Home

The CesiumJS Skills wiki is generated from source-controlled Markdown in the wiki/ directory of the CesiumGS/cesiumjs-skills repository. Changes should be proposed through pull requests to the source repository rather than edited directly in the GitHub wiki.

What This Wiki Covers

This wiki currently focuses on the Cesium AI Evaluation Framework: the architecture, decisions, and operating model for evaluating AI-assisted Cesium development workflows.

The framework starts with CesiumGS/cesiumjs-skills, where skill documents guide coding agents that generate CesiumJS code. The same approach is designed to grow toward MCP/tool-call evaluations, Cesium ion workflow evaluations, and broader benchmarks for AI-assisted geospatial development.

Start Here

1. Architecture Concept Document - Goals, constraints, building blocks, runtime views, decisions, risks, and glossary.
2. Run Skill Evaluations Locally - How to validate manifests and reproduce selected browser scenarios.
3. Add an Evaluation Scenario - How to add public-safe scenario manifests.
4. Public Artifact Policy - Which eval artifacts can be committed publicly.
5. ADR-0001: Skills-First Sequencing - Why the initial implementation starts in cesiumjs-skills.
6. ADR-0002: Pairwise Judge Protocol - Why visual evaluation uses A/B/TIE comparison with three independent judges.
7. ADR-0003: Deterministic Decision Policy - How gates, verdicts, and report scores are separated.

Architecture Overview

flowchart TD
    Scenario["Scenario manifest<br/>prompt, expectations, checks"]
    Candidate["Candidate context<br/>skill, prompt, model, tool config"]
    Baseline["Baseline context<br/>current best or release target"]
    Adapter["Agent or tool adapter"]
    Runner["Evaluation runner"]
    Environment["Browser-backed<br/>Cesium environment"]
    Evidence["Evidence bundle<br/>metadata, logs, scene state, screenshots"]
    Checks["Deterministic checks<br/>schema, API, runtime gates"]
    Judges["Pairwise judges<br/>BASELINE / CANDIDATE / TIE"]
    Decision["Decision engine<br/>keep, discard, or review"]
    Report["Report and history"]
    Summary["Public-safe summary<br/>sanitized results and curated artifacts"]
    Analysis["Regression analysis<br/>coverage gaps, failure patterns"]
    Proposal["Candidate update<br/>skill, docs, examples, tools"]
    Review["Maintainer review<br/>accept, revise, or reject"]

    Scenario -->|defines task| Runner
    Candidate -->|candidate under test| Adapter
    Baseline -->|comparison target| Adapter
    Adapter -->|generates output| Runner
    Runner -->|executes in| Environment
    Environment -->|captures| Evidence
    Evidence -->|programmatic facts| Checks
    Evidence -->|visual and qualitative evidence| Judges
    Checks -->|gates| Decision
    Judges -->|majority verdict| Decision
    Decision -->|records outcome| Report
    Report -->|publishes| Summary
    Report -.->|drives next iteration| Analysis
    Summary -.->|informs maintainers| Analysis
    Analysis -.-> Proposal
    Proposal -.-> Review
    Review -.->|updates context| Candidate
    Review -.->|adds or refines coverage| Scenario

    classDef input fill:#e8f3ff,stroke:#2563eb,color:#0f172a
    classDef execute fill:#ecfdf3,stroke:#16a34a,color:#0f172a
    classDef evidence fill:#fff7ed,stroke:#f97316,color:#0f172a
    classDef policy fill:#f5f3ff,stroke:#7c3aed,color:#0f172a
    classDef output fill:#f8fafc,stroke:#475569,color:#0f172a
    classDef iterate fill:#fff1f2,stroke:#e11d48,color:#0f172a
    class Scenario,Candidate,Baseline input
    class Adapter,Runner,Environment execute
    class Evidence evidence
    class Checks,Judges,Decision policy
    class Report,Summary output
    class Analysis,Proposal,Review iterate

Loading

External Links

• Source repository
• CesiumJS
• CesiumJS documentation
• Model Context Protocol