hy3-learning-agent

[zmuxuny]

Project

Learning Agent is a personal learning Agent Harness powered by the Hy3 API. It turns learning goals into reviewable plans, tracks real task evidence, maintains layered context and memory, and uses the same Agent Runtime for conversations, scheduled heartbeats, and email replies.

Complete source repository: https://github.com/zmuxuny/hy3-learning-agent

Final demo

Play the complete demo online (92.6 seconds · 1080p)

The single video contains both required end-to-end demos. Model decisions come from the TokenHub Hy3 API; tool calls, plan state, evidence checks, progress, and notifications are real application state. Editing only removes model and network wait time.

Hy3's role

Hy3 is called through Tencent TokenHub's OpenAI-compatible API. It decides when requirements are sufficient, delegates bounded planning research, chooses among 37 typed tools, observes tool results, teaches from the learner's current plan state, and decides whether proactive intervention is useful. The project does not train, fine-tune, or locally deploy Hy3.

End-to-end demos

1. Ambiguous goal → clarification card → planning sub-runs → verified learning resources → reviewable proposal → accepted plan and Session handoff.
2. Current-position teaching → real file/code evidence → bounded execution and grading → progress/XP/review update → proactive heartbeat decision.

Verification

• 26 backend tests passed
• Frontend production build passed
• Production dependency audit reported 0 vulnerabilities
• Real Hy3 profile_get → plan_list tool-calling smoke test passed
• SMTP and IMAP connection tests passed
• Responsive browser checks passed at 2560, 1440, 1280, 768, and 375 px widths

The submission file documents implemented capabilities and explicit safety limitations without presenting planned hardening as complete.