resolve via factories.
- Log `request_id` / `job_id`; never log raw report text or identifiers.

Skill: `.cursor/skills/ai-workflows/SKILL.md`