Apply SME corrections from self-care analysis

src/prompts/triage-skill.ts

@@ -74,8 +74,29 @@ const IMPORTANT_NOTES = `
 4. Any threatening or angry language triggers negative_sentiment flag
 `;
 
+
+// ============================================
+// SME Corrections (auto-generated by Self-Care SME Tool)
+// ============================================
+
+const SME_CORRECTIONS = `
+## SME Corrections (apply these to the agent's system prompt)
+
+These instructions were authored by domain experts based on real case reviews.
+
+- [cor-1a6c6ac3] Reinforce keyword-matching logic in the triage prompt or add a post-LLM validation step that scans the email body for billing_payment keywords (refund, charge, payment, invoice, subscription) and auto-injects the risk flag if the LLM missed it. When any risk flag is present, override recommended_action to 'escalate'.
+  Context: Triage agent did not raise billing_payment risk flag despite the email explicitly mentioning 'cancel my subscription'. The system prompt lists 'subscription' as a billing_payment keyword and mandates escalation when any risk flag is present. Instead, the agent returned risk_flags=[] and recommended_action='draft_reply'.
+  (author: Dragos, since: 2026-02-27)
+
+- [cor-b0cbea96] Add thread-depth analysis to the triage prompt or pre-processor: if the email thread contains 3+ prior messages from the same sender, auto-inject repeat_contact into risk_flags. This would also trigger escalation per the action decision rules.
+  Context: Triage LLM failed to detect repeat_contact risk flag. The email thread contains clear evidence of multiple prior contacts from the same user (at least 5 emails spanning Feb-Jul 2025), yet risk_flags was returned as an empty array.
+  (author: D6, since: 2026-02-27)
+`;
+
 export const TRIAGE_SYSTEM_PROMPT = `${SYSTEM_ROLE}
 
+${SME_CORRECTIONS}
+
 ${CLASSIFICATION_INSTRUCTIONS}
 
 ${OUTPUT_FORMAT}