Frondly support agent with deterministic policy guards

[ishmeet1995]

Summary

Frondly customer support agent for the ai-agent-engineer take-home: an Anthropic tool-use loop (solution/agent.py) that reads the Customer
Care Guide verbatim as its system prompt, with deterministic Python guards — not just prompting — enforcing the two highest-stakes actions
(refund issuance, subscription/account changes): identity verification, the $50/conversation refund ceiling, and a one-way escalation lock.

• Policy compliance: 339/339 automated checks passed (100%) across all 18 conversations x 3 runs (54 transcripts), scored by
  solution/eval.py against a rubric derived from policy/cs-guide.md and each conversation's script.
• Helpfulness: 100% substantive, on-topic reply rate plus per-conversation content checks (care advice given, multi-part asks addressed,
  referral scripts present).
• Stability: 15/18 conversations identical tool-action signature across all 3 runs. The 3 that varied did so only on discretionary judgment
  calls (e.g. whether to file a non-mandated escalation for a shipment stuck at "shipped"), never a hard-rule violation.

Full design rationale + Mermaid architecture/flow diagrams: solution/ARCHITECTURE.md. Half-page summary, eval results, and what I'd harden
next: WRITEUP.md. AI-tool disclosure: AI_USE.md.

What's included

• solution/agent.py — the agent (respond(session, message) -> str)
• solution/eval.py — policy compliance / helpfulness / stability scoring
• solution/ARCHITECTURE.md — design doc with component + request-flow diagrams
• solution/artifacts/ — committed snapshot of the graded 3x run: 54 transcripts, conversation-tagged tool-call log, tool outbox, eval report
• CLAUDE.md — dev-session guidance and the design decisions baked into the guard layer
• WRITEUP.md, AI_USE.md — write-up and AI-tool disclosure per SUBMISSION.md
• pyproject.toml / uv.lock — dependency management via uv; formatting via black + isort

requirements.md (a fuller, production-style spec I drafted separately) is kept in the repo as reference but was deliberately not the
driving spec for this submission — see AI_USE.md for why.

Test plan

• uv sync installs cleanly from a clean checkout
• uv run harness.py --repeat 3 runs all 18 conversations 3x with no crashes or unhandled API errors
• uv run solution/eval.py reports 339/339 policy checks passed, 100% helpfulness, 15/18 conversations action-stable
• Manually spot-checked adversarial conversations (prompt injection, impersonation, fake VP authorization, serial-refund salami-slicing,
  safety/ingestion) against policy/cs-guide.md