autoresearchclaw-integration-architecture-decisions

Confirmed integration design for AutoResearchClaw into Oracle: (1) Dedicated researchclaw container on oracle-network, NOT inside oracle-hermes, to isolate LLM-generated experiment code execution. (2) MCP Server invocation pattern (~10 tools) chosen over CLI or ACP because it fits the oracle-* native tooling architecture and returns structured JSON tool responses. (3) Hybrid LLM config: Opus 4.6 for the 8 critical reasoning stages (hypothesis gen, paper writing, peer review), GPT-4o-mini or Gemini Flash for the 15 routine stages — AutoResearchClaw natively supports primary+fallback model config.