Skip to content

Ticket triage and 15-exchange AI context

Open Tiket bantuan. Switch between Tabel and Papan Kanban. Four lanes match existing statuses: open, in progress, resolved and closed. Drag a card to another lane or use its keyboard-accessible status selector. Clicking the title opens the ticket; case links still open the associated received conversation.

Each ticket has a priority (low, normal, high, urgent), category (general, account, app, agenda, assignment, membership, other), optional due date/time in WIB and assigned staff member. Existing tickets start with normal priority, general category and no due date. A due date is a staff target, not an automatic contractual SLA. Only open/in-progress tickets past their due date are overdue.

Search matches ticket ID, subject, description, requester name, email or phone. Filter by status, priority, category, assignee, requester and overdue status; sort by newest, oldest, highest priority or earliest due date. Requester dropdowns follow the existing 200-member result limit; search by name/email/phone for members outside that list. Both table and board show up to 200 filtered tickets, not the entire global backlog. Lane counts describe the displayed results.

Select up to 50 displayed tickets for a bulk status, priority or assignment update. Nothing is sent to WhatsApp. Bulk changes are one database transaction: an invalid/stale ticket or assignee causes the whole batch to roll back. The selection is cleared when data refreshes so an old selection cannot silently act on different results.

Ticket forms, lane moves and bulk actions send the version shown to the admin. If another person has edited the ticket meanwhile, the server returns a conflict instead of overwriting that edit. Refresh the list, or close/reopen the ticket dialog, review the newer version and try again. Legacy integration clients may omit the optional expected_version for compatibility; new clients should always send it. Database row locks serialize the actual edits.

The activity timeline shows actor, WIB timestamp, before/after field changes and internal notes. Clearing assignment or a due date is recorded. Historical events remain compatible. Staff notes remain admin/staff-only, including when a member reads their own ticket through the integration API. WhatsApp natural-language status replies remain owner-filtered and never include internal notes. Bot actions cannot change priority, assignment, targets or staff-only ticket data.

The worker retains the latest 15 completed exchanges for each member, each expiring after 24 hours. The history character budget is 75,000 so all 15 maximum stored text pairs fit; it is not the model’s total advertised token window. The current prompt/FAQ and output budget are separate. Common credential patterns are sanitized; screenshot bytes and member contact profiles are excluded. Google quotas and shared provider backoff remain in effect. Longer prompts can consume more free-tier input quota and take longer to process.

The expanded memory accumulates as new messages are processed; exchanges previously pruned from short-term memory are not reconstructed from the case archive. Thirty-day case records stay separate and are not loaded wholesale into model prompts.

Tests include a full 15-pair provider payload (31 user/model messages including the current question), field/filter behavior, atomic bulk updates, stale-version conflicts, invalid assignees and timezone validation. Live generic Gemma checking identified the application named in the oldest of 15 exchanges. Browser checks use only a temporary database: edit triage metadata, filter overdue work, move cards by selector/drag, apply bulk updates and inspect the timeline on desktop/mobile. Production browser checks perform no business-record updates or outbound messages.

Schema migration is additive. Take a database backup and source copy before rollout; recreate only API, worker and dashboard. WAHA, PostgreSQL and the shared proxy remain in place. Rollback source is server code-backups/before-ticket-tools/app; preserve new columns and data. Existing record retention and credential management stay as documented.