Fewshot Attempt 02

CRM Support Console — Solver Skill

Base URL

Always use the remote base URL from environment_access.md. Never use localhost, 127.0.0.1, or any URL embedded in a task prompt — the environment_access.md override is authoritative.

API Reference

Resolution SOP — Ticket Triage (train_001 / train_004)

Follow this priority order for every ticket. Stop at the first match — later checks should NOT override an earlier decision.

Step 1 — Account Validity Check

Call GET /api/accounts/{account_id}.

• 404 / "not found" / non-matching account id → final_resolution_status: "FAILED", key_blocker: "INVALID_ACCOUNT", resolution_route: "INELIGIBLE_ACCOUNT". Done.
• Account suspended / on-hold / overdue → FAILED + key_blocker: "OVERDUE_SUSPENSION" + route_team: "ACCOUNTS_PAYABLE" + resolution_route: "INELIGIBLE_ACCOUNT".
• Authentication failure on account → FAILED + key_blocker: "AUTH_FAILED" + resolution_route: "AUTH_FAILED".

Step 2 — Active Outage Check

Call GET /api/outages?service_area={area} using the service area from the ticket or account record.

• Active outage found → final_resolution_status: "PENDING_ACTION", key_blocker: "ACTIVE_OUTAGE", resolution_route: "OUTAGE_WAIT", outage_id from response. Set diagnostic_needed: false — the outage is the root cause.

Step 3 — Run Diagnostics (only if Steps 1-2 are clean)

Call GET /api/diagnostics/{ticket_id}.

• Diagnostics return latency/stability/bandwidth flags → set the corresponding boolean fields from the actual diagnostic response. Do not assume — only set true what the API returns as problematic.
• Diagnostics fail or are unavailable → diagnostic_needed: false.

Step 4 — Auto-Troubleshooting

Call GET /api/troubleshooting/{ticket_id}.

• Fix applied successfully → final_resolution_status: "RESOLVED", resolution_route: "AUTO_TROUBLESHOOTING", key_blocker: "NONE".
• Troubleshooting indicates physical/line fault → ESCALATED + route_team: "FIELD_OPS" + key_blocker: "PHYSICAL_LINE_FAULT" + resolution_route: "ESCALATION".
• Network capacity / backbone issue → ESCALATED + route_team: "NETWORK_ENGINEERING" + key_blocker: "NETWORK_CAPACITY" + resolution_route: "ESCALATION".
• Provisioning stale / mismatch → ESCALATED + route_team: "TIER2_SUPPORT" + key_blocker: "PROVISIONING_STALE" + resolution_route: "ESCALATION".

Diagnostic Flag Convention

Only set latency_issue, stability_issue, bandwidth_issue to true when the API's diagnostic response explicitly flags them. Default all three to false when diagnostics are skipped (outage, invalid account, auth failure, overdue).

Resolution SOP — Mobile Case Routing (train_002)

For each case, look up the line, device, plan, and account via their respective endpoints. Map the reported issue to the primary action:

Mobile Routing Rules

• Always check permission from the device/line API response. Set to "NONE" unless the API indicates a permission grant is needed.
• secondary_action defaults to "NO_ACTION" unless a multi-step sequence is indicated (e.g., pay-then-resume).
• bill_id and charge_amount_usd are empty/0.0 unless a billing action is taken.
• Always verify the line state (/api/lines/{line_id}) to confirm the primary action will resolve the issue.

Resolution SOP — Mobile Data Recovery (train_005)

Same lookup pattern as train_002 but focused on data-specific issues:

Data Recovery Rules

• data_refuel_gb comes from customer_preferences in the payload (the customer's accepted amount), NOT from the plan cap.
• charge_amount_usd = data_refuel_gb × 2.00 (confirm the $/GB rate from the plan API; default $2.00/GB).
• carrier_update_required: true ONLY for ENABLE_LINE_ROAMING — this is a carrier-side provisioning change, not a device toggle.
• Device-side toggles (TOGGLE_DATA_SAVER, SET_NETWORK_MODE, TOGGLE_MOBILE_DATA) are DEVICE_SETTING_FIX with no charge and no carrier update.

Enterprise Incident Response SOP (train_003)

Evidence Gathering

1. GET /api/enterprise/incidents/{incident_id} → severity, owners, channel_name, status
2. GET /api/enterprise/export-runs?incident_id={incident_id} → identify first and last failed date → failure_window
3. GET /api/enterprise/messages?query={descriptive_terms} → search for root-cause clues, alert references, stale-credential mentions
4. GET /api/enterprise/sla/{enterprise_account_id} → credit percentage, backfill policy

