Claude Support Agent — Data skill for Claude Code

claude-support-agent

[](https://github.com/rahulkumar96t/claude-support-agent/actions/workflows/ci.yml)

A customer-support agent for a fictional outdoor-gear shop, built directly on the Claude Messages API tool-use loop. The model talks to the customer and decides which tool to call; the code decides what is allowed. Refund eligibility and amounts come from a written policy implemented in Python, refunds above a threshold pause the conversation until a human approves them, every tool only sees the signed-in customer's data, tool output reaches the model as untrusted JSON, and every conversation leaves a PII-redacted JSONL trace. It runs as a terminal chat or a FastAPI service, and the whole test suite runs offline against a scripted model.

Features

• Bounded agent loop on the Messages API: max model calls per request, a per-call timeout and a wall-clock deadline, with explicit stop reasons (completed, awaiting_approval, max_turns, timeout, refusal, out_of_scope).
• Six tools defined from Pydantic models: lookup_order, get_shipment_status, check_refund_eligibility, issue_refund, create_ticket, search_knowledge_base. The JSON schemas are sent with strict: true; the same models validate the arguments before anything runs.
• Written refund policy, enforced in code: `refund-policy.md` (30/60-day windows, excluded categories, one refund per order, $500 self-service cap, review above $100) is implemented in policy.py, and a test checks that the numbers in the document match the code. issue_refund has no amount parameter.
• Human approval gate: a refund above the threshold is held; the conversation pauses with an approval id, a reviewer approves or denies (terminal prompt or POST /approvals/{id}), the policy is re-checked, and the loop resumes with the result.
• **