fix(translator): treat developer role as system in Claude translation (#2407)

openAIToClaude now extracts developer-role messages into systemParts
(alongside system role) and filters them from nonSystemMessages, so
identity context injected via the Responses API developer role reaches
the Claude system field instead of silently becoming an assistant turn.

tests/unit/translator-openai-to-claude.test.ts

@@ -498,3 +498,32 @@ test("OpenAI -> Claude handles assistant tool call messages without reasoning_co
   assert.ok(toolUseBlock, "expected tool_use block to be preserved");
   assert.equal(toolUseBlock.name, `${CLAUDE_OAUTH_TOOL_PREFIX}do_thing`);
 });
+
+test("OpenAI -> Claude treats developer role as system (fix for Responses API → Claude path)", () => {
+  const MARKER = "MAGIC_IDENTITY_847261";
+  const result = openaiToClaudeRequest(
+    "claude-sonnet-4-6",
+    {
+      messages: [
+        { role: "developer", content: `You are ${MARKER}` },
+        { role: "user", content: "What is your identity?" },
+      ],
+    },
+    true
+  );
+
+  // developer content must appear in the system field, not as an assistant turn
+  const systemText =
+    typeof result.system === "string" ? result.system : JSON.stringify(result.system);
+  assert.ok(systemText.includes(MARKER), "developer content must appear in Claude system field");
+
+  // No message in the messages array should contain the marker
+  for (const msg of result.messages) {
+    const msgStr = JSON.stringify(msg);
+    assert.ok(!msgStr.includes(MARKER), "developer content must NOT appear in messages array");
+  }
+
+  // Exactly one user message
+  const userMessages = result.messages.filter((m) => m.role === "user");
+  assert.equal(userMessages.length, 1, "expected exactly one user message");
+});