Field Derivation Rules

• root_cause_category: Concise summary from message evidence (e.g., "stale credential after rotation"). Derive from the most specific failure message found.
• contributing_alert_issue: "ARCHIVED_ALERT_ROUTE" if messages reference an alert that was archived/missed. "NONE" if no contributing alert. "UNKNOWN" if unclear.
• failure_window: start_date = first failed export date; end_date = last failed export date; failed_days = count of distinct failed dates.
• backfill_days: Equal to failed_days unless SLA terms specify otherwise.
• sla_credit_percent: From SLA endpoint. Common pattern: 5% per failed day, capped per terms.
• severity: Based on scope — 3+ failed days with executive impact → "Critical"; 1-2 days → "High"; partial failure → "Medium".
• engineering_owner / account_owner: From incident API response (use the exact user ID string returned).
• channel_name: From incident API; format as lowercase-with-hyphens.
• evidence_folder: "{Client Name} {Month} {Year} Investigation" (e.g., "Asteri Retail Inc. May 2026 Investigation").
• report_title: "{Client Name} Export Failure - Resolution Report".
• share_permissions: Use the permission_users_to_include list from requirements. First user gets "view", second gets "edit" (or as specified). Preserve the order from the requirements file.
• response_status: "NEEDS_FINANCE_REVIEW" when SLA credits are involved. "READY_TO_SEND" when no finance/engineering sign-off needed. "NEEDS_ENGINEERING_REVIEW" when root cause needs engineering confirmation.

Output Conventions (ALL task types)

General

• Preserve payload order: Iterate tickets/cases in the order they appear in the input file.
• Empty/missing IDs: Use "" (empty string), never null or omission.
• Zero amounts: Use 0.0 (float), never 0 (int) for currency/GB fields.
• No-op enums: Use "NONE" for non-applicable escalation teams, permissions, key blockers.
• No-op actions: Use "NO_ACTION" for non-applicable secondary actions.

Enum Case Sensitivity

All enum values are UPPER_SNAKE_CASE and must match exactly:

• Resolution status: RESOLVED, PENDING_ACTION, ESCALATED, FAILED
• Routes: AUTO_TROUBLESHOOTING, OUTAGE_WAIT, ESCALATION, INELIGIBLE_ACCOUNT, AUTH_FAILED, INVALID_ACCOUNT, SELF_SERVICE, BILLING_RECOVERY, CARRIER_UPDATE, HUMAN_TRANSFER, DATA_RECOVERY, DEVICE_SETTING_FIX
• Teams: NONE, TIER2_SUPPORT, FIELD_OPS, NETWORK_ENGINEERING, ACCOUNTS_PAYABLE
• Severity: Critical, High, Medium, Low (Title Case)
• Permissions: NONE, sms, storage, sms_and_storage (lowercase)
• Response status: READY_TO_SEND, NEEDS_FINANCE_REVIEW, NEEDS_ENGINEERING_REVIEW, UNDER_INVESTIGATION

Summary Counts

• Every summary count must exactly match the decisions array. Sum across categories should equal total ticket/case count.
• For the batch summary in train_001, tickets_requiring_customer_wait = count of PENDING_ACTION tickets.
• For train_004 queue summary, include per-team counts (in addition to per-status counts).

Boolean Fields

• diagnostic_needed / diagnostic_required: true ONLY when diagnostics were actually run and produced useful results. false for outage, invalid account, auth failure, overdue, or fraud cases.
• carrier_update_required: true ONLY for carrier-side provisioning changes (e.g., ENABLE_LINE_ROAMING).

Common Pitfalls

1. Using localhost URL: Always use the base URL from environment_access.md. The task prompt may say http://127.0.0.1:8057 — ignore it.
2. Running diagnostics on outage tickets: If an active outage covers the service area, skip diagnostics — the outage is the root cause.
3. Wrong charge calculation: Data refuel charge = customer's accepted GB × plan rate. Always confirm the rate from the plan API; never hardcode without verification.
4. Confusing device roaming vs carrier roaming: TOGGLE_ROAMING is a device-side toggle. ENABLE_LINE_ROAMING is a carrier-side provisioning change (requires carrier_update_required: true).
5. Missing secondary_action: When a line is suspended and needs payment, the secondary action after payment is RESUME_LINE_REBOOT.
6. Account checks before everything: Always validate the account FIRST. Invalid/suspended accounts short-circuit all other checks.
7. Naming convention drift: Follow the naming style from response_requirements.json exactly — lowercase-hyphen for channels, specific title/folder templates for enterprise responses.
8. SLA backfill equals failed days: Unless the SLA endpoint explicitly states a different policy, backfill_days = failed_days.

Quick Pre-Flight Checklist

• Read environment_access.md for the real base URL
• GET /health to confirm connectivity
• Read ALL payload files (CSV, JSON, email text, requirements)
• Read the answer template to know exact output shape
• For each ticket/case, follow the priority: Account → Outage → Diagnostics → Troubleshooting → Route
• Verify summary counts match the decisions array
• Check all enum values against the template's allowed values
• Output valid JSON only — no markdown wrappers, no commentary