aozyildirim/agena/stale-ticket-triage

Stale Ticket Triage

Backlogs accumulate dead rows. Most teams have a recurring Friday meeting where someone walks every ticket older than X days and decides whether to close, snooze, or keep it. That hour is the perfect AI workload — the decision is shallow but the volume is high.

How to apply this pattern

1. Define stale: a ticket is stale when its updated_at is older than

the org's threshold (default 30 days) AND its status is still in the active set (not Closed / Done / Cancelled).

1. Schedule the scan (chip-pick: every 6h / 12h / daily 9am / weekly

Sundays / monthly). Each run hits the source-platform API for every project the org cares about, lists stale issues, and sends each through a short LLM call.

1. System prompt the LLM with three explicit verdicts —

close / snooze / keep — plus a one-sentence reason. Be conservative on close: only pick it when the ticket itself signals resolution (links a merged PR, mentions a follow-up). Default to snooze when unsure; never close from silence alone.

1. Persist each verdict as a triage decision row keyed by

(org, source, external_id) so re-runs are idempotent. Status transitions: pending → applied / skipped / overridden.

1. One-click bulk approve in the UI: the human reviews the AI verdicts

in a list, hits "Apply all AI suggestions", and the system writes back to Jira / Azure DevOps in a single batch.

Example LLM reply format

VERDICT: close
REASON: Closed by PR #4221 (merged 38 days ago); customer hasn't
        responded since the fix shipped.

VERDICT: snooze
REASON: Still relevant but no recent customer activity; revisit when
        the related epic resumes.

Notes

• Audit trail is non-negotiable. Every AI verdict and every human

override needs a row with timestamp + user id. Bulk-approve must not destroy history; it appends.

• Source-side scan (read tickets directly from Jira / Azure DevOps)

scales better than relying on imported task records, because most stale tickets were never imported into the AI tool.

• Threshold is per-workspace. Solo product teams might want 14 days;

enterprise backlogs need 60-90.

• The verdict prompt is short on purpose — long prompts encourage the

model to add commentary and skip the structured reply.

• Pair with reporter-routing rules so security tickets bypass the

triage flow entirely — those should never be auto-closed.