fix(ai-moderation): read delta.reasoning_content in stream aggregation — image vision never returned text

Root cause: 9router combo 'multimodal' routes to cloudflare-ai/@cf/google/
gemma-4-26b-a4b-it which streams ALL output in delta.reasoning_content
(content:"") and finishes with 'length' at max_tokens. llmClient only read
delta.content, so llmVision returned empty → every image moderation fell back
to text-only analysis ('Meskipun analisis gambar gagal' in every ai_analysis).

Fix: extractChunkText() prefers delta.content then falls back to
delta.reasoning_content (also handles message/text/response fields), with
unit tests for the exact 9router chunk shape. Verified live against a real
DB image: oc/mimo-v2.5-free (new first model in the multimodal combo) returns
a proper description in delta.content.

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

@@ -56,7 +56,7 @@ export async function withLlmConcurrency<T>(fn: () => Promise<T>): Promise<T> {
  */
 type LLMResponseChunk = {
   choices?: Array<{
-    delta?: { content?: string | null };
+    delta?: { content?: string | null; reasoning_content?: string | null };
     message?: { content?: string | null };
     finish_reason?: string | null;
     text?: string;
@@ -67,6 +67,29 @@ type LLMResponseChunk = {
   finish_reason?: string;
 };
 
+/**
+ * Extract the textual payload from a single streaming chunk. Prefers
+ * `delta.content`; falls back to `delta.reasoning_content` (DeepSeek-style /
+ * Cloudflare gemma stream ALL output there with content:"") so reasoning-only
+ * models still produce usable aggregated text. Exported for unit tests.
+ */
+export function extractChunkText(
+  chunk: LLMResponseChunk | null | undefined,
+): string {
+  if (!chunk) return "";
+  const choice = chunk.choices?.[0];
+  return (
+    choice?.delta?.content ||
+    choice?.delta?.reasoning_content ||
+    choice?.message?.content ||
+    choice?.text ||
+    chunk?.message?.content ||
+    chunk?.response ||
+    chunk?.content ||
+    ""
+  );
+}
+
 // ---------------------------------------------------------------------------
 // Lazy singleton — created on first use so that config is always resolved.
 // ---------------------------------------------------------------------------
@@ -167,15 +190,7 @@ export async function llmChat(
             let finishReason = "stop";
             for await (const chunk of response as unknown as AsyncIterable<LLMResponseChunk>) {
               const choice = chunk?.choices?.[0];
-              const textChunk =
+              content += extractChunkText(chunk);
-                choice?.delta?.content ||
-                choice?.message?.content ||
-                choice?.text ||
-                chunk?.message?.content ||
-                chunk?.response ||
-                chunk?.content ||
-                "";
-              content += textChunk;
               const fr = choice?.finish_reason || chunk?.finish_reason;
               if (fr) finishReason = fr;
             }

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

@@ -0,0 +1,70 @@
+// ═══════════════════════════════════════════════════════════════════════════
+// llmClient chunk extraction — reasoning_content fallback (pure, no network)
+// ═══════════════════════════════════════════════════════════════════════════
+// Regression: 9router "multimodal" combo routed to cloudflare gemma-4-26b
+// which streams ALL output in delta.reasoning_content with content:"" — the
+// old extractor returned empty text → llmVision reported "Vision API null
+// response" → every image moderation batch fell back to text-only analysis
+// (LLM kept writing "Meskipun analisis gambar gagal").
+import { describe, expect, it } from "vitest";
+import { extractChunkText } from "../src/modules/ai-moderation/llmClient.js";
+
+describe("extractChunkText — streaming chunk text extraction", () => {
+  it("reads delta.content (standard OpenAI streaming)", () => {
+    expect(
+      extractChunkText({
+        choices: [{ delta: { content: "halo" }, finish_reason: null }],
+      }),
+    ).toBe("halo");
+  });
+
+  it("falls back to delta.reasoning_content when content is empty — reasoning-only models (cloudflare gemma)", () => {
+    // Exact shape seen from 9router → cloudflare-ai/@cf/google/gemma-4-26b:
+    // {"choices":[{"delta":{"content":"","reasoning_content":"Task","role":"assistant"},"finish_reason":null,...}]}
+    expect(
+      extractChunkText({
+        choices: [
+          {
+            delta: { content: "", reasoning_content: "Task" },
+            finish_reason: null,
+          },
+        ],
+      }),
+    ).toBe("Task");
+  });
+
+  it("prefers content over reasoning when both present (deepseek-style final answer)", () => {
+    expect(
+      extractChunkText({
+        choices: [
+          {
+            delta: { content: "jawaban akhir", reasoning_content: "pikiran" },
+            finish_reason: null,
+          },
+        ],
+      }),
+    ).toBe("jawaban akhir");
+  });
+
+  it("handles Anthropic-style message.content", () => {
+    expect(extractChunkText({ message: { content: "via message" } })).toBe(
+      "via message",
+    );
+  });
+
+  it("handles top-level content / response fields (local LLM proxies)", () => {
+    expect(extractChunkText({ content: "top-level" })).toBe("top-level");
+    expect(extractChunkText({ response: "via response" })).toBe("via response");
+  });
+
+  it("returns empty string for null/undefined/empty chunks", () => {
+    expect(extractChunkText(null)).toBe("");
+    expect(extractChunkText(undefined)).toBe("");
+    expect(extractChunkText({})).toBe("");
+    expect(
+      extractChunkText({
+        choices: [{ delta: { content: "", reasoning_content: null } }],
+      }),
+    ).toBe("");
+  });
+});