fix: handle message_id with extra wrapping quotes and request strict JSON output without reasoning

src/moderation/llmModerationClient.ts

@@ -132,6 +132,13 @@ export function parseModerationResponse(
       }
 
       let finalId = String(message_id).trim();
+      // Remove wrapping double quotes if any (common in some LLM outputs)
+      if (finalId.startsWith('"') && finalId.endsWith('"')) {
+        finalId = finalId.slice(1, -1).trim();
+      }
+      if (finalId.startsWith("'") && finalId.endsWith("'")) {
+        finalId = finalId.slice(1, -1).trim();
+      }
       if (finalId.startsWith("[") && finalId.endsWith("]")) {
         finalId = finalId.slice(1, -1).trim();
       }
@@ -335,6 +342,7 @@ Each result must have:
 - score: confidence score from 0 to 1
 - analysis: brief explanation
 
+Do not include reasoning, analysis steps, markdown, prose, XML tags, or comments.
 Return ONLY valid JSON, no other text.`;
 
   // Check for image attachments to support multimodal analysis
@@ -456,11 +464,11 @@ Return ONLY valid JSON, no other text.`;
                   content: messageContent,
                 },
               ],
-              temperature: 0.6,
-              top_p: 0.95,
-              max_tokens: 65536,
-              reasoning_budget: 16384,
-              chat_template_kwargs: { enable_thinking: true },
+              temperature: 0,
+              top_p: 1,
+              max_tokens: 8192,
+              response_format: { type: "json_object" },
+              chat_template_kwargs: { enable_thinking: false },
             }),
           },
         );

tests/moderation/llmModerationClient.test.ts

@@ -320,6 +320,26 @@ describe("parseModerationResponse", () => {
     expect(result).toHaveLength(1);
     expect(result[0].messageId).toBe("m1");
   });
+
+  it("handles message_id returned with extra wrapping quotes", () => {
+    const result = parseModerationResponse(
+      JSON.stringify({
+        results: [
+          {
+            message_id: '"test-msg-1"',
+            status: "clean",
+            flags: [],
+            score: 0.1,
+            analysis: "OK",
+          },
+        ],
+      }),
+      ["test-msg-1"],
+    );
+
+    expect(result).toHaveLength(1);
+    expect(result[0].messageId).toBe("test-msg-1");
+  });
 });
 
 describe("runModerationAnalysis", () => {
@@ -359,6 +379,45 @@ describe("runModerationAnalysis", () => {
     expect(result.raw).toEqual(mockResponse);
   });
 
+  it("requests strict JSON output without thinking", async () => {
+    const mockResponse = {
+      choices: [
+        {
+          message: {
+            content: JSON.stringify({
+              results: [
+                {
+                  message_id: "m1",
+                  status: "clean",
+                  flags: [],
+                  score: 0.1,
+                  analysis: "OK",
+                },
+              ],
+            }),
+          },
+        },
+      ],
+    };
+
+    global.fetch = vi.fn().mockResolvedValue({
+      ok: true,
+      text: async () => JSON.stringify(mockResponse),
+      json: async () => mockResponse,
+    });
+
+    await runModerationAnalysis({
+      targets: [createMessageRecord()],
+      contextText: "test context",
+    });
+
+    const requestBody = JSON.parse((global.fetch as any).mock.calls[0][1].body);
+    expect(requestBody.temperature).toBe(0);
+    expect(requestBody.response_format).toEqual({ type: "json_object" });
+    expect(requestBody.reasoning_budget).toBeUndefined();
+    expect(requestBody.chat_template_kwargs).toEqual({ enable_thinking: false });
+  });
+
   it("throws on non-ok HTTP response", async () => {
     global.fetch = vi.fn().mockResolvedValue({
       ok: false,