Athena soul refactor

Date: 2026-06-29 Status: Implemented on Chowmes VPS

Trigger

Arijit reported that Athena's Telegram responses sounded robotic and repetitive despite a defined personality.

Finding

The live SOUL.md over-indexed on ceremony, gravitas, and role language. A smoke test for a casual greeting produced:

Arijit. I am here.

The board is clear, and I am ready to engage with whatever strategic imperative you bring.

That confirmed the prompt itself was encouraging theatrical, stiff language.

The live soul file also still described the old OpenRouter/DeepSeek/Claude routing policy even though Chowmes had moved to Gemini plus Algolia inference.

Change

SOUL.md was refactored to:

• prefer alive, plain, warm, specific speech over ceremonial authority.
• ban stock roleplay phrases such as "The board is clear", "strategic imperative", "I stand ready", and "How may I assist".
• add anti-robot rules for casual, frustrated, and operational messages.
• update model posture to current routing:
• default Gemini Flash.
• Gemini Pro for authority/deep work.
• Algolia inference only for low-end fast/casual work and cheap housekeeping.

Verification

After deploying the new SOUL.md, clearing the stale Telegram session, and running health checks:

• Telegram send-test succeeded.
• Casual Athena smoke returned: Here. I'm ready. What's on your mind?
• Frustration-handling smoke acknowledged the frustration before asking for the next symptom.

Caveat

This is a first-pass personality repair, not a finished voice system. Athena should be judged from real Telegram exchanges over the next few days and tuned against actual examples of robotic responses.