Marvel Snap Cybernetic Loop

A fully automated, privacy-preserving pipeline that ingests Marvel Snap screenshots, reconstructs games, analyzes play patterns, and publishes curated public artifacts.

Built by Bryan Chasko and the HeraldStack multi-platform agent architecture.

---

What This Does

Screenshots → Classification → Metadata → Game Records → Analysis → Public Artifacts

You play Marvel Snap. You take screenshots. Everything after that is automated.

The pipeline classifies screenshots, extracts game state (cards, locations, turns, cubes), reconstructs full matches, computes competitive analytics, and publishes anonymized results to your website and GitHub.

Architecture

graph TD
    A[Screenshots] --> B[Local Staging]
    B --> C{Snap?}
    C -->|Yes| D[UI-TARS 1.5 7B]
    C -->|No| E[Archive]
    D --> F[Metadata JSON]
    F --> G[Seed OSS 36B]
    G --> H[Game Record]
    H --> I[Qdrant]
    I --> J[Analysis Engine]
    J --> K[Public Artifacts]
    K --> L[bryanchasko.com]
    K --> M[GitHub]

    style A fill:#1a1a2e,stroke:#e94560,color:#eee
    style D fill:#533483,stroke:#e94560,color:#eee
    style G fill:#533483,stroke:#e94560,color:#eee
    style I fill:#16213e,stroke:#0f3460,color:#eee
    style L fill:#0f3460,stroke:#e94560,color:#eee

Loading

Privacy Model

Not a privacy architecture. Just: don't give ByteDance your Google credentials.

Zone	What Lives Here	Who Touches It
Local	Google creds, raw screenshots	Goose-cli local agents only
Hybrid	Sanitized game screenshots	ByteDance models via OpenRouter
Public	Anonymized game records, analysis	The world

Data flows one direction: local → hybrid → public. Never reversed.

ByteDance Models (via OpenRouter)

Model	Alias	Use Case	Cost
UI-TARS 1.5 7B	bytedance-vision	Screenshot classification, game UI recognition	$0.10/$0.20 per M tokens
Seed OSS 36B	bytedance-reason	Game reconstruction, analysis reasoning	TBD

Daily cost cap: $0.01. That's ~50-100 classifications per day.

Pipeline Layers

Layer	Recipe	What It Does
1. Ingestion	snap-ingest.yaml	Watch local folder, queue new screenshots
2. Classification	snap-classify.yaml	Is this Marvel Snap? (UI-TARS)
3. Reconstruction	snap-reconstruct.yaml	Build turn-by-turn game record (Seed OSS)
4. Analysis	snap-analyze.yaml	Cube efficiency, deck stats, misplays
5. Artifacts	snap-publish.yaml	Generate public-safe content, open PRs
Orchestrator	snap-pipeline.yaml	Run all layers in sequence with CSWR

Spec-Driven Development

This project uses spec-driven development. Specs are the source of truth. Code serves specs.

specs/
  SPEC-INDEX.md                    # Map of all specs
  marvel-snap-pipeline/
    design.md                      # Architecture (Design-First)
    tasks.md                       # Executable tasks (generate-task-issues.py format)
  private-ingestion/
    requirements.md                # EARS notation requirements
  classification-metadata/
    requirements.md
  game-reconstruction/
    requirements.md
  vendor-gateway/
    requirements.md

Requirements use EARS notation (WHEN/THE SYSTEM SHALL) for testability.

CSWR — Conversation-Scoped Work Reference

Every agent action is anchored to a CSWR before it does anything. The CSWR ties conversations to GitHub issues, specs, and sessions.

CSWR = { issue_id, spec_id, conversation_id }

Implemented via prompt-ledger.sh in the gander runtime.

Execution Platform

Runs on the gander — the goose-cli collective of the HeraldStack.

Existing infrastructure leveraged:

• goose-proxy.py — heuristic model dispatcher
• Qdrant — vector store for game records and semantic search
• Valkey — Redis-compatible cache for cost tracking
• Cedar policies — governance (no-secrets, branch-naming)
• Docker hardening — cap_drop, no-new-privileges, resource limits
• 38+ MCP launchers — filesystem, GitHub, vision-server, Qdrant, AWS

Project Governance

• CONSTITUTION.md — 8 immutable architectural principles
• SECURITY.md — public repo rules
• specs/SPEC-INDEX.md — spec map and validation checklists

The Team

Built by the HeraldStack haunting:

Agent	Role
Harald (he/him)	Anchor, coordination
Stratia (she/her)	Architect, recipe design
Ellow	GooseCLI implementation
Myrren	OpenRouter model routing
Kade-Vox	Security
Ralph Wiggum	QA validation

Links

• bryanchasko.com — Website (game dashboard target)
• builder.aws.com/community/@bryanchasko — Blog
• goosecli-heraldstack-gander — Gander runtime
• OpenRouter ByteDance models — Model catalog

License

MIT