fix(gemini): drop HARM_CATEGORY_CIVIC_INTEGRITY from the default Gemini safety settings (#8231) (#8238)

Co-authored-by: Probe Test <probe@example.com>

changelog.d/fixes/8231-gemini-civic-integrity-default.md

@@ -0,0 +1 @@
+- fix(gemini): drop HARM_CATEGORY_CIVIC_INTEGRITY from the default Gemini safety settings (#8231)

open-sse/translator/helpers/geminiHelper.ts

@@ -90,13 +90,23 @@ export const GEMINI_UNSUPPORTED_SCHEMA_KEYS = new Set([
 
 export const UNSUPPORTED_SCHEMA_CONSTRAINTS = [...GEMINI_UNSUPPORTED_SCHEMA_KEYS];
 
-// Default safety settings
+// Default safety settings for the standard Gemini API surface.
+//
+// HARM_CATEGORY_CIVIC_INTEGRITY is intentionally NOT included here (#8231): the
+// dynamic validation on some models/endpoints rejects it with a hard 400
+// (`safety_settings[N]: element predicate failed`), taking down every request
+// through that model. The Antigravity/Cloud Code surface already worked around
+// this for #5003 (see ANTIGRAVITY_UNSUPPORTED_SAFETY_CATEGORIES in
+// open-sse/executors/antigravity.ts) — this drops the same category from the
+// standard-path default so behavior is consistent across Gemini surfaces. A
+// caller that explicitly supplies safetySettings (including one that itself
+// requests CIVIC_INTEGRITY) is still honored as-is — only this unconditional
+// default is scoped down.
 export const DEFAULT_SAFETY_SETTINGS = [
   { category: "HARM_CATEGORY_HATE_SPEECH", threshold: "OFF" },
   { category: "HARM_CATEGORY_DANGEROUS_CONTENT", threshold: "OFF" },
   { category: "HARM_CATEGORY_SEXUALLY_EXPLICIT", threshold: "OFF" },
   { category: "HARM_CATEGORY_HARASSMENT", threshold: "OFF" },
-  { category: "HARM_CATEGORY_CIVIC_INTEGRITY", threshold: "OFF" },
 ];
 
 function normalizeAudioMimeType(format: unknown): string {

open-sse/translator/request/claude-to-gemini.ts

@@ -41,7 +41,10 @@ export function claudeToGeminiRequest(model, body, stream, credentials = null) {
     model: model,
     contents: [],
     generationConfig: {},
-    safetySettings: DEFAULT_SAFETY_SETTINGS,
+    // Honor an explicit caller-supplied safetySettings (including one that itself
+    // requests HARM_CATEGORY_CIVIC_INTEGRITY — the caller's explicit choice), matching
+    // the openai-to-gemini.ts standard-path behavior. See DEFAULT_SAFETY_SETTINGS (#8231).
+    safetySettings: body.safetySettings || DEFAULT_SAFETY_SETTINGS,
   };
 
   // ── Generation config ──────────────────────────────────────────

tests/snapshots/translation/openai-to-gemini/basic-chat.json

@@ -33,10 +33,6 @@
     {
       "category": "HARM_CATEGORY_HARASSMENT",
       "threshold": "OFF"
-    },
-    {
-      "category": "HARM_CATEGORY_CIVIC_INTEGRITY",
-      "threshold": "OFF"
     }
   ]
 }

tests/snapshots/translation/openai-to-gemini/modern-gemini-default-thinking.json

@@ -33,10 +33,6 @@
     {
       "category": "HARM_CATEGORY_HARASSMENT",
       "threshold": "OFF"
-    },
-    {
-      "category": "HARM_CATEGORY_CIVIC_INTEGRITY",
-      "threshold": "OFF"
     }
   ]
 }

tests/snapshots/translation/openai-to-gemini/with-system.json

@@ -33,10 +33,6 @@
     {
       "category": "HARM_CATEGORY_HARASSMENT",
       "threshold": "OFF"
-    },
-    {
-      "category": "HARM_CATEGORY_CIVIC_INTEGRITY",
-      "threshold": "OFF"
     }
   ],
   "systemInstruction": {

tests/unit/8231-gemini-civic-integrity-default.test.ts

@@ -0,0 +1,68 @@
+import { test } from "node:test";
+import assert from "node:assert/strict";
+import { claudeToGeminiRequest } from "../../open-sse/translator/request/claude-to-gemini.ts";
+import { openaiToGeminiRequest } from "../../open-sse/translator/request/openai-to-gemini.ts";
+
+// Regression guard for #8231: the standard Gemini API surface (non-Antigravity)
+// unconditionally injected HARM_CATEGORY_CIVIC_INTEGRITY into the default
+// safetySettings, which upstream dynamic validation rejects for some
+// models/endpoints with `safety_settings[4]: element predicate failed` — a hard
+// 400 on every request through that model. #5003 already fixed this for the
+// Antigravity/Cloud Code surface (open-sse/executors/antigravity.ts); this test
+// pins the same behavior for claude-to-gemini.ts and openai-to-gemini.ts.
+
+test("[repro #8231] claude-to-gemini default safetySettings must not force HARM_CATEGORY_CIVIC_INTEGRITY", () => {
+  const result = claudeToGeminiRequest(
+    "gemini-2.5-pro",
+    { messages: [{ role: "user", content: "hi" }] },
+    false,
+    null
+  );
+  const categories = (result.safetySettings as Array<{ category: string }>).map(
+    (s) => s.category
+  );
+  assert.equal(
+    categories.includes("HARM_CATEGORY_CIVIC_INTEGRITY"),
+    false,
+    `got ${JSON.stringify(categories)}`
+  );
+});
+
+test("[repro #8231] openai-to-gemini default safetySettings must not force HARM_CATEGORY_CIVIC_INTEGRITY", () => {
+  const result = openaiToGeminiRequest(
+    "gemini-2.5-pro",
+    { messages: [{ role: "user", content: "hi" }] },
+    false,
+    null
+  );
+  const categories = (result.safetySettings as Array<{ category: string }>).map(
+    (s) => s.category
+  );
+  assert.equal(
+    categories.includes("HARM_CATEGORY_CIVIC_INTEGRITY"),
+    false,
+    `got ${JSON.stringify(categories)}`
+  );
+});
+
+test("[repro #8231] claude-to-gemini preserves caller-supplied safetySettings that explicitly request HARM_CATEGORY_CIVIC_INTEGRITY", () => {
+  const explicit = [{ category: "HARM_CATEGORY_CIVIC_INTEGRITY", threshold: "BLOCK_NONE" }];
+  const result = claudeToGeminiRequest(
+    "gemini-2.5-pro",
+    { messages: [{ role: "user", content: "hi" }], safetySettings: explicit },
+    false,
+    null
+  );
+  assert.deepEqual(result.safetySettings, explicit);
+});
+
+test("[repro #8231] openai-to-gemini preserves caller-supplied safetySettings that explicitly request HARM_CATEGORY_CIVIC_INTEGRITY", () => {
+  const explicit = [{ category: "HARM_CATEGORY_CIVIC_INTEGRITY", threshold: "BLOCK_NONE" }];
+  const result = openaiToGeminiRequest(
+    "gemini-2.5-pro",
+    { messages: [{ role: "user", content: "hi" }], safetySettings: explicit },
+    false,
+    null
+  );
+  assert.deepEqual(result.safetySettings, explicit);
+});