fix(ai): append system prompt to codex bridge message instead of converting to input

Previously the system prompt was converted to an input message in convertMessages,
then stripped out by filterPiSystemPrompts. Now the system prompt is passed directly
to transformRequestBody and appended after CODEX_PI_BRIDGE in the bridge message.

packages/ai/test/total-tokens.test.ts

@@ -24,8 +24,9 @@ const oauthTokens = await Promise.all([
 	resolveApiKey("github-copilot"),
 	resolveApiKey("google-gemini-cli"),
 	resolveApiKey("google-antigravity"),
+	resolveApiKey("openai-codex"),
 ]);
-const [anthropicOAuthToken, githubCopilotToken, geminiCliToken, antigravityToken] = oauthTokens;
+const [anthropicOAuthToken, githubCopilotToken, geminiCliToken, antigravityToken, openaiCodexToken] = oauthTokens;
 
 // Generate a long system prompt to trigger caching (>2k bytes for most providers)
 const LONG_SYSTEM_PROMPT = `You are a helpful assistant. Be concise in your responses.
@@ -533,4 +534,27 @@ describe("totalTokens field", () => {
 			},
 		);
 	});
+
+	// =========================================================================
+	// OpenAI Codex (OAuth)
+	// =========================================================================
+
+	describe("OpenAI Codex (OAuth)", () => {
+		it.skipIf(!openaiCodexToken)(
+			"gpt-5.2-xhigh - should return totalTokens equal to sum of components",
+			{ retry: 3, timeout: 60000 },
+			async () => {
+				const llm = getModel("openai-codex", "gpt-5.2-xhigh");
+
+				console.log(`\nOpenAI Codex / ${llm.id}:`);
+				const { first, second } = await testTotalTokensWithCache(llm, { apiKey: openaiCodexToken });
+
+				logUsage("First request", first);
+				logUsage("Second request", second);
+
+				assertTotalTokensEqualsComponents(first);
+				assertTotalTokensEqualsComponents(second);
+			},
+		);
+	});
 });