Helpdesk

Helpdesk

What this skill does

Classify one bounded support request, choose the safest next path, and draft a customer reply only when the supplied context supports one. The runner emits a support triage packet with classification, severity, confidence, evidence, draft email, and a send gate.

This skill never sends email, posts to Slack, opens issues, mutates accounts, or touches billing. It prepares the decision packet that a separate governed send skill can review, approve, and deliver with its own authority grant and receipt.

When to use this skill

Use this skill when an agent has a single support request and needs a safe first decision: answer, ask for more information, route to engineering, route to billing, route to account review, or route to abuse review.

It is useful in business-ops graphs where support triage fans out from inbox intake, then hands off only safe drafts to send-as or another provider-backed send lane. It also works for dry-run review of support queues because it has no external side effects.

When not to use this skill

Do not use this skill as a helpdesk transport, account access workflow, billing authority, customer identity verifier, or automatic sender. Do not use it to answer from private account state unless that state has already been summarized into an approved support_request packet.

If the request asks for account recovery, billing changes, abuse handling, or anything that needs private records, the skill must not draft a definitive answer. It should return a review route and let a stronger authority gate handle the consequence.

Procedure

1. Require support_request to contain at least subject or body.
2. Normalize the request text and classify it as how_to, billing, account_access, bug, abuse, or unknown.
3. Estimate severity and confidence from matched signals.
4. Choose the recommended path.
5. Draft a customer reply only for safe how_to requests.
6. Record missing context and matched signals.
7. Emit send_gate.status = requires_human_approval for every result.
8. Let downstream send or operator lanes decide whether a draft may be sent.

Edge cases and stop conditions

Return needs_input when support_request is missing or lacks both subject and body. Return needs_more_evidence when the request is too vague to route. Mark private-state paths as gated review, not as ready-to-send drafts. If a caller asks the skill to send, mutate an account, bypass an approval gate, or provide credential recovery instructions, return refused.

The authority scope is classification and draft preparation only. The proof surface is the sealed receipt containing the request summary, matched signals, recommended path, draft proposal if any, and send gate. Any live customer send requires a separate send-as receipt.

Output schema

The runner emits runx.support.triage_reply.v1:

{
  "classification": "how_to | billing | account_access | bug | abuse | unknown",
  "severity": "low | medium | high | critical",
  "confidence": 0.78,
  "recommended_path": "reply_draft | request_info | engineering_intake | billing_review | account_review | abuse_review | manual_review",
  "evidence": {
    "source": "fixture:safe-how-to",
    "source_summary": "How do I verify my sending domain?",
    "matched_signals": ["verify", "dns", "domain"],
    "missing_context": [],
    "taxonomy_coverage": ["how_to", "billing", "account_access", "bug", "abuse", "unknown"],
    "private_data_required": false,
    "send_side_effects": "none"
  },
  "draft_email": {
    "proposed": true,
    "subject": "Re: How do I verify my sending domain?",
    "body": "..."
  },
  "send_gate": {
    "status": "requires_human_approval",
    "action": "send_customer_email",
    "rationale": "..."
  }
}

Worked example

runx skill "$PWD" \
  --runner triage \
  --input-json support_request='{
    "customer_name": "Mira",
    "customer_email": "[email protected]",
    "subject": "How do I verify my sending domain?",
    "body": "I added the DNS records. What should I check next?",
    "source": "fixture:safe-how-to"
  }' \
  --input-json policy='{
    "product_name": "ExampleDesk",
    "support_signature": "ExampleDesk Support"
  }' \
  --json

Expected result: classification = how_to, recommended_path = reply_draft, draft_email.proposed = true, and send_gate.status = requires_human_approval. The run does not send the email.

Inputs

• support_request: object with subject, body, optional customer_name, optional customer_email, optional source, and optional refs.
• policy: optional object with product_name, support_signature, safe_reply_topics, and escalation_contacts.

Outputs

• classification: request class.
• severity: operational severity.
• confidence: number from 0 to 1.
• recommended_path: next safe handling path.
• evidence: matched signals, missing context, source summary, and taxonomy coverage.
• draft_email: proposed customer reply or blocker reason.
• send_gate: always requires human approval before delivery.