πŸ’Έ Model Strategy (Hermes + Paperclip)

Decision date: 2026-07-30 Owner: Pat (cost mandate)

The Stack

Use CaseModelProviderCost
Primary (Hermes daily)glm-5.2z.aiLowest
FallbackMinimax M3OpenRouter$0.30/M tokens
Heavy reasoningGPT-5.6 LunaOpenRouterModerate
Heavy codingGPT-5.6 TerraOpenRouterModerate
Claude (special)Sonnet 5 / Opus 5 / Haiku 4.5OpenRouterVariable
Sprint bookendsClaude Opus 5OpenRouter$1-3/run

Hard Rules

  1. NO x.ai as fallback. Explicitly prohibited by Pat.
  2. Cost-conscious always. β€œI don’t need a 10,000 hospital bill.”
  3. Minimax M3 is multimodal and well-rounded β€” good enough for most Paperclip agent work.
  4. Opus 5 only for deep thinking at sprint start/end bookends.

Hermes Aliases

  • /model heavy β†’ GPT-5.6 Luna
  • /model coding β†’ GPT-5.6 Terra
  • /model cheap β†’ Minimax M3
  • /model claude-heavy β†’ Claude Opus 5
  • /model claude-reasoning β†’ Claude Sonnet 5
  • /model claude-fast β†’ Claude Haiku 4.5
  • /model sonnet β†’ Claude Sonnet 5

Paperclip Agent Models (OpenCode)

All Paperclip agents using opencode_local adapter are now set to:

  • Model: openrouter/deepseek/deepseek-v4-flash-0731
  • OpenCode version: 1.18.11
  • OpenCode global config: /root/.config/opencode/opencode.jsonc
  • Default model for ad-hoc runs: same deepseek v4 flash 0731

Agent working directories:

  • CEO, Chief of Staff β†’ /root/repos/brain
  • Huible Tech Lead, Clinical Advisor β†’ /root/repos/huible
  • InvestInMe PM, R&D Lead, Project Builder β†’ /root/repos/investinme

All agents set to idle status, errors cleared. Ready for wipe/rebuild.