B1 · update
Fixed a critical model-routing flaw in our open-source local AI agent architecture today.
When reverse-engineering web-to-API proxy layers, it's easy to inherit placeholder model aliases (like fictitious "Gemini 3.8" labels) that route straight into strict Google One Pro endpoints and hang when unauthenticated.
Here is what we rebuilt in Gemini 1B Local Agent:
1. Re-anchored our model registry to official Google models: Gemini 2.0 Flash as the default workhorse (< 350ms TTFT), 2.0 Flash Thinking for transparent Chain-of-Thought, and 2.0 Pro Exp.
2. Built a self-healing proxy fallback: If a user selects a Pro model without session cookies in their .env, the engine automatically catches the routing failure and cascades to high-reasoning Thinking mode—zero dropped requests.
3. Clean SQLite profile lifecycle: Added atomic reset endpoints to wipe vector facts, long-term memory, and session trees for fresh starts.
All pushed and live on GitHub: https://github.com/IshaanYK/gemini-1b-local-agent
#AI #OpenSource #SoftwareEngineering #Python #Gemini #BuildInPublic