Remove MiMo V2 Pro (nous) provider from all wizard configs — it was added
during the evaluation attempt (#447) and "config-murdered" the fleet.
Restore the canonical golden state provider chain:
Kimi K2.5 → Gemini 2.5 Pro (OpenRouter) → Ollama gemma4
Changes:
- Create wizards/timmy/config.yaml (was missing — Timmy resurrected)
- Update wizards/allegro/config.yaml: strip nous, normalize to golden state
- Update wizards/ezra/config.yaml: strip nous, preserve max_turns: 90
- Update wizards/bezalel/config.yaml: strip nous, add openrouter+ollama,
preserve custom telegram/webhook, personality kawaii, and session_reset
- All wizards now have no Anthropic references and correct provider chain
Acceptance criteria met:
- [x] All wizards resurrected from checked-in configs (Timmy created, others cleaned)
- [x] Provider chain verified: Kimi K2.5 → Gemini 2.5 Pro → Ollama gemma4
- [x] No Anthropic/nous/mimo references in any running config
- [ ] request_log telemetry (handled by thin_config Ansible, blocking dep done)
- [ ] Ezra Telegram token propagation (infrastructure, out of scope for this PR)
- [ ] Duplicate agents resolution (separate fleet audit issue, explicitly non-blocking)
Closes#448
Add 'nous' provider configuration to all three wizard houses (Allegro,
Ezra, Bezalel) with MiMo V2 Pro model (xiaomi/mimo-v2-pro) for free
tier evaluation on Nous Portal.
- Adds nous provider to fallback_providers chain (after ollama)
- Model: xiaomi/mimo-v2-pro with 131K context window
- Endpoint: https://inference.nousresearch.com/v1
- API key env: NOUS_API_KEY
- Timeout: 120s
- Reason documented for evaluation tracking (#447)
This is the smallest concrete fix to enable Nous Portal free tier
evaluation. Actual latency/reliability comparison will be captured
through request_log telemetry in subsequent evaluation runs.
Closes#447
- Adds fallback_providers with kimi-coding:kimi-k2.5 as fallback1
- Followed by anthropic and openrouter fallbacks
- Aligns with #lazzyPit epic for automated resilience
