Create a2a sin agent

👑 ULTIMATE A2A SIN Factory — Agents + Teams + Coders

OpenCode mirror: sourced from ~/.config/opencode/skills/create-a2a-sin-agent Dieser Skill ersetzt: /create-a2a, /create-a2a-team, /create-a2a-sin-coder

---

📋 Was dieser Skill kann

Erstelle	Zweck	Workflow
SIN Agent	Einzelner A2A Agent (z.B. A2A-SIN-Google-Docs)	Template-SIN-Agent
SIN Team	Team Manager (z.B. Team-SIN-Survey)	Template-SIN-Team
SIN Coder	Entwickler-Agent (z.B. A2A-SIN-Frontend)	Template-SIN-Agent + Coder-Mandate

Alle drei nutzen jetzt EIN Template: Template-SIN-Agent (fusioniert aus Template-SIN-Agent, Template-SIN-Agent-Worker, Template-SIN-Worker).

---

🧠 Agent Configuration System (v5)

Jeder neue Agent wird dem richtigen A2A Team zugeordnet und erbt dessen Modell-Konfiguration:

Team-Register (oh-my-sin.json)

Team	Manager	Config-Datei	Primaer-Modell
Team Coding	A2A-SIN-Zeus	my-sin-team-code.json	claude-sonnet-4-6
Team Worker	A2A-SIN-Team-Worker	my-sin-team-worker.json	gemini-3-flash
Team Infrastructure	A2A-SIN-Team-Infrastructure	my-sin-team-infrastructure.json	gpt-5.4

Subagenten-Modelle (oh-my-openagent.json)

Subagent	Modell	Fallback-Kette
explore	nvidia-nim/stepfun-ai/step-3.5-flash	gemini-3-flash → gpt-5.4 → gemini-3.1-pro → claude-sonnet → qwen
librarian	nvidia-nim/stepfun-ai/step-3.5-flash	gemini-3-flash → gpt-5.4 → gemini-3.1-pro → claude-sonnet → qwen

PARALLEL-EXPLORATION MANDATE (PRIORITY -4.5)

Bei grossen Codebases (100k+ Zeilen) MUSS der Agent 5-10 parallele explore + 5-10 librarian-Agenten starten. Ein einzelner Agent liefert nur ~20% Abdeckung.

---

🛠️ Workflow 1: SIN Agent erstellen

# 1. Spec erstellen
cp agent-spec.example.json my-agent-spec.json
# Bearbeite: name, slug, team, manager, description, capabilities

# 2. Agent generieren
node /Users/jeremy/dev/OpenSIN-backend/scripts/create-sin-a2a-agent.mjs --spec /abs/path/to/my-agent-spec.json

# 3. Dashboard Page scaffolen
node /Users/jeremy/dev/OpenSIN-backend/scripts/scaffold-a2a-agent-page.mjs --slug sin-<agentname>

# 4. Consumer Auth einrichten
~/.config/opencode/skills/create-a2a-sin-agent/scripts/setup_consumer_auth.sh /abs/path/to/new-agent-root

# 5. MCP Surface generieren
MCP_SCRIPTS=~/.config/opencode/skills/create-a2a-mcp/scripts
node $MCP_SCRIPTS/mcp-scaffold.mjs --agent-root /abs/path/to/new-agent-root --slug sin-<agentname>

# 6. Dependencies pruefen
node ~/.config/opencode/skills/create-a2a-sin-agent/scripts/a2a-preflight.mjs --agent-root /abs/path/to/new-agent-root --targets deps,install,assets,hf

Pflicht-Dateien (jeder SIN Agent)

Datei	Zweck
agent.json	Identitaet, capabilities, marketplace object
A2A-CARD.md	Agent Card fuer Discovery
mcp-config.json	MCP Server Konfiguration
clients/opencode-mcp.json	OpenCode MCP Client
src/runtime.ts	Runtime mit opencode CLI Integration
src/mcp-server.ts	MCP Server Oberflaeche
src/a2a-http.ts	A2A HTTP Endpoint
src/metadata.ts	Metadata mit marketplace object
src/cli.ts	Native CLI Entry Point
scripts/hf_pull_script.py	Token Pull Script (Producer-Consumer)
scripts/complete-install.sh	Idempotenter Installations-Script
governance/repo-governance.json	Repo Governance
governance/pr-watcher.json	PR Watcher Contract
platforms/registry.json	Platform Registry
n8n-workflows/inbound-intake.json	Inbound Work Intake
docs/03_ops/inbound-intake.md	Inbound Intake Doku
scripts/watch-pr-feedback.sh	PR Feedback Watcher

