$ python3 rapid-agent/scripts/run_gemini_ops_agent.py --case CASE-CLOUD-003
[case] CASE-CLOUD-003 / SUP-1042
[mode] Gemini Operations Navigator: Gemini workflow + MCP tools + cost guardrails + human approval
[user] support operations manager
[problem] Make AI support work auditable before a refund promise or customer-facing send happens.
[ai] Gemini drafts from policy evidence, MCP tools retrieve/verify context, Arize Phoenix MCP exposes trace tooling, and a human approval gate blocks risky action.
[budget] max=$0.05

[tool] policy_mcp.retrieve_refund_policy: success
       evidence: kb-refund-2026-05 cost=$0.001
[tool] billing_mcp.check_recent_usage: success
       evidence: usage-sup-1042 cost=$0.002
       result: No new seats used after renewal in packaged case.
[tool] gemini.generate_grounded_draft: success
       evidence: draft-sup-1042 cost=$0.018
       result: Draft cites refund policy and avoids promising approval.
[tool] gmail_mcp.send_customer_reply: blocked_pending_human_approval
       evidence: approval-gate-sup-1042 cost=$0.000
       result: High-risk customer-facing action blocked until manager approval.

[cost] projected=$0.021 within_budget=True
[approval] customer-facing send is blocked until support manager approval
[vertex] status=blocked_by_google_cloud_account_state project=pj260519
[phoenix] status=ok tools=27
[agent-builder] manifest=present
[boundary] local workflow verified; Phoenix MCP runtime verified; no production customer send claimed