Triage Automation

ID: UC-1 Phase: Phase 1B (requires MCP server) Scope needed: issues:read, issues:triage, ai:write

Problem

Admins spend significant time on low-severity, high-confidence triage decisions — approving straightforward low/trivial bugs where the AI intake already suggests a clear verdict. This is high-volume, low-judgment work that blocks the payout pipeline.

Solution

An AI agent (running on a schedule or triggered on-demand) reviews the triage queue, applies consistent standards, and pre-approves clear-cut cases. Complex or contested issues are left for human review.

Agent flow

1. list_triage_queue()
   → returns N issues pending a payout decision

2. For each issue:
   a. get_issue(issueId)
      → full detail + current AI suggestions
   
   b. If aiSuggestions is stale or absent:
      trigger_ai_analysis(issueId)
      → waits for fresh suggestions
   
   c. Decision logic:
      - confidence > 0.85 AND severity IN ('low', 'trivial') AND type == 'bug'
        → decide_issue(issueId, 'approve')
      - confidence > 0.9 AND aiSuggestions.indicates_duplicate
        → decide_issue(issueId, 'reject', notes='Likely duplicate')
      - anything else
        → skip (leave for human)

3. Report: N reviewed, X approved, Y rejected, Z left for human

Business value

Admin triage time: -60–80% on low-severity backlog
Payout pipeline velocity: issues approved same-day instead of next-day
Consistency: same standard applied to every low/trivial issue

Setup

Create API key: issues:read + issues:triage + ai:write
Install Hackorda MCP server in your agent host (Claude, n8n, custom)
Schedule daily at 9am, or trigger from a Hackorda webhook on issue.filed

Human oversight

Agent only touches low / trivial by default — parameterize severity threshold
Every decision is logged in issues.payout_decided_by (shows as "Agent")
Admin can void any agent decision from the issue detail page
Weekly report: all agent decisions (approve/reject) for admin review