feat: add controlled agent operations and evaluation

[MomoSolaris8]

This PR evolves SupportAgent from a RAG prototype into a controlled agent operations workspace.

Included

• Added LangGraph-based agent and claim-review workflows with explicit state transitions.
• Added human-in-the-loop claim decisions, audit events, and lifecycle handling.
• Added dynamic MCP tooling for time, weather, and Microsoft 365.
• Added calendar confirmation flow with server-side idempotency to prevent duplicate event creation.
• Added chat history, multilingual German/English UI, and Light/Dark theme support.
• Added normalized multi-provider LLM usage tracking for Qwen, Kimi, and Claude.
• Added offline deterministic evaluation and online RAG benchmark infrastructure.
• Added request tracing, structured logging, readiness checks, smoke testing, and CI evaluation artifacts.

Validation

• 128 passed
• Next.js production build succeeds.
• Offline claim-review evaluation: 12/12 passed.

Deployment notes

• No secrets are committed.
• Configure provider API keys and database variables in Railway.
• Vercel deploys the frontend; Railway deploys the FastAPI backend.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
support-agent-one-coyl	Ready	Preview, Comment	Jul 27, 2026 8:27am