Gemini Routing Cutover

Date: 2026-06-29 Status: Implemented

Decision

Chowmes should stop relying on OpenRouter as the active provider route and use the direct Gemini API through GEMINI_API_KEY.

Live Routing

• Default Athena / Hermes model: gemini-2.5-flash
• Default provider: gemini
• Deep / judge / boardroom / delegation model: gemini-2.5-pro
• Context length override: 131072
• CI synthesis provider: gemini
• CI synthesis model: gemini-2.5-flash

Profile Coverage

The following Hermes profile configs were updated to Gemini and given access to the existing GEMINI_API_KEY:

• Default / Athena
• Vulcan
• Arjuna
• Kubera
• Prometheus
• Strategic

Only Default / Athena and Vulcan are active gateways. Arjuna, Kubera, Prometheus, and Strategic remain stopped until needed.

Model Boundary

GLM / Z.ai is not live yet. Hermes supports provider id zai with aliases glm, z-ai, z.ai, and zhipu, but no ZAI_API_KEY / GLM_API_KEY is configured.

Direct DeepSeek is not live yet. Hermes supports provider id deepseek, but no DEEPSEEK_API_KEY is configured.

CI Runtime Fix

The CI cron wrappers must use /opt/hermes/.venv/bin/python when available. The Hermes container system Python does not have yaml, so using bare python3 can break CI even when the model provider works.

Verification

• Direct Gemini API smoke test reached the Gemini endpoint.
• Hermes gemini-2.5-flash one-shot returned successfully.
• Hermes gemini-2.5-pro one-shot returned successfully.
• Default route one-shot returned GEMINI_DEFAULT_OK.
• Vulcan profile one-shot returned VULCAN_GEMINI_OK.
• CI Gemini provider preflight exited 0.
• Chowmes provider watch exited 0 under Gemini routing.
• Chowmes health check passed and reported auxiliary model gemini-2.5-flash.

Backup

Live config backup was created at:

/opt/data/backups/model-routing-20260629004353