---

🛠️ Workflow 2: SIN Team erstellen

# Team generieren (CLI-first, 90% weniger Tokens!)
~/.config/opencode/skills/create-a2a-sin-agent/scripts/generate-team.sh \
  sin-team-<name> \
  "Team - <Name>" \
  "SIN-Team-<Name>" \
  "Beschreibung des Teams"

Team Manager Pflicht-Dateien

Datei	Zweck
agent.json	Team Manager Identitaet
A2A-CARD.md	Team Card
src/manager.ts	Hermes Dispatch Logik
src/cli.ts	Native CLI

---

🛠️ Workflow 3: SIN Coder erstellen

Coder Agents nutzen denselben Workflow wie SIN Agents, ZUSAETZLICH mit:

Coder-spezifische Mandate

1. GitHub App Bot Lane: githubBot.enabled=true mit eigenem Bot-Persona
2. LangGraph.js Supervisor: Multi-Agent Graph mit CodeGenerationNode, TestingNode, IntegrationNode
3. Master Boss Coder Prompt: CEO-level Identitaet mit 10 Millionen Developer-Wissen
4. Browser + GUI Testing: sinInChrome + sin-computer-use Integration
5. 7-Layer Security Hardening: Pre-commit/pre-push hooks, leak prevention CI
6. Inbound-Work + PR-Watcher Contract: Ab Tag 1

Coder-spezifische Dateien

Datei	Zweck
config/github-app-routing.example.json	GitHub App Bot Routing
governance/coder-dispatch-matrix.json	Coder Spezialisierungs-Matrix
.githooks/pre-commit	Secret + External Code Detection
.githooks/pre-push	Commit Message Leak Detection
.secrets.baseline	Detect-Secrets Baseline
.github/workflows/leak-prevention.yml	CI Security Scanning

---

🚨 LLM-AUFRUF PFLICHT-ARCHITEKTUR

JEDER A2A Agent ruft LLMs AUSSCHLIESSLICH ueber die opencode CLI auf:

import subprocess, json

def call_llm(prompt: str, timeout: int = 120) -> str:
    result = subprocess.run(
        ["opencode", "run", prompt, "--format", "json"],
        capture_output=True, text=True, timeout=timeout,
    )
    parts = []
    for line in result.stdout.splitlines():
        try:
            ev = json.loads(line)
            if ev.get("type") == "text":
                parts.append(ev.get("part", {}).get("text", ""))
        except json.JSONDecodeError:
            pass
    return "".join(parts).strip()

REGELN — SOFORTIGER BAN BEI VERSTOSS:

• opencode run --format json = EINZIG ERLAUBT
• OCI-Proxy direkt = VERBOTEN
• Gemini API direkt = PERMANENT VERBOTEN
• Anthropic API direkt = PERMANENT VERBOTEN
• Modell: openai/gpt-5.4 + --fallback nvidia/minimaxai/minimax-m2.7

---

🏗️ Unified Template: Template-SIN-Agent

Drei Templates wurden zu EINEM fusioniert:

• Template-SIN-Agent (Basis)
• Template-SIN-Agent-Worker (Worker)
• Template-SIN-Worker (Minimal)

→ Jetzt: Template-SIN-Agent (all-in-one)

Jeder neue Agent nutzt dieses eine Template und konfiguriert sich via agent.json:

• type: "agent" — Standard Agent
• type: "worker" — Worker Agent (minimal)
• type: "coder" — Coder Agent (mit LangGraph + Security)
• type: "team-manager" — Team Manager (mit Hermes Dispatch)

---

🛡️ Required Validations

Vor "complete" MUessen ALLE bestehen:

npm --prefix /abs/path/to/agent-root run build
node /abs/path/to/agent-root/dist/src/cli.js run-action '{"action":"agent.help"}'
node /abs/path/to/agent-root/dist/src/cli.js print-card
cd /Users/jeremy/dev/OpenSIN-backend && npm run test:a2a:fleet
cd /Users/jeremy/dev/OpenSIN-backend && npm run sync:a2a:control-plane-projection
node ~/.config/opencode/skills/create-a2a-sin-agent/scripts/a2a-verify-hf-vm-readiness.mjs \
  --agent-root /abs/path/to/agent-root \
  --health-url https://<public>/health

---

📦 Vollstaendige Dokumentation

→ Agent Configuration Guide

---

Letzte Aktualisierung: 2026-04-14 | ULTIMATE SKILL v1.0