fix(gateway): flagged high/critical severity always eligible for auto-delete

The LLM's recommended_action is conservative — for a flagged message at
high/critical severity it frequently emits 'review' (screenshot/context
ambiguity) even when the violation is severe (harassment, SARA,
threats). autoDeleteEligibility trusted that value, so serious violations
slipped through undeleted (e.g. harassment flagged high but
review → not eligible).

Fix: flagged + high/critical severity bypasses the recommended-action
check entirely (always eligible); the action check now only gates
warn/flagged-medium. deriveRecommendedAction also returns delete for
flagged high/critical BEFORE consulting the stored LLM action. +5
regression tests.

services/discord-gateway/src/modules/ai-moderation/autoDeleteEligibility.ts

@@ -35,8 +35,17 @@ export function deriveSeverity(msg: MessageRecord): string {
 
 /** Derive recommended action from legacy messages that lack structured AI fields. */
 export function deriveRecommendedAction(msg: MessageRecord): string {
-  if (msg.ai_recommended_action) return msg.ai_recommended_action;
   const severity = deriveSeverity(msg);
+  // Flagged at high/critical severity is ALWAYS delete — the stored
+  // recommended_action from the LLM is conservative (often "review") and
+  // must not override severity for severe violations.
+  if (
+    msg.ai_status === "flagged" &&
+    (severity === "critical" || severity === "high")
+  ) {
+    return "delete";
+  }
+  if (msg.ai_recommended_action) return msg.ai_recommended_action;
   if (
     msg.ai_status === "flagged" &&
     (severity === "critical" || severity === "high" || severity === "medium")
@@ -177,19 +186,36 @@ export function isEligibleForAutoDelete(
     return false;
   }
 
-  // Recommended action check
+  // Recommended action check.
-  const recommendedAction =
+  // CRITICAL: the LLM's `recommended_action` is CONSERVATIVE — for a flagged
-    analysisResult?.recommendedAction ?? deriveRecommendedAction(message);
+  // message at high/critical severity it frequently emits "review" (it sees a
+  // screenshot/context ambiguity and hedges) even when the violation itself is
+  // severe. Trusting that value lets serious violations (harassment, SARA,
+  // threats) slip through undeleted. So: flagged + high/critical severity is
+  // ALWAYS eligible regardless of the LLM's recommended action. The action
+  // check only gates warn/flagged-medium (where a review is legitimate).
   if (
-    recommendedAction !== "delete" &&
+    status === "flagged" &&
-    recommendedAction !== "escalate" &&
+    (severity === "high" || severity === "critical")
-    recommendedAction !== "warn"
   ) {
     logger.debug(
-      { messageId: message.id, recommendedAction },
+      { messageId: message.id, status, severity },
-      "Message not eligible for auto-delete: recommended action is not delete/escalate/warn",
+      "Message eligible for auto-delete: flagged with high/critical severity",
     );
-    return false;
+  } else {
+    const recommendedAction =
+      analysisResult?.recommendedAction ?? deriveRecommendedAction(message);
+    if (
+      recommendedAction !== "delete" &&
+      recommendedAction !== "escalate" &&
+      recommendedAction !== "warn"
+    ) {
+      logger.debug(
+        { messageId: message.id, recommendedAction },
+        "Message not eligible for auto-delete: recommended action is not delete/escalate/warn",
+      );
+      return false;
+    }
   }
 
   // Categories check

services/discord-gateway/tests/autoDeleteEligibility.test.ts

@@ -0,0 +1,68 @@
+import { describe, expect, it } from "vitest";
+import {
+  deriveRecommendedAction,
+  isEligibleForAutoDelete,
+} from "../src/modules/ai-moderation/autoDeleteEligibility.js";
+import type { MessageRecord } from "../src/shared/moderation-types.js";
+
+const baseMessage: MessageRecord = {
+  id: "test-msg-1",
+  channel_id: "chan-1",
+  thread_id: null,
+  guild_id: "guild-1",
+  user_id: "user-1",
+  username: "tester",
+  content: "kau ngehina aku hitam kah ?",
+  created_at: Date.now(),
+  ai_status: "flagged",
+  ai_severity: "high",
+  ai_recommended_action: "review",
+  ai_moderation_flags: '["harassment"]',
+  ai_confidence: 0.95,
+  ai_analysis: "konfrontatif",
+  ai_categories: null,
+  ai_moderation_score: null,
+  ai_analyzed_at: Date.now(),
+  deleted_at: null,
+} as unknown as MessageRecord;
+
+describe("autoDeleteEligibility — flagged high severity with conservative LLM action", () => {
+  it("flagged + high severity is eligible even when LLM said review", () => {
+    const eligible = isEligibleForAutoDelete(baseMessage);
+    expect(eligible).toBe(true);
+  });
+
+  it("flagged + high severity derives delete regardless of stored review action", () => {
+    expect(deriveRecommendedAction(baseMessage)).toBe("delete");
+  });
+
+  it("flagged + medium severity with review action is NOT eligible", () => {
+    const medium = {
+      ...baseMessage,
+      ai_severity: "medium",
+    } as unknown as MessageRecord;
+    const eligible = isEligibleForAutoDelete(medium);
+    expect(eligible).toBe(false);
+  });
+
+  it("warn status with review action is NOT eligible", () => {
+    const warn = {
+      ...baseMessage,
+      ai_status: "warn",
+      ai_severity: "low",
+      ai_recommended_action: "warn",
+    } as unknown as MessageRecord;
+    const eligible = isEligibleForAutoDelete(warn);
+    expect(eligible).toBe(true); // warn action is allowed
+  });
+
+  it("clean status is never eligible", () => {
+    const clean = {
+      ...baseMessage,
+      ai_status: "clean",
+      ai_severity: "none",
+      ai_recommended_action: "none",
+    } as unknown as MessageRecord;
+    expect(isEligibleForAutoDelete(clean)).toBe(false);
+  });
+});