Verified 2026-06-05: Provider config confirmed from
/root/.hermes/profiles/tai/config.yaml— default modeldeepseek-v4-flashviaopencode-gowithopenrouteras fallback. Reasoning policy (medium default, xhigh for Rei QA) matches config. OpenRouter configured with free-tier models only. MoA plugin v0.2.0 presets (deepseek, diverse, freebie, budget) confirmed accurate.
Tai MoA & Provider Architecture
Multi-provider Mixture-of-Agents system and overall Hermes provider setup.
Active Provider
- Primary: OpenCode Go (
deepseek-v4-flash) — $10/mo, 14 models - Fallback: OpenRouter (free tier only)
- Format: OpenAI-compatible chat completions
Reasoning Policy
| Context | Effort | Notes |
|---|---|---|
| Global default | medium |
All general chat |
| Rei QA/code-audit | xhigh |
Rei profile only |
| Cron jobs | Free models | OpenCode Go sub-models |
| Daily digest | Free models | OpenRouter free tier |
⚠️ Cost history: Pro + xhigh burned $10 in <12h. OpenRouter = free-tier models ONLY (deepseek-v4-flash:free, gemini-2.0-flash:free, llama-4-maverick:free, qwen3-235b:free). No frontier models on OpenRouter.
Tai MoA Plugin (v0.2.0)
Multi-provider council system. Uses OpenCode Go + OpenRouter together.
| Preset | Models | Use Case |
|---|---|---|
deepseek |
V4 Pro + Kimi K2.6 + GLM-5 → free Flash agg | Default, balanced |
diverse |
4 model families via OpenCode Go | Broader perspectives |
freebie |
All free models | Cost-sensitive |
budget |
Free + OpenCode Go | Cheap-but-capable |
Config at plugins.tai_moa.presets in config.yaml.
Cost Guardrails
- Medium reasoning global default
- xhigh only for Rei QA
- Cron jobs on free/cheapest models
- Non-critical crons paused when cost is a concern
- Actionable cost data preferred over analysis
- OpenRouter wallet-safe: free tiers only
Other Provider Notes
- Kimi key: NOT configured — OpenCode Go serves as the “kimicode” lane
- Brave/Tavily: Not configured for web backend
- Tinker-Atropos: Backburner — only for active RL experiments