skin-tone-detection

---

name: skin-tone-detection description: Classifies patient skin tone using Fitzpatrick/Monk scale with Gemini 3 Flash and outputs confidence score license: MIT compatibility: opencode metadata: audience: developers workflow: clinical-pipeline

What I do

I classify patient skin tone using the Fitzpatrick (I-VI) and Monk (1-10) scales. I use Gemini 3 Flash for classification and return a confidence score that determines downstream calibration decisions.

When to use me

Use this when:

• An image has been verified and needs skin tone classification
• You need to determine the appropriate fairness calibration threshold
• Confidence score is needed for safety routing decisions

Key Concepts

• Fitzpatrick Scale: I-VI classification (very fair to darkest)
• Monk Scale: 1-10 numerical skin tone representation
• Confidence Score: 0-1 float determining if standard or safety calibration applies
• Threshold: 0.65 - below this triggers safety calibration

Source Files

• services/vision.ts: Vision model integration
• types.ts: FitzpatrickType enum

Code Patterns

• Use Gemini 3 Flash API for classification
• Return confidence with classification result
• Route to Safety-Calibration-Agent if confidence < 0.65

Operational Constraints

• All inference must return confidence score (0-1)
• If low confidence, MUST route through Safety-Calibration-Agent
• Store fitzpatrick_type in WorldState for downstream agents